Ipopt Ation Manual

manual_IPOPT

User Manual: Pdf

Open the PDF directly: View PDF .
Page Count: 94

Download
Open PDF In Browser	View PDF

Introduction to Ipopt:
A tutorial for downloading, installing, and using Ipopt
Revision number of this document: Revision : 2567
April 16, 2015
Abstract
This document is a guide to using Ipopt 3.11. It includes instructions on how to obtain and
compile Ipopt, a description of the interface, user options, etc., as well as a tutorial on how to solve
a nonlinear optimization problem with Ipopt.

History of this document
The initial version of this document was created by Yoshiaki Kawajir1 as a course project for 47852
Open Source Software for Optimization, taught by Prof. François Margot at Tepper School of Business,
Carnegie Mellon University. After this, Carl Laird2 has added significant portions, including the very
nice tutorials. The current version is maintained by Stefan Vigerske3 and Andreas Wächter4 .

Contents
1 Introduction
1.1 Mathematical Background . . . . . .
1.2 Availability . . . . . . . . . . . . . .
1.3 Prerequisites . . . . . . . . . . . . .
1.4 How to use Ipopt . . . . . . . . . .
1.5 More Information and Contributions
1.6 History of Ipopt . . . . . . . . . . .

.
.
.
.
.
.

2 Installing Ipopt
2.1 Getting System Packages (Compilers, ...) . .
2.2 Getting the Ipopt Code . . . . . . . . . . . .
2.2.1 Getting the Ipopt code via subversion
2.2.2 Getting the Ipopt code as a tarball .
2.3 Download External Code . . . . . . . . . . .
2.3.1 Download BLAS, LAPACK and ASL
2.3.2 Download HSL Subroutines . . . . . .
2.3.3 Obtaining the MUMPS Linear Solver
2.3.4 Obtaining the Linear Solver Pardiso .
2.3.5 Obtaining the Linear Solver WSMP .
2.3.6 Using the Linear Solver Loader . . . .
2.3.7 Obtaining METIS . . . . . . . . . . .
1 then

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh PA
Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh PA
3 GAMS Software GmbH
4 Department of Industrial Engineering and Management Sciences, Northwestern University
2 then

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.

3
3
3
4
6
7
8

.
.
.
.
.
.
.
.
.
.
.
.

8
8
9
9
9
9
10
10
11
11
12
12
12

2.4
2.5

2.6
2.7
2.8

2.9

Compiling and Installing Ipopt . . . . . . . . . . . . . . . . . .
Installation on Windows . . . . . . . . . . . . . . . . . . . . . .
2.5.1 Installation with Cygwin using GNU compilers . . . . .
2.5.2 Installation with Cygwin using the MSVC++ compiler .
2.5.3 Installation with MSYS/MinGW . . . . . . . . . . . . .
Compiling and Installing the Java Interface JIpopt . . . . . .
Compiling and Installing the R Interface ipoptr . . . . . . . .
Compiling and Installing the MATLAB interface . . . . . . . .
2.8.1 Setting up mex . . . . . . . . . . . . . . . . . . . . . . .
2.8.2 Adjusting configuration and build of Ipopt . . . . . . .
2.8.3 Building the MATLAB interface . . . . . . . . . . . . .
2.8.4 Making MATLAB aware of the mex file . . . . . . . . .
2.8.5 Additional notes . . . . . . . . . . . . . . . . . . . . . .
2.8.6 Troubleshooting . . . . . . . . . . . . . . . . . . . . . .
Expert Installation Options for Ipopt . . . . . . . . . . . . . .

3 Interfacing your NLP to Ipopt
3.1 Using Ipopt through AMPL . . . . . . . .
3.1.1 Using Ipopt from the command line
3.2 Interfacing with Ipopt through code . . . .
3.3 The C++ Interface . . . . . . . . . . . . . .
3.3.1 Coding the Problem Representation
3.3.2 Coding the Executable (main) . . .
3.3.3 Compiling and Testing the Example
3.3.4 Additional methods in TNLP . . . . .
3.4 The C Interface . . . . . . . . . . . . . . . .
3.5 The Fortran Interface . . . . . . . . . . . .
3.6 The Java Interface . . . . . . . . . . . . . .
3.7 The R Interface . . . . . . . . . . . . . . . .
3.8 The MATLAB Interface . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

12
14
14
16
16
17
18
18
18
19
20
20
20
21
21

.
.
.
.
.
.
.
.
.
.
.
.
.

23
24
25
26
28
28
37
39
40
44
47
47
51
55

4 Special Features
4.1 Derivative Checker . . . . . . . . . . . . . . . . . . .
4.2 Quasi-Newton Approximation of Second Derivatives
4.3 Warm-Starting Capabilities via AMPL . . . . . . . .
4.4 sIpopt: Optimal Sensitivity Based on Ipopt . . . .

.
.
.
.

56
56
57
57
58

.
.
.
.
.
.
.
.
.
.
.
.
.

5 Ipopt Options

6 Ipopt Output
6.1 Diagnostic Tags for Ipopt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60
63

A Triplet Format for Sparse Matrices

B The Smart Pointer Implementation: SmartPtr

C Options Reference
C.1 Output . . . . . .
C.2 Termination . . . .
C.3 NLP Scaling . . . .
C.4 NLP . . . . . . . .
C.5 Initialization . . .
C.6 Barrier Parameter

.
.
.
.
.
.

67
67
69
70
71
72
73

C.7 Multiplier Updates . .
C.8 Line Search . . . . . .
C.9 Warm Start . . . . . .
C.10 Restoration Phase . .
C.11 Linear Solver . . . . .
C.12 Hessian Perturbation .
C.13 Quasi-Newton . . . . .
C.14 Derivative Test . . . .
C.15 MA27 Linear Solver .
C.16 MA57 Linear Solver .
C.17 MA77 Linear Solver .
C.18 MA86 Linear Solver .
C.19 MA97 Linear Solver .
C.20 MUMPS Linear Solver
C.21 Pardiso Linear Solver .
C.22 WSMP Linear Solver .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

D Options available via the AMPL Interface

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

76
76
77
78
79
80
81
83
83
84
85
86
87
90
90
91
92

The following names used in this document are trademarks or registered trademarks: AMPL, IBM, Intel, Matlab,
Microsoft, MKL, Visual Studio C++, Visual Studio C++ .NET

Introduction

Ipopt (Interior Point Optimizer, pronounced “Eye–Pea–Opt”) is an open source software package for
large-scale nonlinear optimization. It can be used to solve general nonlinear programming problems of
the form
min

f (x)

s.t.

g L ≤ g(x) ≤ g U

x∈Rn

(1)

x ≤x≤x ,

(2)
(3)

where x ∈ Rn are the optimization variables (possibly with lower and upper bounds, xL ∈ (R ∪ {−∞})n
and xU ∈ (R ∪ {+∞})n ), f : Rn −→ R is the objective function, and g : Rn −→ Rm are the general
nonlinear constraints. The functions f (x) and g(x) can be linear or nonlinear and convex or non-convex
(but should be twice continuously differentiable). The constraints, g(x), have lower and upper bounds,
g L ∈ (R ∪ {−∞})m and g U ∈ (R ∪ {+∞})m . Note that equality constraints of the form gi (x) = ḡi can
be specified by setting giL = giU = ḡi .

1.1

Mathematical Background

Ipopt implements an interior point line search filter method that aims to find a local solution of (1)-(3).
The mathematical details of the algorithm can be found in several publications [6, 8, 13, 11, 10].

1.2

Availability

The Ipopt package is available from COIN-OR (http://www.coin-or.org) under the EPL (Eclipse
Public License) open-source license and includes the source code for Ipopt. This means, it is available
free of charge, also for commercial purposes. However, if you give away software including Ipopt code (in
source code or binary form) and you made changes to the Ipopt source code, you are required to make
those changes public and to clearly indicate which modifications you made. After all, the goal of open

source software is the continuous development and improvement of software. For details, please refer to
the Eclipse Public License.
Also, if you are using Ipopt to obtain results for a publication, we politely ask you to point out in
your paper that you used Ipopt, and to cite the publication [13]. Writing high-quality numerical software
takes a lot of time and effort, and does usually not translate into a large number of publications, therefore
we believe this request is only fair :). We also have space at the Ipopt project home page where we list
publications, projects, etc., in which Ipopt has been used. We would be very happy to hear about your
experiences

1.3

Prerequisites

In order to build Ipopt, some third party components are required:
• BLAS (Basic Linear Algebra Subroutines). Many vendors of compilers and operating systems
provide precompiled and optimized libraries for these dense linear algebra subroutines. You can
also get the source code for a simple reference implementation from www.netlib.org and have the
Ipopt distribution compile it automatically. However, it is strongly recommended to use some
optimized BLAS implemetion, for large problems this can make a runtime difference of an order of
magnitude!
Examples for efficient BLAS implementations are:
– From hardware vendors:
∗ ACML (AMD Core Math Library) by AMD
∗ ESSL (Engineering Scientific Subroutine Library) by IBM
∗ MKL (Math Kernel Library) by Intel
∗ Sun Performance Library by Sun
– Generic:
∗ Atlas (Automatically Tuned Linear Algebra Software)
∗ GotoBLAS
You find more information on the web by googling them.
Note: BLAS libraries distributed with Linux are usually not optimized.
• LAPACK (Linear Algebra PACKage). Also for LAPACK, some vendors offer precompiled and
optimized libraries. But like with BLAS, you can get the source code from http://www.netlib.org
and have the Ipopt distribution compile it automatically.
Note that currently LAPACK is only required if you intend to use the quasi-Newton options in
Ipopt. You can compile the code without LAPACK, but an error message will then occur if you
try to run the code with an option that requires LAPACK. Currently, the LAPACK routines that
are used by Ipopt are only DPOTRF, DPOTRS, and DSYEV.
Note: LAPACK libraries distributed with Linux are usually not optimized.
• A sparse symmetric indefinite linear solver. Ipopt needs to obtain the solution of sparse, symmetric,
indefinite linear systems, and for this it relies on third-party code.
Currently, the following linear solvers can be used:
– MA27 from the HSL Mathematical Software Library
(see http://www.hsl.rl.ac.uk).
– MA57 from the HSL Mathematical Software Library
(see http://www.hsl.rl.ac.uk).

– HSL MA77 from the HSL Mathematical Software Library
(see http://www.hsl.rl.ac.uk).
– HSL MA86 from the HSL Mathematical Software Library
(see http://www.hsl.rl.ac.uk).
– HSL MA97 from the HSL Mathematical Software Library
(see http://www.hsl.rl.ac.uk).
– MUMPS (MUltifrontal Massively Parallel sparse direct Solver)
(see http://graal.ens-lyon.fr/MUMPS)
– The Parallel Sparse Direct Solver (PARDISO)
(see http://www.pardiso-project.org).
– The Watson Sparse Matrix Package (WSMP)
(see http://researcher.ibm.com/view_project.php?id=1426)
You should include at least one of the linear solvers above in order to run Ipopt, and if you want
to be able to switch easily between different alternatives, you can compile Ipopt with all of them.
The Ipopt library also has mechanisms to load the linear solvers MA27, MA57, HSL MA77,
HSL MA86, HSL MA97, and Pardiso from a shared library at runtime, if the library has not
been compiled with them (see Section 2.3.6).
NOTE: The solution of the linear systems is a central ingredient in Ipopt and the
optimizer’s performance and robustness depends on your choice. The best choice
depends on your application, and it makes sense to try different options. Most of the
solvers also rely on efficient BLAS code (see above), so you should use a good BLAS
library tailored to your system. Please keep this in mind, particularly when you are
comparing Ipopt with other optimization codes.
If you are compiling MA57, HSL MA77, HSL MA86, HSL MA97, or MUMPS within the Ipopt
build system, you should also include the METIS linear system ordering package.
Interfaces to other linear solvers might be added in the future; if you are interested in contributing
such an interface please contact us! Note that Ipopt requires that the linear solver is able to
provide the inertia (number of positive and negative eigenvalues) of the symmetric matrix that is
factorized.
• Furthermore, Ipopt can also use the HSL package MC19 for scaling of the linear systems before
they are passed to the linear solver. This may be particularly useful if Ipopt is used with MA27
or MA57. However, it is not required to have MC19 to compile Ipopt; if this routine is missing,
the scaling is never performed.
• ASL (AMPL Solver Library). The source code is available at www.netlib.org, and the Ipopt
makefiles will automatically compile it for you if you put the source code into a designated space.
NOTE: This is only required if you want to use Ipopt from AMPL and want to compile the Ipopt
AMPL solver executable.
For more information on third-party components and how to obtain them, see Section 2.3.
Since the Ipopt code is written in C++, you will need a C++ compiler to build the Ipopt library.
We tried very hard to write the code as platform and compiler independent as possible.
In addition, the configuration script also searches for a Fortran compiler, since some of the dependencies above are written in Fortran. If all third party dependencies are available as self-contained libraries,
those compilers are in principle not necessary. Also, it is possible to use the Fortran-to-C compiler f2c
from http://www.netlib.org/f2c to convert Fortran 77 code to C, and compile the resulting C files
with a C compiler and create a library containing the required third party dependencies.
When using GNU compilers, we recommend you use the same version numbers for gcc, g++, and
gfortran. For gfortran specifically, we recommend versions newer than 4.5.2 (versions 4.5.1, 4.5.2, and
5

before 4.2.0 are known to have bugs that caused issues with some of the newer Fortran 90 HSL linear
solvers).

1.4

How to use Ipopt

If desired, the Ipopt distribution generates an executable for the modeling environment AMPL. As well,
you can link your problem statement with Ipopt using interfaces for C++, C, or Fortran. Ipopt can
be used with most Linux/Unix environments, and on Windows using Visual Studio .NET, Cygwin or
MSYS/MinGW. In Section 3 this document demonstrates how to solve problems using Ipopt. This
includes installation and compilation of Ipopt for use with AMPL as well as linking with your own code.
Additionally, the Ipopt distribution includes interfaces for
• CUTEr5 (for solving problems modeled in SIF),
• Java, which allows you to use Ipopt from Java, see the files in the Ipopt/contrib/JavaInterface
directory,
• Matlab (mex interface), which allows you to use Ipopt from Matlab, see
https://projects.coin-or.org/Ipopt/wiki/MatlabInterface,
• and the R project for statistical computing, see the files in the Ipopt/contrib/RInterface directory.
There is also software that facilitates use of Ipopt maintained by other people, among them are:
• ADOL-C (automatic differentiation)
ADOL-C facilitates the evaluation of first and higher derivatives of vector functions that are defined
by computer programs written in C or C++. It comes with examples that show how to use it in
connection with Ipopt, see https://projects.coin-or.org/ADOL-C.
• AIMMS (modeling environment)
The AIMMSlinks project on COIN-OR, maintained by Marcel Hunting, provides an interface for
Ipopt within the AIMMS modeling tool, see https://projects.coin-or.org/AIMMSlinks.
• CasADi
CasADi is a symbolic framework for automatic differentiation and numeric optimization and comes
with Ipopt, see http://casadi.org.
• CppAD (automatic differentiation)
Given a C++ algorithm that computes function values, CppAD generates an algorithm that computes corresponding derivative values (of arbitrary order using either forward or reverse mode). It
comes with an example that shows how to use it in connection with Ipopt, see https://projects.
coin-or.org/CppAD.
• GAMS (modeling environment)
The GAMSlinks project on COIN-OR, maintained by Stefan Vigerske, includes a GAMS interface
for Ipopt, see https://projects.coin-or.org/GAMSlinks.
• JuliaOpt
Julia is a high-level, high-performance dynamic programming language for technical computing.
JuliaOpt, see http://juliaopt.org, is an umbrella group for Julia-based optimization- related
projects. It includes the algebraic modeling language JuMP and an interface to Ipopt.
5 see

http://cuter.rl.ac.uk/cuter-www

• MADOPT (Modelling and Automatic Differentiation for Optimisation)
Light-weight C++ and Python modelling interfaces implementing expression building using operator overloading and automatic differentiation, see https://github.com/stanle/madopt
• .NET :
An interface to the C# language is available here: http://code.google.com/p/csipopt
• OPTimization Interface (OPTI) Toolbox
OPTI is a free Matlab toolbox for constructing and solving linear, nonlinear, continuous and discrete
optimization problem and comes with Ipopt.
• Optimization Services
The Optimization Services (OS) project provides a set of standards for representing optimization
instances, results, solver options, and communication between clients and solvers, incl. Ipopt, in a
distributed environment using Web Services, see https://projects.coin-or.org/OS.
• PyIpopt:
An interface to the python language is available here: https://github.com/xuy/pyipopt
• Scilab (free Matlab-like environment):
A Scilab interface is available here: http://forge.scilab.org/index.php/p/sci-ipopt

1.5

More Information and Contributions

More and up-to-date information can be found at the Ipopt homepage,
http://projects.coin-or.org/Ipopt.
Here, you can find FAQs, some (hopefully useful) hints, a bug report system etc. The website is
managed with Wiki, which means that every user can edit the webpages from the regular web browser.
In particular, we encourage Ipopt users to share their experiences and usage hints on
the “Success Stories” and “Hints and Tricks” pages, or to list the publications discussing
applications of Ipopt in the “Papers related to Ipopt” page6 . In particular, if you have trouble
getting Ipopt work well for your optimization problem, you might find some ideas here. Also, if you had
some difficulties to solve a problem and found a way around it (e.g., by reformulating your problem or
by using certain Ipopt options), it would be very nice if you help other users by sharing your experience
at the “Hints and Tricks” page.
Ipopt is an open source project, and we encourage people to contribute code (such as interfaces to
appropriate linear solvers, modeling environments, or even algorithmic features). If you are interested in
contributing code, please have a look at the COIN-OR contributions webpage7 and contact the Ipopt
project leader.
There is also a mailing list for Ipopt, available from the webpage
http://list.coin-or.org/mailman/listinfo/ipopt,
where you can subscribe to get notified of updates, to ask general questions regarding installation and
usage, or to share your experience with Ipopt. You might want to look at the archives before posting a
question. An easy way to search the archive with Google is to specify
“site:http://list.coin-or.org/pipermail/ipopt”
in addition to your keywords in the search string.
6 Since

we had some malicious hacker attacks destroying the content of the web pages in the past, you are now required
to enter a user name and password; simply follow the instructions on top of the main project page.
7 see http://www.coin-or.org/contributions.html

We try to answer questions posted to the mailing list in a reasonable manner. Please understand that
we cannot answer all questions in detail, and because of time constraints, we may not be able to help
you model and debug your particular optimization problem.
A short tutorial on getting started with Ipopt is also available [9].

1.6

History of Ipopt

The original Ipopt (Fortran version) was a product of the dissertation research of Andreas Wächter [8],
under the supervision of Lorenz T. Biegler at the Chemical Engineering Department at Carnegie Mellon
University. The code was made open source and distributed by the COIN-OR initiative, which is now a
non-profit corporation. Ipopt has been actively developed under COIN-OR since 2002.
To continue natural extension of the code and allow easy addition of new features, IBM Research
decided to invest in an open source re-write of Ipopt in C++. With the help of Carl Laird, who came
to the Mathematical Sciences Department at IBM Research as a summer intern in 2004 and 2005 during
his PhD studies, the code was re-implemented from scratch.
The new C++ version of the Ipopt optimization code (Ipopt 3.0.0 and beyond) was maintained at
IBM Research and remains part of the COIN-OR initiative. The development on the Fortran version has
ceased, but the source code can still be downloaded from http://www.coin-or.org/download/source/
Ipopt-Fortran.

Installing Ipopt

The following sections describe the installation procedures on UNIX/Linux systems. For installation
instructions on Windows see Section 2.5.
Additional hints on installing Ipopt and its various interfaces is available on the Ipopt and CoinHelp
wiki pages, in particular
• Ipopt compilation hints:
https://projects.coin-or.org/Ipopt/wiki/CompilationHints
• Current configuration and installation issues for COIN-OR projects:
https://projects.coin-or.org/BuildTools/wiki/current-issues

2.1

Getting System Packages (Compilers, ...)

Many Linux distributions will come with all necessary tools. All you should need to do is check the
compiler versions. On a Debian-based distribution, you can obtain all necessary tools with the following
command:
sudo apt-get install gcc g++ gfortran subversion patch wget
Replace apt-get with your relevant package manager, e.g. yum for Red Hat-based distributions, zypper
for SUSE, etc. The g++ and gfortran compilers may need to be specified respectively as gcc-c++ and
gcc-gfortran with some package managers.
On Mac OS X, you need either the Xcode Command Line Tools, available at https://developer.
apple.com/downloads after registering as an Apple Developer, or a community alternative such as
https://github.com/kennethreitz/osx-gcc-installer/downloads to install the gcc and g++ compilers. It has been reported, that gcc/g++ 4.2 and older is not sufficient for using the HSL codes. If
you have a recent version of Xcode installed, the Command Line Tools are available under Preferences,
Downloads. In Xcode 3.x, the Command Line Tools are contained in the optional item “UNIX Dev
Support” during Xcode installation. These items unfortunately do not come with a Fortran compiler,
but you can get gfortran from http://gcc.gnu.org/wiki/GFortranBinaries#MacOS. We have been
able to compile Ipopt using default Xcode versions of gcc and g++ and a newer version of gfortran
from this link, but consistent version numbers may be an issue in future cases.
8

2.2

Getting the Ipopt Code

Ipopt is available from the COIN-OR subversion repository. You can either download the code using
svn (the subversion client similar to CVS) or simply retrieve a tarball (compressed archive file). While
the tarball is an easy method to retrieve the code, using the subversion system allows users the benefits
of the version control system, including easy updates and revision control.
2.2.1

Getting the Ipopt code via subversion

Of course, the subversion client must be installed on your system if you want to obtain the code this
way (the executable is called svn); it is already installed by default for many recent Linux distributions.
Information about subversion and how to download it can be found at http://subversion.apache.org.
To obtain the Ipopt source code via subversion, change into the directory in which you want to create
a subdirectory Ipopt with the Ipopt source code. Then follow the steps below:
1. Download the code from the repository
$ svn co https://projects.coin-or.org/svn/Ipopt/stable/3.11 CoinIpopt
Note: The $ indicates the command line prompt, do not type $, only the text following it.
2. Change into the root directory of the Ipopt distribution
$ cd CoinIpopt
In the following, “$IPOPTDIR” will refer to the directory in which you are right now (output of pwd).
2.2.2

Getting the Ipopt code as a tarball

To use the tarball, follow the steps below:
1. Download the desired tarball from http://www.coin-or.org/download/source/Ipopt, it has the
form Ipopt-x.y.z.tgz, where x.y.z is the version number, such as 3.11.0. There might also be
daily snapshot from the stable branch. The number of the latest official release can be found on
the Ipopt Trac page.
2. Issue the following commands to unpack the archive file:
$ gunzip Ipopt-x.y.z.tgz
$ tar xvf Ipopt-x.y.z.tar
Note: The $ indicates the command line prompt, do not type $, only the text following it.
3. Rename the directory you just extracted:
$ mv Ipopt-x.y.z CoinIpopt
4. Change into the root directory of the Ipopt distribution
$ cd CoinIpopt
In the following, “$IPOPTDIR” will refer to the directory in which you are right now (output of pwd).

2.3

Download External Code

Ipopt uses a few external packages that are not included in the Ipopt source code distribution, namely
ASL (the AMPL Solver Library if you want to compile the Ipopt AMPL solver executable), Blas, Lapack.
Ipopt also requires at least one linear solver for sparse symmetric indefinite matrices. There are
different possibilities, see Sections 2.3.2–2.3.5. It is important to keep in mind that usually the
largest fraction of computation time in the optimizer is spent for solving the linear system,
and that your choice of the linear solver impacts Ipopt’s speed and robustness. It might be
worthwhile to try different linear solver to experiment with what is best for your application.
Since this third party software is released under different licenses than Ipopt, we cannot distribute
their code together with the Ipopt packages and have to ask you to go through the hassle of obtaining
9

it yourself (even though we tried to make it as easy for you as we could). Keep in mind that it is still
your responsibility to ensure that your downloading and usage of the third party components conforms
with their licenses.
Note that you only need to obtain the ASL if you intend to use Ipopt from AMPL. It is not required
if you want to specify your optimization problem in a programming language (C++, C, or Fortran). Also,
currently, Lapack is only required if you intend to use the quasi-Newton options implemented in Ipopt.
2.3.1

Download BLAS, LAPACK and ASL

Note: It is highly recommended that you obtain an efficient implementation of the BLAS
library, tailored to your hardware; Section 1.3 lists a few options. Assuming that your precompiled efficient BLAS library is libmyblas.a in $HOME/lib, you need to add the flag --with-blas="-L$HOME/lib
-lmyblas" when you run configure (see Section 2.4). Some of those libraries also include LAPACK.
If you have the download utility wget installed on your system (or ftp on Mac OS X), retrieving source
code for BLAS (the inefficient reference implementation, not required if you have a precompiled library),
as well as LAPACK and ASL is straightforward using scripts included with the ipopt distribution. These
scripts download the required files from the Netlib Repository (http://www.netlib.org).
$
$
$
$
$
$

cd $IPOPTDIR/ThirdParty/Blas
./get.Blas
cd ../Lapack
./get.Lapack
cd ../ASL
./get.ASL

If you do not have wget (or ftp on Mac OS X) installed on your system, please read the INSTALL.* files
in the $IPOPTDIR/ThirdParty/Blas, $IPOPTDIR/ThirdParty/Lapack and $IPOPTDIR/ThirdParty/ASL
directories for alternative instructions.
If you are having firewall issues with wget, try opening the get. scripts and replace the
line wgetcmd=wget with wgetcmd="wget --passive-ftp".
If you are getting permissions errors from tar, try opening the get. scripts and replace any
instances of tar xf with tar --no-same-owner -xf.
2.3.2

Download HSL Subroutines

There are two versions of HSL available:
HSL Archive contains outdated codes that are freely available for personal commercial or non-commercial usage. Note that you may not redistribute these codes in either source or binary form without
purchasing a licence from the authors. This version includes MA27, MA28, and MC19.
HSL 2011 contains more modern codes that are freely available for academic use only. This version
includes the codes from the HSL Archive and additionally MA57, HSL MA77, HSL MA86, and
HSL MA97. Ipopt supports the HSL 2011 codes from 2012 and 2013, the support for the versions
from 2012 may be dropped in a future release.
To obtain the HSL code, you can follow the following steps:
1. Go to http://hsl.rl.ac.uk/ipopt.
2. Choose whether to download either the Archive code or the HSL 2011 code. To download, select
the relevant “source” link.
3. Follow the instructions on the website, read the license, and submit the registration form.
4. Wait for an email containing a download link (this should take no more than one working day).

You may either:
• Compile the HSL code as part of Ipopt. See the instructions below.
• Compile the HSL code separately either before or after the Ipopt code and use the shared library
loading mechanism. See the documentation distributed with the HSL package for information on
how to do so.
To compile the HSL code as part of Ipopt, unpack the archive, then move and rename the resulting
directory so that it becomes $IPOPTDIR/ThirdParty/HSL/coinhsl. Ipopt may then be configured as
normal.
Note: Whereas it is essential to have at least one linear solver, the package MC19 could be omitted
(with the consequence that you cannot use this method for scaling the linear systems arising inside the
Ipopt algorithm). By default, MC19 is only used to scale the linear system when using one of the
HSL solvers, but it can also be switched on for other linear solvers (which usually have internal scaling
mechanisms). Further, also the package MA28 can be omitted, since it is used only in the experimental
dependency detector, which is not used by default.
Note: If you are an academic or a student, we recommend you download the HSL 2011 package as
this ensures you have access to the full range of solvers. MA57 can be considerably faster than MA27 on
some problems.
Yet another note: If you have a precompiled library containing the HSL codes, you can specify the
directory with the header files and the linker flags for this library with the --with-hsl-incdir and
--with-hsl-lib flags for the configure script described in Section 2.4.
2.3.3

Obtaining the MUMPS Linear Solver

You can also use the (public domain) sparse linear solver MUMPS. Please visit the MUMPS home
page http://graal.ens-lyon.fr/MUMPS for more information about the solver. MUMPS is provided as
Fortran 90 and C source code. You need to have a Fortran 90 compiler (for example, the GNU compiler
gfortran is a free one) to be able to use it.
You can obtain the MUMPS code by running the script $IPOPTDIR/ThirdParty/Mumps/get.Mumps if
you have wget (or ftp on Mac OS X) installed in your system. Alternatively, you can get the latest version
from the MUMPS home page and extract the archive in the directory $IPOPTDIR/ThirdParty/Mumps.
The extracted directory usually has the MUMPS version number in it, so you need to rename it to MUMPS
such that you have a file called $IPOPTDIR/ThirdParty/Mumps/MUMPS/README.
Once you put the MUMPS source code into the correct place, the Ipopt configuration scripts will
automatically detect it and compile MUMPS together with Ipopt, if your Fortran compiler is able to
compile Fortran 90 code.
Note: MUMPS will perform better with METIS, see Section 2.3.7.
Note: MUMPS uses interally a fake implementation of MPI. If you are using Ipopt within an MPI
program together with MUMPS, the code will not run. You will have to modify the MUMPS sources so
that the MPI symbols inside the MUMPS code are renamed.
2.3.4

Obtaining the Linear Solver Pardiso

If you would like to compile Ipopt with the Parallel Sparse Direct Linear Solver (Pardiso), you need to
obtain the Pardiso library from http://www.pardiso-project.org for your operating system. (Alternatively, experimental support for Pardiso from Intel’s MKL library is available.)
From http://www.pardiso-project.org, you can obtain a limited time license of Pardiso for academic or evaluation purposes or buy a non-profit or commercial license. Make sure you read the license
agreement before filling out the download form.
Please consult Appendix 2.9 to find out how to configure your Ipopt installation to work with Pardiso.

2.3.5

Obtaining the Linear Solver WSMP

If you would like to compile Ipopt with the Watson Sparse Matrix Package (WSMP), you need to
obtain the WSMP library for your operating system. Information about WSMP can be found at http:
//www.research.ibm.com/projects/wsmp.
At this website you can download the library for several operating systems including a trial license
key for 90 days that allows you to use WSMP for “educational, research, and benchmarking purposes
by non-profit academic institutions” or evaluation purposes by commercial organizations; make sure you
read the license agreement before using the library. Once you obtained the library and license, please
check if the version number of the library matches the one on the WSMP website.
If a newer version is announced on that website, you can (and probably should) request the current
version by sending a message to wsmp@watson.ibm.com. Please include the operating system and other
details to describe which particular version of WSMP you need.
Note: Only the interface to the shared-memory version of WSMP is currently supported.
Please consult Appendix 2.9 to find out how to configure your Ipopt installation to work with WSMP.
2.3.6

Using the Linear Solver Loader

By default, Ipopt will be compiled with a mechanism, the Linear Solver Loader, which can dynamically
load shared libraries with MA27, MA57, HSL MA77, HSL MA86, HSL MA97, or the Pardiso linear solver
at runtime8 . This means, if you obtain one of those solvers after you compiled Ipopt, you don’t need to
recompile to use it. Instead, you can just put a shared library called libhsl.so or libpardiso.so into
the shared library search path, LD LIBRARY PATH. These are the names on most UNIX platforms, including
Linux. On Mac OS X, the names are libhsl.dylib, libpardiso.dylib, and DYLD LIBRARY PATH. On
Windows, the names are libhsl.dll, libpardiso.dll, and PATH.
The Pardiso shared library can be downloaded from the Pardiso website. To create a shared library
containing the HSL linear solvers, read the instructions in $IPOPTDIR/ThirdParty/HSL/INSTALL.HSL.
2.3.7

Obtaining METIS

The linear solvers MA57, HSL MA77, HSL MA86, HSL MA97, and MUMPS can make use of the matrix
ordering algorithms implemented in METIS (see http://glaros.dtc.umn.edu/gkhome/metis/metis/
overview). If you are using one of these linear solvers, you should obtain the METIS source code and
put it into $IPOPTDIR/ThirdParty/Metis. Read the INSTALL.Metis file in that directory, and if you
have the wget utility (or ftp on Mac OS X) installed on your system, you can download the code by
running the ./get.Metis script.
Note, that only the older METIS 4.x version9 is supported by MA57, HSL MA77, HSL MA86,
HSL MA97, MUMPS, and the build system. The ./get.Metis script takes care of downloading the right
METIS version.

2.4

Compiling and Installing Ipopt

Ipopt can be easily compiled and installed with the usual configure, make, make install commands.
We follow the procedure that is used for most of the COIN-OR projects, based on the GNU autotools.
At https://projects.coin-or.org/CoinHelp you can find a general description of the tools.
Below are the basic steps for the Ipopt compilation that should work on most systems. For special
compilations and for some troubleshooting see Appendix 2.9 and consult the generic COIN-OR help
page https://projects.coin-or.org/CoinHelp before submitting a ticket or sending a message to the
mailing list.
8 This is not enabled if you compile Ipopt with the MS Visual Studio project files provided in the Ipopt distribution. Further, if you have problems compiling this new feature, you can disable this by specifying --disable-linear-solver-loader
for the configure script
9 http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/OLD/metis-4.0.3.tar.gz

1. Create a directory where you want to compile Ipopt, for example
$ mkdir $IPOPTDIR/build
and go into this direcrory
$ cd $IPOPTDIR/build
Note: You can choose any location, including $IPOPTDIR itself, as the location of your compilation.
However, on COIN-OR we recommend to keep the source and compiled files separate.
2. Run the configure script
$ $IPOPTDIR/configure
One might have to give options to the configure script, e.g., in order to choose a non-default
compiler, or to tell it where some third party code is installed, see Appendix 2.9.
If the last output line reads “configure: Main configuration of Ipopt successful” then everything worked fine. Otherwise, look at the screen output, have a look at the config.log output
files and/or consult Appendix 2.9.
The default configure (without any options) is sufficient for most users that downloaded the source
code for the linear solver. If you want to see the configure options, consult Appendix 2.9, and also
visit the generic COIN-OR configuration instruction page at
https://projects.coin-or.org/CoinHelp/wiki/user-configure
3. Build the code
$ make
Note: If you are using GNU make, you can also try to speed up the compilation by using the -jN
flag (e.g., make -j3), where N is the number of parallel compilation jobs. A good number for N is
the number of available processors plus one. Under some circumstances, this fails, and you might
have to re-issue the command, or omit the -j flag.
4. If you want, you can run a short test to verify that the compilation was successful. For this, you
just enter
$ make test
This will test if the AMPL solver executable works (if you got the ASL code) and if the included
C++, C, and Fortran examples work.
Note: The configure script is not able to automatically determine the C++ runtime libraries for
the C++ compiler. For certain compilers we enabled default values for this, but those might not
exist or be wrong for your compiler. In that case, the C and Fortran example in the test will most
probably fail to compile. If you don’t want to hook up the compiled Ipopt library to some Fortran or
C code that you wrote you don’t need to worry about this. If you do want to link the Ipopt library
with a C or Fortran compiler, you need to find out the C++ runtime libraries (e.g., by running the
C++ compiler in verbose mode for a simple example program) and run configure again, and this
time specify all C++ runtime libraries with the CXXLIBS variable (see also Appendix 2.9).
5. Install Ipopt
$ make install
This installs
• the Ipopt AMPL solver executable (if ASL source was downloaded) in $IPOPTDIR/build/bin,
• the Ipopt library (libipopt.so, libipopt.a or similar) and all its dependencies (MUMPS,
HSL, Metis libraries) in $IPOPTDIR/build/lib,
• text files ipopt addlibs cpp.txt, ipopt addlibs c.txt, and ipopt addlibs f.txt in
$IPOPTDIR/build/share/coin/doc/Ipopt each containing a line with linking flags that are
required for linking code with the Ipopt library for C++, C, and Fortran main programs,
respectively. (This is only for convenience if you want to find out what additional flags are
required, for example, to include the Fortran runtime libraries with a C++ compiler.)
13

• the necessary header files in $IPOPTDIR/build/include/coin.
You can change the default installation directory (here $IPOPTDIR/build) to something else (such
as /usr/local) by using the --prefix switch for configure.
6. (Optional) Install Ipopt for use with CUTEr
If you have CUTEr already installed on your system and you want to use Ipopt as a solver for
problems modeled in SIF, type
$ make cuter
This assumes that you have the environment variable MYCUTER defined according to the CUTEr
instructions. After this, you can use the script sdipo as the CUTEr script to solve a SIF model.
Note: The above procedures show how to compile the code in directories separate from the source
files. This comes in handy when you want to compile the code with different compilers, compiler options,
or different operating system that share a common file system. To use this feature, change into the
directory where you want to compile the code, and then type $IPOPTDIR/configure with all the options.
For this, the directories with the Ipopt source must not have any configuration and compiled code.

2.5

Installation on Windows

There are several ways to install Ipopt on Windows systems. The first two options, described in Sections 2.5.1 and 2.5.2, are to use Cygwin (see http://www.cygwin.com), which offers a comprehensive
UNIX-like environment on Windows and in which the installation procedure described earlier in this
section can be used. If you want to use the (free) GNU compilers, follow the instructions in Section 2.5.1.
If you have the Microsoft C++ compiler and possibly a “native” Fortran compiler (e.g., the Intel Fortran
compiler) and want to use those to compile Ipopt, please see Section 2.5.2. If you use MSYS/MinGW
(a light-weight UNIX-like environment for Windows), please consider the notes in Section 2.5.3. If you
want to compile the Ipopt mex interface to MATLAB, then we recommend to use the MSYS/MinGW
option.
Note: Some binaries for Ipopt are available on the COIN-OR website at http://www.coin-or.org/
download/binary/Ipopt. There, also precompiled versions of Ipopt as DLLs (generated from the MSVS
solution in Ipopt’s subdirectory $IPOPTDIR/Ipopt/MSVisualStudio/v8-ifort) are available. Look at
the README files for details. An example how to use these DLLs from your own MSVS project is in
$IPOPTDIR/Ipopt/MSVisualStudio/BinaryDLL-Link-Example.
2.5.1

Installation with Cygwin using GNU compilers

Cygwin is a Linux-like environment for Windows; if you don’t know what it is you might want to have a
look at the Cygwin homepage, http://www.cygwin.com.
It is possible to build the Ipopt AMPL solver executable in Cygwin for general use in Windows. You
can also hook up Ipopt to your own program if you compile it in the Cygwin environment10 .
If you want to compile Ipopt under Cygwin, you first have to install Cygwin on your Windows system.
This is pretty straight forward; you simply download the “setup” program from http://www.cygwin.com
and start it.
Then you do the following steps (assuming here that you don’t have any complications with firewall
settings etc - in that case you might have to choose some connection settings differently):
1. Click next
2. Select “install from the internet” (default) and click next
3. Select a directory where Cygwin is to be installed (you can leave the default) and choose all other
things to your liking, then click next
10 It

is also possible to build an Ipopt DLL that can be used from non-cygwin compilers, but this is not (yet?) supported.

4. Select a temp dir for Cygwin setup to store some files (if you put it on your desktop you will later
remember to delete it)
5. Select “direct connection” (default) and click next
6. Select some mirror site that seems close by to you and click next
7. OK, now comes the complicated part:
You need to select the packages that you want to have installed. By default, there are already
selections, but the compilers are usually not pre-chosen. You need to make sure that you select the
GNU compilers (for Fortran, C, and C++), Subversion, and some additional tools. For this, get
the following packages from the associated branches:
• “Devel”: gcc4
• “Devel”: gcc4-fortran
• “Devel”: pkg-config
• “Devel”: subversion
• “Archive”: unzip
• “Utils”: patch
• “Web”: wget
When a Resolving Dependencies window comes up, be sure to “Select required packages (RECOMMENDED)”. This will automatically also select some other packages.
8. Then you click on next, and Cygwin will be installed (follow the rest of the instructions and choose
everything else to your liking). At a later point you can easily add/remove packages with the setup
program.
9. The version of the GNU Make utility provided by the Cygwin installer will not work. Therefore,
you need to download the fixed version from http://www.cmake.org/files/cygwin/make.exe and
save it to C:\cygwin\bin. Double-check this new version by typing make --version in a Cygwin
terminal (see next point). If you get an error -bash: /usr/bin/make: Bad address, then try
http://www.cmake.org/files/cygwin/make.exe-cygwin1.7 instead, rename it to make.exe and
move it to C:\cygwin\bin. (Replace C:\cygwin with your installation location if different.)
10. Now that you have Cygwin, you can open a Cygwin window, which is like a UNIX shell window.
11. Now you just follow the instructions in the beginning of Section 2: You download the Ipopt
code into your Cygwin home directory (from the Windows explorer that is usually something like
C:\Cygwin\home\your user name). After that you obtain the third party code (as on Linux/UNIX),
type
./configure
and
make install
in the correct directories, and hopefully that will work. The Ipopt AMPL solver executable will
be in the subdirectory bin (called “ipopt.exe”). If you want to set the installation, type
make test

2.5.2

Installation with Cygwin using the MSVC++ compiler

This section describes how you can compile Ipopt with the Microsoft Visual C++ compiler under Cygwin.
Here you have two options for compiling the Fortran code in the third party dependencies:
• Using a Windows Fortran compiler, e.g., the Intel Fortran compiler, which is also able to compile
Fortran 90 code. This would allow you to compile the MUMPS linear solver if you desire to do so.
• Using the f2c Fortran to C compiler, available for free at Netlib (see http://www.netlib.org/f2c).
This can only compile Fortran 77 code (i.e., you won’t be able to compile MUMPS). Before doing
the following installation steps, you need to follow the instructions in
$IPOPTDIR/BuildTools/compile f2c/INSTALL.
Once you have settled on this, do the following:
1. Follow the instructions in Section 2.5.1 until Step 11 and stop after your downloaded the third
party code.
2. Now you need to make sure that Cygwin knows about the native compilers. For this you need to
edit the file cygwin.bat in the Cygwin base directory (usually C:\cygwin). Here you need to add
a line like the following:
call "C:\Program Files\Microsoft Visual Studio 8\VC\vcvarsall.bat"
On my computer, this sets the environment variables so that I can use the MSVC++ compiler.
If you want to use also a native Fortran compiler, you need to include something like this
call "C:\Program Files\Intel\Fortran\compiler80\IA32\BIN\ifortvars.bat"
You might have to search around a bit. The important thing is that, after your change, you can
type “cl” in a newly opened Cygwin windows, and it finds the Microsoft C++ compiler (and if
you want to use it, the Fortran compiler, such as the Intel’s ifort).
3. Run the configuration script, and tell it that you want to use the native compilers:
./configure --enable-doscompile=msvc
Make sure the last message is
Main Ipopt configuration successful
4. Now you can compile the code with
make,
test the installation with
make test,
and install everything with
make install
2.5.3

Installation with MSYS/MinGW

You can compile Ipopt also under MSYS/MinGW, which is another, more light-weight UNIX-like environment for Windows. It can be obtained from http://www.mingw.org/.
If you want to use MSYS/MinGW to compile Ipopt with native Windows compilers (see Section 2.5.2), all you need to install is the basic version11 . If you also want to use the GNU compilers, you
need to install those as well, of course.
11 a convenient Windows install program is available from http://sourceforge.net/projects/mingw/files/Installer/
mingw-get-inst/

A compilation with the GNU compilers works just like with any other UNIX system, as described in
Section 2.4. That is, during the installation, select (at least) the C Compiler, C++ Compiler, Fortran
Compiler, MSYS Basic System, and the MinGW Developer ToolKit. Additionally, wget and unzip
should be installed with the following command in an MSYS terminal:
mingw-get install msys-wget msys-unzip
If you want to use the native MSVC++ compiler (with f2c or a native Fortran compiler), you
essentially follow the steps outlined in Section 2.5.2. Additionally, you need to make sure that the
environment variables are set for the compilers (see step 2), this time adding the line to the msys.bat
file.
For a 64-bit build, you will need to install also a MinGW-64 distribution. We recommend TDM-GCC,
which is available from http://sourceforge.net/projects/tdm-gcc/files/TDM-GCC%20Installer/
tdm-gcc-webdl.exe/download. Install MinGW-64 in a different folder than your existing 32-bit MinGW
installation! The components you need are: core (under gcc), c++ (under gcc), fortran (under gcc),
openmp (under gcc, necessary if you want to use any multi-threaded linear solvers), binutils, and
mingw64-runtime.
After MinGW-64 is installed, open the file C:\MinGW\msys\1.0\etc\fstab, and replace the line
C:\MinGW\

/mingw

with
C:\MinGW64\

/mingw

(Replace paths with your installation locations if different.)

2.6

Compiling and Installing the Java Interface JIpopt
based on documentation by Rafael de Pelegrini Soares 12

JIpopt uses the Java Native Interface (JNI), which is a programming framework that allows Java
code running in the Java Virtual Machine (JVM) to call and be called by native applications and libraries
written in languages such as C and C++. JIpopt requires Java 5 or higher.
After building and installing Ipopt, the JIpopt interface can be build by setting the environment
variable JAVA HOME to the directory that contains your JDK, changing to the JIpopt directory in your
Ipopt build, and issuing make, e.g.,
export JAVA_HOME=/usr/lib/jvm/java-1.5.0
cd $IPOPTDIR/build/Ipopt/contrib/JavaInterface
make
This will generate the Java class org/coinor/Ipopt.class, which you will need to make available
in your Java code (i.e., add $IPOPTDIR/build/Ipopt/contrib/JavaInterface to your CLASSPATH) and
the shared object lib/libjipopt.so (on Linux/UNIX) or lib/libjipopt.dylib (on Mac OS X) or the
DLL lib/jipopt.dll (on Windows). In order to test your JIpopt library you can run two example
problems by issuing the command make test inside the JIpopt directory.
NOTE: The JIpopt build procedure currently cannot deal with spaces in the path to the JDK. If
you are on Windows and have Java in a path like C:\Program Files\Java, try setting JAVA HOME to the
DOS equivalent C:\Progra~1 (or similar).
NOTE: JIpopt needs to be able to load the Ipopt library dynamically at runtime. Therefore, Ipopt
must have been compiled with the -fPIC compiler flag. While per default, an Ipopt shared library is
compiled with this flag, for a configuration of Ipopt in debug mode (--enable-debug) or as static library
(--disable-shared), the configure flag --with-pic need to be used to enable compilation with -fPIC.
12 VRTech

Industrial Technologies

2.7

Compiling and Installing the R Interface ipoptr

The ipoptr interface can be build after Ipopt has been build and installed. In the best case, it is
sufficient to execute the following command in R:
install.packages(’$IPOPTDIR/build/Ipopt/contrib/RInterface’, repos=NULL, type=’source’)
In certain situations, however, it can be necessary to setup the dynamic library load path to the path
where the Ipopt library has been installed, e.g.,
LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$IPOPTDIR/build/lib
NOTE: R needs to be able to load the Ipopt library dynamically at runtime. Therefore, Ipopt
must have been compiled with the -fPIC compiler flag. While per default, an Ipopt shared library is
compiled with this flag, for a configuration of Ipopt in debug mode (--enable-debug) or as static library
(--disable-shared), the configure flag --with- pic need to be used to enable compilation with -fPIC.
After installation of the ipoptr package, it should be possible to load the package in R and to view
the help page:
> library(’ipoptr’)
> ?ipoptr

2.8

Compiling and Installing the MATLAB interface
based on documentation by Peter Carbonetto13 , Tony Kelman14 , and Ray Zimmerman

The MATLAB interface to Ipopt uses the mex interface of MATLAB. It has been tested on MATLAB versions 7.2 through 7.7. It might very well work on earlier versions of MATLAB, but there is also
a good chance that it will not. It is unlikely that the software will run with versions prior to MATLAB
6.5.
NOTE: The MATLAB interface does not support MATLAB 8.3 (aka, R2014a), as it has not
been adapted to MATLAB changes in the buildsystem for MEX files.
First, note that some binaries of Ipopt mex files are available for download at http://www.coin-or.
org/download/binary/Ipopt. Further, the OPTI Toolbox (http://www.i2c2.aut.ac.nz/Wiki/OPTI)
comes with a precompiled MATLAB interface for Ipopt on Windows.
2.8.1

Setting up mex

To build the interface by yourself, you will need to have MATLAB installed on your system and have
it configured to build mex files, see http://www.mathworks.com/support/tech-notes/1600/1605.html
for detail on how to set this up.
Ipopt 3.11 has added Makefile options to automate fixes for commonly-encountered issues with
building the MATLAB interface. On Mac OS X or Windows, the file mexopts.sh (mexopts.bat on
Windows) will need to be modified. This is performed automatically by calling make mexopts in the
$IPOPTDIR/build/Ipopt/contrib/MatlabInterface/src directory. No changes will be made if you already have a mexopts.sh file in that directory. If you need to make these modifications manually, follow
the steps below.
For Mac OS X, the following procedure has been reported: First, one executes a command like
/Applications/MATLAB_R2012.app/bin/mex -setup
This creates a mexopts.sh file in the /.matlab/R2010 directory. Copy that file to the directory
$IPOPTDIR/build/Ipopt/contrib/MatlabInterface/src and modify it as follows.
13 University
14 University

of British Columbia
of California, Berkeley

• In the maci section (32 bit builds) or the maci64 section (64 bit builds), change both instances of
libgfortran.dylib to libgfortran.a in the FC LIBDIR line (in case your Fortran compiler only
comes with static libraries).
• Remove all occurrences of -isysroot $SDKROOT or -Wl,-syslibroot,$SDKROOT in case the hardcoded version of the Xcode SDK that Matlab expects does not match what you have installed on
your system.
• Remove all occurrences of -arch $ARCHS in case you are using a GNU compiler that does not
recognize these Apple-specific flags.
On Windows, if you are using the GNU compilers via MinGW, then you will need to use the gnumex
project. First, execute the script ./get.Gnumex from the $IPOPTDIR/Ipopt/contrib/MatlabInterface
directory. Then, after configuring Ipopt, go to $IPOPTDIR/build/contrib/MatlabInterface/src and
execute make gnumex. This will start an instance of MATLAB and open the gnumex tool. Check that
the options are filled out appropriately for your MinGW installation, click “Make options file,” then close
this new instance of MATLAB after gnumex has created mexopts.bat.
Calling make mexopts will automatically make the necessary changes to this new mexopts.bat file.
If you would like to do so manually, the changes are as follows.
• Change COMPILER=gcc to COMPILER=g++
• Change GM MEXLANG=c to GM MEXLANG=cxx
• Add the contents of the LIBS= line from the MATLAB interface Makefile to GM ADD LIBS
• If you want to statically link the standard libraries into the mex file, add -static to GM ADD LIBS
2.8.2

Adjusting configuration and build of Ipopt

The configure script of Ipopt attempts to automatically locate the directory where MATLAB is installed
by querying the location of the matlab executable. You can also manually specify the MATLAB home
directory when calling the configure script with the flag --with-matlab-home. You can determine this
home directory by the command matlabroot within MATLAB.
In practice, it has been found easier to install and use the MATLAB interface by disabling compilation
of the shared libraries, and use only static libraries instead. However, these static libraries need to be
built in a way that allow using them in a shared library, i.e., they need to build with position-independent
code. This is achieved with the configure script flags
--disable-shared --with-pic
On Mac OS X, it has been reported that additionally the following flags for configure should be used:
ADD_CFLAGS="-fno-common -fexceptions -no-cpp-precomp"
ADD_CXXFLAGS="-fno-common -fexceptions -no-cpp-precomp"
ADD_FFLAGS="-fexceptions -fbackslash"
With Ipopt 3.11, a site script for configure has been added to the MATLAB interface. This
script takes care of setting configure options in a way that is appropriate for building an Ipopt mex
file that is useable via MATLAB. Therefore, instead of setting configure options as described in the
previous section, it should be sufficient to create a directory $IPOPTDIR/build/share, copy the site
file $IPOPTDIR/contrib/MatlabInterface/MatlabInterface.site to that directory, and rename it to
config.site before running configure. Alternatively, you can set an environment variable CONFIG SITE
that points to the site file.
This site script sets the configure flags (if not specified by the user) --disable-shared --with-pic
--with-blas=BUILD --with-lapack=BUILD. The first two flags are discussed above. We also specify

that the reference versions of BLAS and LAPACK should be used by default because of a commonly
observed issue on 64-bit Linux systems. If Ipopt configure finds BLAS and/or LAPACK libraries already
installed then it will use them. However, MATLAB includes its own versions of BLAS and LAPACK,
which on 64-bit systems are incompatible with the expected interface used by Ipopt and the BLAS and
LAPACK packages available in many Linux distributions. If the Ipopt mex file is compiled in such a way
that the BLAS and LAPACK libraries are dynamically linked as shared libraries (as found in installed
Linux packages), those library dependencies will be overridden by MATLAB’s incompatible versions.
This can be avoided by statically linking BLAS and LAPACK into the Ipopt mex file, which the above
combination of configure flags will do. Note, that this issue does not appear to affect Mac OS X versions
of MATLAB, so if you would like to use the Apple optimized BLAS and LAPACK libraries you can
override these settings and specify --with-blas=’-framework vecLib’ --with-lapack=’-framework
vecLib’.
The site script also tests whether the compilers on your system are capable of statically linking the
standard C++ and Fortran libraries into a shared library. This is possible with GCC versions 4.5.0 or
newer on Mac OS X or Windows, 4.7.3 or newer (when GCC itself is built --with-pic) on Linux. If
this is the case, then the site script will set appropriate configure flags and options in the MATLAB
interface Makefile to statically link all standard libraries into the Ipopt mex file. This should allow a
single mex file to work with a variety of versions of MATLAB, and on computers that do not have the
same compiler versions installed. If this static linking of standard libraries causes any issues, you can
disable it with the configure flag --disable-matlab-static.
2.8.3

Building the MATLAB interface

After configuring, building, and installing Ipopt itself, it is time to build the MATLAB interface.
For that, Ipopt’s configure has setup a directory $IPOPTDIR/build/contrib/MatlabInterface/src
which contains a Makefile. You may need to edit this file a little bit to suit for your system setup.
You will find that most of the variables such as CXX and CXXFLAGS have been automatically (and
hopefully, correctly) set according to the flags specified during your initial call to configure script.
However, you may need to modify MATLAB HOME, MEXSUFFIX and MEX as explained in the comments
of the Makefile. For example, on Mac OS X, it has been reported that all duplicates of strings like
-L/usr/lib/gcc/i686-apple-darwin11/4.2.1/../../.. should be removed from the LIBS line.
Once you think you’ve set up the Makefile properly, type make install in the same directory as the
Makefile. If you didn’t get any errors, then you should have ended up with a mex file. The mex file will
be called ipopt.$MEXEXT, where $MEXEXT is mexglx for 32-bit Linux, mexa64 for 64-bit Linux, mexw32
for 32-bit Windows, etc.
2.8.4

Making MATLAB aware of the mex file

In order to use the mex file in MATLAB, you need to tell MATLAB where to find it. The best way to
do this is to type
addpath sourcedir
in the MATLAB command prompt, where sourcedir is the location of the mex file you created. (For
more information, type help addpath in MATLAB. You can also achieve the same thing by modifying
the MATLABPATH environment variable in the UNIX command line, using either the export command (in
Bash shell) or the setenv command (in C-shell).
2.8.5

Additional notes

Starting with version 7.3, MATLAB can handle 64-bit addressing, and the authors of MATLAB have
modified the implementation of sparse matrices to reflect this change. However, the row and column
indices in the sparse matrix are converted to signed integers, and this could potentially cause problems
when dealing with large, sparse matrices on 64-bit platforms with MATLAB version 7.3 or greater.
20

As MATLAB (version R2008a or newer) includes its own HSL MA57 library, Ipopt’s configure can
be setup to enable using this library in Ipopt’s MA57 solver interface. To enable this, one should specify
the configure option --enable-matlab-ma57. Note, that using this option is not advisable if one also
provides the source for MA57 via ThirdParty/HSL.
2.8.6

Troubleshooting

The installation procedure described above does involve a certain amount of expertise on the part of the
user. If you are encountering problems, it is highly recommended that you follow the standard installation
of Ipopt first, and then test the installation by running some examples, either in C++ or in AMPL.
What follows are a list of common errors encountered, along with a suggested remedy.
Problem: compilation is successful, but MATLAB crashes
Remedy: Even if you didn’t get any errors during compilation, there’s still a possibility that you
didn’t link the mex file properly. In this case, executing Ipopt in MATLAB will cause MATLAB to
crash. This is a common problem, and usually arises because you did not choose the proper compiler
or set the proper compilation flags (e.g. --with-pic) when you ran the configure script at the very
beginning.
Problem: MATLAB fails to link to Ipopt shared library
Remedy: You might encounter this problem if you try to execute one of the examples in MATLAB,
and MATLAB complains that it cannot find the Ipopt shared library. The installation script has been
set up so that the mex file you are calling knows where to look for the Ipopt shared library. However, if
you moved the library then you will run into a problem. One way to fix this problem is to modify the
LDFLAGS variable in the MATLAB interface Makefile (see above) so that the correct path of the Ipopt
library is specified. Alternatively, you could modify the LD LIBRARY PATH environment variable so that
the location of the Ipopt library is included in the path. If none of this is familiar to you, you might
want to do a web search to find out more.
Problem: mwIndex is not defined
Remedy: You may get a compilation error that says something to the effect that mwIndex is not
defined. This error will surface on a version of MATLAB prior to 7.3. The solution is to add the flag
-DMWINDEXISINT to the CXXFLAGS variable in the MATLAB interface Makefile (see above).

2.9

Expert Installation Options for Ipopt

The configuration script and Makefiles in the Ipopt distribution have been created using GNU’s autoconf
and automake. They attempt to automatically adapt the compiler settings etc. to the system they are
running on. We tested the provided scripts for a number of different machines, operating systems and
compilers, but you might run into a situation where the default setting does not work, or where you need
to change the settings to fit your particular environment.
In general, you can see the list of options and variables that can be set for the configure script by
typing configure --help. Also, the generic COIN-OR help pages are a valuable resource of information:
https://projects.coin-or.org/CoinHelp
Below a few particular options are discussed:
• The configure script tries to determine automatically, if you have BLAS and/or LAPACK already
installed on your system (trying a few default libraries), and if it does not find them, it makes sure
that you put the source code in the required place.
However, you can specify a BLAS library (such as your local ATLAS library15 ) explicitly, using the
--with-blas flag for configure. For example,
./configure --with-blas="-L$HOME/lib -lf77blas -lcblas -latlas"
15 see

http://math-atlas.sourceforge.net

To tell the configure script to compile and use the downloaded BLAS source files even if a BLAS
library is found on your system, specify --with-blas=BUILD.
Similarly, you can use the --with-lapack switch to specify the location of your LAPACK library,
or use the keyword BUILD to force the Ipopt makefiles to compile LAPACK together with Ipopt.
• Similarly, if you have a precompiled library containing the HSL packages, you can specify the
directory with the CoinHslConfig.h header file with the --with-hsl-incdir flag and the linker
flags with the --with-hsl-lib flag. Analogously, use --with-asl-incdir and --with-asl-lib
for building against a precompiled AMPL solver library.
• The HSL codes HSL MA86 and HSL MA97 can run in parallel if compiled with OpenMP support.
By default, this is not enabled by Ipopt’s configure so far. To enable OpenMP with GNU compilers,
it has been reported that the following configure flags should be used:
ADD_CFLAGS=-fopenmp ADD_FFLAGS=-fopenmp ADD_CXXFLAGS=-fopenmp
• If you want to compile Ipopt with the linear solver Pardiso (see Section 2.3.4) from the Pardiso
project website, you need to specify the link flags for the library with the --with-pardiso flag,
including required additional libraries and flags. For example, if you want to compile Ipopt with
the parallel version of Pardiso (located in $HOME/lib) on an AIX system in 64bit mode, you should
add the flag
--with-pardiso="-qsmp=omp $HOME/lib/libpardiso_P4AIX51_64_P.so"
If you are using the parallel version of Pardiso, you need to specify the number of processors it should
run on with the environment variable OMP_NUM_THREADS, as described in the Pardiso manual.
Experimental: If you want to compile Ipopt with the Pardiso library that is included in Intel MKL, you have to specify the MKL libs for the --with-pardiso configure flag and add
-DHAVE_PARDISO_MKL -DHAVE_PARDISO_PARALLEL to the compiler flags, e.g.,
--with-blas="-L${MKLDIR} -Wl,--start-group -lmkl_def -lmkl_lapack -lmkl_intel_lp64 \
-lmkl_intel_thread -lmkl_core -lippcoreem64t -Wl,--end-group -liomp5 -lpthread"
--with-pardiso="-L${MKLDIR} -Wl,--start-group -lmkl_def -lmkl_lapack -lmkl_intel_lp64 \
-lmkl_intel_thread -lmkl_core -lippcoreem64t -Wl,--end-group -liomp5 -lpthread"
ADD_CXXFLAGS="-DHAVE_PARDISO_MKL -DHAVE_PARDISO_PARALLEL"
• If you want to compile Ipopt with the linear solver WSMP (see Section 2.3.5), you need to specify
the link flags for the library with the --with-wsmp flag, including required additional libraries and
flags. For example, if you want to compile Ipopt with WSMP (located in $HOME/lib) on an Intel
IA32 Linux system, you should add the flag
--with-wsmp="$HOME/lib/wsmp/wsmp-Linux/lib/IA32/libwsmp.a -lpthread"
• If you want to compile Ipopt with a precompiled MUMPS library (see Section 2.3.3), you need to
specify the directory containing the MUMPS header files with the --with-mumps-incdir flag, e.g.,
--with-mumps-incdir="$HOME/MUMPS/include"
and you also need to provide the link flags for MUMPS with the --with-mumps-lib flag.
• If you want to specify that you want to use particular compilers, you can do so by adding the
variables definitions for CXX, CC, and F77 to the ./configure command line, to specify the C++,
C, and Fortran compiler, respectively. For example,
./configure CXX=g++-4.2.0 CC=gcc-4.2.0 F77=gfortran-4.2.0
In order to set the compiler flags, you should use the variables CXXFLAGS, CFLAGS, FFLAGS. Note,
that the Ipopt code uses “dynamic cast”. Therefore it is necessary that the C++ code is compiled

including RTTI (Run-Time Type Information). Some compilers need to be given special flags to
do that (e.g., “-qrtti=dyna” for the AIX xlC compiler).
Please also check the generic COIN-OR help page at
https://projects.coin-or.org/CoinHelp/wiki/user-configure#GivingOptions
for the description of more variables that can be set for configure.
• By default, the Ipopt library is compiled as a shared library, on systems where this is supported.
If you want to generate a static library, you need to specify the --disable-shared flag. If you
want to compile both shared and static libraries, you should specify the --enable-static flag.
• If you want to link the Ipopt library with a main program written in C or Fortran, the C and
Fortran compiler doing the linking of the executable needs to be told about the C++ runtime
libraries. Unfortunately, the current version of autoconf does not provide the automatic detection
of those libraries. We have hard-coded some default values for some systems and compilers, but
this might not work all the time.
If you have problems linking your Fortran or C code with the Ipopt library libipopt.a and
the linker complains about missing symbols from C++ (e.g., the standard template library), you
should specify the C++ libraries with the CXXLIBS variable. To find out what those libraries are,
it is probably helpful to link a simple C++ program with verbose compiler output.
For example, for the Intel compilers on a Linux system, you might need to specify something like
./configure CC=icc F77=ifort CXX=icpc \
CXXLIBS=’-L/usr/lib/gcc-lib/i386-redhat-linux/3.2.3 -lstdc++’
• Compilation in 64bit mode sometimes requires some special consideration. For example, for compilation of 64bit code on AIX, we recommend the following configuration
./configure AR=’ar -X64’ NM=’nm -X64’ \
CC=’xlc -q64’ F77=’xlf -q64’ CXX=’xlC -q64’\
CFLAGS=’-O3 -bmaxdata:0x3f0000000’ \
FFLAGS=’-O3 -bmaxdata:0x3f0000000’ \
CXXFLAGS=’-qrtti=dyna -O3 -bmaxdata:0x3f0000000’
(Alternatively, a simpler solution for AIX is to set the environment variable OBJECT MODE to 64.)
• It is possible to compile the Ipopt library in a debug configuration, by specifying --enable-debug.
Then the compilers will use the debug flags (unless the compilation flag variables are overwritten
in the configure command line)
Also, you can tell Ipopt to do some additional runtime sanity checks, by specifying the flag
--with-ipopt-checklevel=1.
This usually leads to a significant slowdown of the code, but might be helpful when debugging
something.

Interfacing your NLP to Ipopt

Ipopt has been designed to be flexible for a wide variety of applications, and there are a number of ways
to interface with Ipopt that allow specific data structures and linear solver techniques. Nevertheless, the
authors have included a standard representation that should meet the needs of most users.
This tutorial will discuss six interfaces to Ipopt, namely the AMPL modeling language [3] interface,
and the C++, C, Fortran, Java, and R code interfaces. AMPL is a 3rd party modeling language tool that
allows users to write their optimization problem in a syntax that resembles the way the problem would
be written mathematically. Once the problem has been formulated in AMPL, the problem can be easily

solved using the (already compiled) Ipopt AMPL solver executable, ipopt. Interfacing your problem by
directly linking code requires more effort to write, but can be far more efficient for large problems.
We will illustrate how to use each of the four interfaces using an example problem, number 71 from
the Hock-Schittkowsky test suite [5],
min

x∈<4

s.t.

x1 x4 (x1 + x2 + x3 ) + x3

(4)

x1 x2 x3 x4 ≥ 25

(5)

x21

x22

x23

x24

= 40

(6)

1 ≤ x1 , x2 , x3 , x4 ≤ 5,

(7)

x0 = (1, 5, 5, 1)

(8)

with the starting point
and the optimal solution
x∗ = (1.00000000, 4.74299963, 3.82114998, 1.37940829).
You can find further, less documented examples for using Ipopt from your own source code in the
Ipopt/examples subdirectory.

3.1

Using Ipopt through AMPL

Using the AMPL solver executable is by far the easiest way to solve a problem with Ipopt. The user must
simply formulate the problem in AMPL syntax, and solve the problem through the AMPL environment.
There are drawbacks, however. AMPL is a 3rd party package and, as such, must be appropriately
licensed (a free student version for limited problem size is available from the AMPL website, http:
//www.ampl.com). Furthermore, the AMPL environment may be prohibitive for very large problems.
Nevertheless, formulating the problem in AMPL is straightforward and even for large problems, it is
often used as a prototyping tool before using one of the code interfaces.
This tutorial is not intended as a guide to formulating models in AMPL. If you are not already familiar
with AMPL, please consult [3].
The problem presented in equations (4)–(8) can be solved with Ipopt with the following AMPL
model.
# tell ampl to use the ipopt executable as a solver
# make sure ipopt is in the path!
option solver ipopt;
# declare the variables and their bounds,
# set notation could be used, but this is straightforward
var x1 >= 1, <= 5;
var x2 >= 1, <= 5;
var x3 >= 1, <= 5;
var x4 >= 1, <= 5;
# specify the objective function
minimize obj:
x1 * x4 * (x1 + x2 + x3) + x3;
# specify the constraints
s.t.
inequality:
x1 * x2 * x3 * x4 >= 25;
24

equality:
x1^2 + x2^2 + x3^2 +x4^2 = 40;
# specify
let x1 :=
let x2 :=
let x3 :=
let x4 :=

the starting point
1;
5;
5;
1;

# solve the problem
solve;
# print
display
display
display
display

the solution
x1;
x2;
x3;
x4;

The line, “option solver ipopt;” tells AMPL to use Ipopt as the solver. The Ipopt executable
(installed in Section 2.4) must be in the PATH for AMPL to find it. The remaining lines specify the
problem in AMPL format. The problem can now be solved by starting AMPL and loading the mod file:
$ ampl
> model hs071_ampl.mod;
.
.
.
The problem will be solved using Ipopt and the solution will be displayed.
At this point, AMPL users may wish to skip the sections about interfacing with code, but should read
Section 5 concerning Ipopt options, and Section 6 which explains the output displayed by Ipopt.
3.1.1

Using Ipopt from the command line

It is possible to solve AMPL problems with Ipopt directly from the command line. However, this requires
a file in format .nl produced by ampl. If you have a model and data loaded in Ampl, you can create the
corresponding .nl file with name, say, myprob.nl by using the Ampl command:
write gmyprob
There is a small .nl file available in the Ipopt distribution. It is located at Ipopt/test/mytoy.nl.
We use this file in the remainder of this section. We assume that the file mytoy.nl is in the current
directory and that the command ipopt is a shortcut for running the ipopt binary available in the bin
directory of the installation of Ipopt.
We list below commands to perform basic tasks from the Linux prompt.
• To solve mytoy.nl from the Linux prompt, use:
ipopt mytoy
• To see all command line options for Ipopt, use:
ipopt -=
• To see more detailed information on all options for Ipopt:
ipopt mytoy ’print options documentation yes’
25

• To run ipopt, setting the maximum number of iterations to 2 and print level to 4:
ipopt mytoy ’max iter 2 print level 4’
If many options are to be set, they can be collected in a file ipopt.opt that is automatically read by
Ipopt if present in the current directory, see Section 5.

3.2

Interfacing with Ipopt through code

In order to solve a problem, Ipopt needs more information than just the problem definition (for example,
the derivative information). If you are using a modeling language like AMPL, the extra information is
provided by the modeling tool and the Ipopt interface. When interfacing with Ipopt through your own
code, however, you must provide this additional information. The following information is required by
Ipopt:
1. Problem dimensions
• number of variables
• number of constraints
2. Problem bounds
• variable bounds
• constraint bounds
3. Initial starting point
• Initial values for the primal x variables
• Initial values for the multipliers (only required for a warm start option)
4. Problem Structure
• number of nonzeros in the Jacobian of the constraints
• number of nonzeros in the Hessian of the Lagrangian function
• sparsity structure of the Jacobian of the constraints
• sparsity structure of the Hessian of the Lagrangian function
5. Evaluation of Problem Functions
Information evaluated using a given point (x, λ, σf coming from Ipopt)
• Objective function, f (x)
• Gradient of the objective ∇f (x)
• Constraint function values, g(x)
• Jacobian of the constraints, ∇g(x)T
Pm
• Hessian of the Lagrangian function, σf ∇2 f (x) + i=1 λi ∇2 gi (x)
(this is not required if a quasi-Newton options is chosen to approximate the second derivatives)
The problem dimensions and bounds are straightforward and come solely from the problem definition.
The initial starting point is used by the algorithm when it begins iterating to solve the problem. If Ipopt
has difficulty converging, or if it converges to a locally infeasible point, adjusting the starting point may
help. Depending on the starting point, Ipopt may also converge to different local solutions.
Providing the sparsity structure of derivative matrices is a bit more involved. Ipopt is a nonlinear
programming solver that is designed for solving large-scale, sparse problems. While Ipopt can be customized for a variety of matrix formats, the triplet format is used for the standard interfaces in this
26

tutorial. For an overview of the triplet format for sparse matrices, see Appendix A. Before solving the
problem, Ipopt needs to know the number of nonzero elements and the sparsity structure (row and
column indices of each of the nonzero entries) of the constraint Jacobian and the Lagrangian function
Hessian. Once defined, this nonzero structure MUST remain constant for the entire optimization procedure. This means that the structure needs to include entries for any element that could ever be nonzero,
not only those that are nonzero at the starting point.
As Ipopt iterates, it will need the values for Item 5 in Section 3.2 evaluated at particular points.
Before we can begin coding the interface, however, we need to work out the details of these equations
symbolically for example problem (4)-(7).
The gradient of the objective f (x) is given by

x1 x4 + x4 (x1 + x2 + x3 )


x1 x4




x1 x4 + 1
x1 (x1 + x2 + x3 )


and the Jacobian of the constraints g(x) is

x2 x3 x4
2x1

x1 x3 x4
2x2

x1 x2 x4
2x3

x1 x2 x3
2x4

We also need to determine the Hessian of the Lagrangian16 . The Lagrangian function for the NLP
(4)-(7) is defined as f (x) + g(x)T λ and the Hessian of the Lagrangian function is, technically, ∇2 f (xk ) +
Pm
2
i=1 λi ∇ gi (xk ). However, we introduce a factor (σf ) in front of the objective term so that Ipopt can
ask for the Hessian of the objective or the constraints independently, if required. Thus, for Ipopt the
symbolic form of the Hessian of the Lagrangian is
σf ∇2 f (xk ) +

m
X

λi ∇2 gi (xk )

(9)

i=1

and for the example problem this becomes
2x4

x4
σf 

x4
2x1 + x2 + x3


x4
0
0
x1



2x1 + x2 + x3
0


x1
+λ1  x3 x4
 x2 x4

x1
0
x2 x3

x3 x4
0
x1 x4
x1 x3

x2 x4
x1 x4
0
x1 x2



x2 x3
2


0
x1 x3 
+λ2 
 0
x1 x2 
0
0

0
2
0
0

0
0
2
0


0
0 

0 
2

where the first term comes from the Hessian of the objective function, and the second and third term
from the Hessian of the constraints (5) and (6), respectively. Therefore, the dual variables λ1 and λ2 are
the multipliers for constraints (5) and (6), respectively.
The remaining sections of the tutorial will lead you through the coding required to solve example
problem (4)–(7) using, first C++, then C, and finally Fortran. Completed versions of these examples can
be found in $IPOPTDIR/Ipopt/examples under hs071 cpp, hs071 c, hs071 f.
As a user, you are responsible for coding two sections of the program that solves a problem using
Ipopt: the main executable (e.g., main) and the problem representation. Typically, you will write an
executable that prepares the problem, and then passes control over to Ipopt through an Optimize or
Solve call. In this call, you will give Ipopt everything that it requires to call back to your code whenever
it needs functions evaluated (like the objective function, the Jacobian of the constraints, etc.). In each
of the three sections that follow (C++, C, and Fortran), we will first discuss how to code the problem
representation, and then how to code the executable.
16 If a quasi-Newton option is chosen to approximate the second derivatives, this is not required. However, if second
derivatives can be computed, it is often worthwhile to let Ipopt use them, since the algorithm is then usually more robust
and converges faster. More on the quasi-Newton approximation in Section 4.2.

3.3

The C++ Interface

This tutorial assumes that you are familiar with the C++ programming language, however, we will lead
you through each step of the implementation. For the problem representation, we will create a class that
inherits off of the pure virtual base class, TNLP (IpTNLP.hpp). For the executable (the main function) we
will make the call to Ipopt through the IpoptApplication class (IpIpoptApplication.hpp). In addition, we will also be using the SmartPtr class (IpSmartPtr.hpp) which implements a reference counting
pointer that takes care of memory management (object deletion) for you (for details, see Appendix B).
After “make install” (see Section 2.4), the header files are installed in $IPOPTDIR/include/coin
(or in $PREFIX/include/coin if the switch --prefix=$PREFIX was used for configure).
3.3.1

Coding the Problem Representation

We provide the required information by coding the HS071 NLP class, a specific implementation of the
TNLP base class. In the executable, we will create an instance of the HS071 NLP class and give this class to
Ipopt so it can evaluate the problem functions through the TNLP interface. If you have any difficulty as
the implementation proceeds, have a look at the completed example in the Ipopt/examples/hs071 cpp
directory.
Start by creating a new directory MyExample under examples and create the files hs071 nlp.hpp and
hs071 nlp.cpp. In hs071 nlp.hpp, include IpTNLP.hpp (the base class), tell the compiler that we are
using the Ipopt namespace, and create the declaration of the HS071 NLP class, inheriting off of TNLP.
Have a look at the TNLP class in IpTNLP.hpp; you will see eight pure virtual methods that we must
implement. Declare these methods in the header file. Implement each of the methods in HS071 NLP.cpp
using the descriptions given below. In hs071 nlp.cpp, first include the header file for your class and tell
the compiler that you are using the Ipopt namespace. A full version of these files can be found in the
Ipopt/examples/hs071 cpp directory.
It is very easy to make mistakes in the implementation of the function evaluation methods, in particular
regarding the derivatives. Ipopt has a feature that can help you to debug the derivative code, using finite
differences, see Section 4.1.
Note that the return value of any bool-valued function should be true, unless an error occurred, for
example, because the value of a problem function could not be evaluated at the required point.
Method get nlp info with prototype
virtual bool get_nlp_info(Index& n, Index& m, Index& nnz_jac_g,
Index& nnz_h_lag, IndexStyleEnum& index_style)
Give Ipopt the information about the size of the problem (and hence, the size of the arrays that it needs
to allocate).
• n: (out), the number of variables in the problem (dimension of x).
• m: (out), the number of constraints in the problem (dimension of g(x)).
• nnz jac g: (out), the number of nonzero entries in the Jacobian.
• nnz h lag: (out), the number of nonzero entries in the Hessian.
• index style: (out), the numbering style used for row/col entries in the sparse matrix format
(C STYLE: 0-based, FORTRAN STYLE: 1-based; see also Appendix A).
Ipopt uses this information when allocating the arrays that it will later ask you to fill with values. Be
careful in this method since incorrect values will cause memory bugs which may be very difficult to find.
Our example problem has 4 variables (n), and 2 constraints (m). The constraint Jacobian for this small
problem is actually dense and has 8 nonzeros (we still need to represent this Jacobian using the sparse
matrix triplet format). The Hessian of the Lagrangian has 10 “symmetric” nonzeros (i.e., nonzeros in the
28

lower left triangular part.). Keep in mind that the number of nonzeros is the total number of elements
that may ever be nonzero, not just those that are nonzero at the starting point. This information is set
once for the entire problem.
bool HS071_NLP::get_nlp_info(Index& n, Index& m, Index& nnz_jac_g,
Index& nnz_h_lag, IndexStyleEnum& index_style)
{
// The problem described in HS071_NLP.hpp has 4 variables, x[0] through x[3]
n = 4;
// one equality constraint and one inequality constraint
m = 2;
// in this example the Jacobian is dense and contains 8 nonzeros
nnz_jac_g = 8;
// the Hessian is also dense and has 16 total nonzeros, but we
// only need the lower left corner (since it is symmetric)
nnz_h_lag = 10;
// use the C style indexing (0-based)
index_style = TNLP::C_STYLE;
return true;
}

Method get bounds info with prototype
virtual bool get_bounds_info(Index n, Number* x_l, Number* x_u,
Index m, Number* g_l, Number* g_u)
Give Ipopt the value of the bounds on the variables and constraints.
• n: (in), the number of variables in the problem (dimension of x).
• x l: (out) the lower bounds xL for x.
• x u: (out) the upper bounds xU for x.
• m: (in), the number of constraints in the problem (dimension of g(x)).
• g l: (out) the lower bounds g L for g(x).
• g u: (out) the upper bounds g U for g(x).
The values of n and m that you specified in get nlp info are passed to you for debug checking. Setting a
lower bound to a value less than or equal to the value of the option nlp lower bound inf will cause Ipopt
to assume no lower bound. Likewise, specifying the upper bound above or equal to the value of the option
nlp upper bound inf will cause Ipopt to assume no upper bound. These options, nlp lower bound inf
and nlp upper bound inf, are set to −1019 and 1019 , respectively, by default, but may be modified by
changing the options (see Section 5).
In our example, the first constraint has a lower bound of 25 and no upper bound, so we set the lower
bound of constraint [0] to 25 and the upper bound to some number greater than 1019 . The second
constraint is an equality constraint and we set both bounds to 40. Ipopt recognizes this as an equality
constraint and does not treat it as two inequalities.
bool HS071_NLP::get_bounds_info(Index n, Number* x_l, Number* x_u,
Index m, Number* g_l, Number* g_u)
{
// here, the n and m we gave IPOPT in get_nlp_info are passed back to us.
// If desired, we could assert to make sure they are what we think they are.

assert(n == 4);
assert(m == 2);
// the variables have lower bounds of 1
for (Index i=0; i<4; i++)
x_l[i] = 1.0;
// the variables have upper bounds of 5
for (Index i=0; i<4; i++)
x_u[i] = 5.0;
// the first constraint g1 has a lower bound of 25
g_l[0] = 25;
// the first constraint g1 has NO upper bound, here we set it to 2e19.
// Ipopt interprets any number greater than nlp_upper_bound_inf as
// infinity. The default value of nlp_upper_bound_inf and nlp_lower_bound_inf
// is 1e19 and can be changed through ipopt options.
g_u[0] = 2e19;
// the second constraint g2 is an equality constraint, so we set the
// upper and lower bound to the same value
g_l[1] = g_u[1] = 40.0;
return true;
}

Method get starting point with prototype
virtual bool get_starting_point(Index n, bool init_x, Number* x,
bool init_z, Number* z_L, Number* z_U,
Index m, bool init_lambda, Number* lambda)
Give Ipopt the starting point before it begins iterating.
• n: (in), the number of variables in the problem (dimension of x).
• init x: (in), if true, this method must provide an initial value for x.
• x: (out), the initial values for the primal variables, x.
• init z: (in), if true, this method must provide an initial value for the bound multipliers z L and
zU .
• z L: (out), the initial values for the bound multipliers, z L .
• z U: (out), the initial values for the bound multipliers, z U .
• m: (in), the number of constraints in the problem (dimension of g(x)).
• init lambda: (in), if true, this method must provide an initial value for the constraint multipliers,
λ.
• lambda: (out), the initial values for the constraint multipliers, λ.
The variables n and m are passed in for your convenience. These variables will have the same values
you specified in get nlp info.
Depending on the options that have been set, Ipopt may or may not require bounds for the primal
variables x, the bound multipliers z L and z U , and the constraint multipliers λ. The boolean flags init x,
init z, and init lambda tell you whether or not you should provide initial values for x, z L , z U , or λ
respectively. The default options only require an initial value for the primal variables x. Note, the initial
(i)
(i)
values for bound multiplier components for “infinity” bounds (xL = −∞ or xU = ∞) are ignored.
30

In our example, we provide initial values for x as specified in the example problem. We do not provide
any initial values for the dual variables, but use an assert to immediately let us know if we are ever asked
for them.
bool HS071_NLP::get_starting_point(Index n, bool init_x, Number* x,
bool init_z, Number* z_L, Number* z_U,
Index m, bool init_lambda,
Number* lambda)
{
// Here, we assume we only have starting values for x, if you code
// your own NLP, you can provide starting values for the dual variables
// if you wish to use a warmstart option
assert(init_x == true);
assert(init_z == false);
assert(init_lambda == false);
// initialize to the given starting point
x[0] = 1.0;
x[1] = 5.0;
x[2] = 5.0;
x[3] = 1.0;
return true;
}

Method eval f

with prototype

virtual bool eval_f(Index n, const Number* x,
bool new_x, Number& obj_value)
Return the value of the objective function at the point x.
• n: (in), the number of variables in the problem (dimension of x).
• x: (in), the values for the primal variables, x, at which f (x) is to be evaluated.
• new x: (in), false if any evaluation method was previously called with the same values in x, true
otherwise.
• obj value: (out) the value of the objective function (f (x)).
The boolean variable new x will be false if the last call to any of the evaluation methods (eval *) used
the same x values. This can be helpful when users have efficient implementations that calculate multiple
outputs at once. Ipopt internally caches results from the TNLP and generally, this flag can be ignored.
The variable n is passed in for your convenience. This variable will have the same value you specified
in get nlp info.
For our example, we ignore the new x flag and calculate the objective.
bool HS071_NLP::eval_f(Index n, const Number* x, bool new_x, Number& obj_value)
{
assert(n == 4);
obj_value = x[0] * x[3] * (x[0] + x[1] + x[2]) + x[2];
return true;
}

Method eval grad f

with prototype

virtual bool eval_grad_f(Index n, const Number* x, bool new_x,
Number* grad_f)
31

Return the gradient of the objective function at the point x.
• n: (in), the number of variables in the problem (dimension of x).
• x: (in), the values for the primal variables, x, at which ∇f (x) is to be evaluated.
• new x: (in), false if any evaluation method was previously called with the same values in x, true
otherwise.
• grad f: (out) the array of values for the gradient of the objective function (∇f (x)).
The gradient array is in the same order as the x variables (i.e., the gradient of the objective with
respect to x[2] should be put in grad f[2]).
The boolean variable new x will be false if the last call to any of the evaluation methods (eval *) used
the same x values. This can be helpful when users have efficient implementations that calculate multiple
outputs at once. Ipopt internally caches results from the TNLP and generally, this flag can be ignored.
The variable n is passed in for your convenience. This variable will have the same value you specified
in get nlp info.
In our example, we ignore the new x flag and calculate the values for the gradient of the objective.
bool HS071_NLP::eval_grad_f(Index n, const Number* x, bool new_x, Number* grad_f)
{
assert(n == 4);
grad_f[0]
grad_f[1]
grad_f[2]
grad_f[3]

=
=
=
=

x[0]
x[0]
x[0]
x[0]

*
*
*
*

x[3] + x[3] * (x[0] + x[1] + x[2]);
x[3];
x[3] + 1;
(x[0] + x[1] + x[2]);

return true;
}

Method eval g

with prototype

virtual bool eval_g(Index n, const Number* x,
bool new_x, Index m, Number* g)
Return the value of the constraint function at the point x.
• n: (in), the number of variables in the problem (dimension of x).
• x: (in), the values for the primal variables, x, at which the constraint functions, g(x), are to be
evaluated.
• new x: (in), false if any evaluation method was previously called with the same values in x, true
otherwise.
• m: (in), the number of constraints in the problem (dimension of g(x)).
• g: (out) the array of constraint function values, g(x).
The values returned in g should be only the g(x) values, do not add or subtract the bound values g L
or g U .
The boolean variable new x will be false if the last call to any of the evaluation methods (eval *) used
the same x values. This can be helpful when users have efficient implementations that calculate multiple
outputs at once. Ipopt internally caches results from the TNLP and generally, this flag can be ignored.
The variables n and m are passed in for your convenience. These variables will have the same values
you specified in get nlp info.
In our example, we ignore the new x flag and calculate the values of constraint functions.
32

bool HS071_NLP::eval_g(Index n, const Number* x, bool new_x, Index m, Number* g)
{
assert(n == 4);
assert(m == 2);
g[0] = x[0] * x[1] * x[2] * x[3];
g[1] = x[0]*x[0] + x[1]*x[1] + x[2]*x[2] + x[3]*x[3];
return true;
}

Method eval jac g

with prototype

virtual bool eval_jac_g(Index n, const Number* x, bool new_x,
Index m, Index nele_jac, Index* iRow,
Index *jCol, Number* values)
Return either the sparsity structure of the Jacobian of the constraints, or the values for the Jacobian of
the constraints at the point x.
• n: (in), the number of variables in the problem (dimension of x).
• x: (in), the values for the primal variables, x, at which the constraint Jacobian, ∇g(x)T , is to be
evaluated.
• new x: (in), false if any evaluation method was previously called with the same values in x, true
otherwise.
• m: (in), the number of constraints in the problem (dimension of g(x)).
• n ele jac: (in), the number of nonzero elements in the Jacobian (dimension of iRow, jCol, and
values).
• iRow: (out), the row indices of entries in the Jacobian of the constraints.
• jCol: (out), the column indices of entries in the Jacobian of the constraints.
• values: (out), the values of the entries in the Jacobian of the constraints.
The Jacobian is the matrix of derivatives where the derivative of constraint g (i) with respect to variable
x is placed in row i and column j. See Appendix A for a discussion of the sparse matrix format used
in this method.
If the iRow and jCol arguments are not NULL, then Ipopt wants you to fill in the sparsity structure of
the Jacobian (the row and column indices only). At this time, the x argument and the values argument
will be NULL.
If the x argument and the values argument are not NULL, then Ipopt wants you to fill in the values
of the Jacobian as calculated from the array x (using the same order as you used when specifying the
sparsity structure). At this time, the iRow and jCol arguments will be NULL;
The boolean variable new x will be false if the last call to any of the evaluation methods (eval *) used
the same x values. This can be helpful when users have efficient implementations that calculate multiple
outputs at once. Ipopt internally caches results from the TNLP and generally, this flag can be ignored.
The variables n, m, and nele jac are passed in for your convenience. These arguments will have the
same values you specified in get nlp info.
In our example, the Jacobian is actually dense, but we still specify it using the sparse format.
(j)

bool HS071_NLP::eval_jac_g(Index n, const Number* x, bool new_x,
Index m, Index nele_jac, Index* iRow, Index *jCol,
Number* values)
{

if (values == NULL) {
// return the structure of the Jacobian
// this
iRow[0]
iRow[1]
iRow[2]
iRow[3]
iRow[4]
iRow[5]
iRow[6]
iRow[7]

particular Jacobian is dense
= 0; jCol[0] = 0;
= 0; jCol[1] = 1;
= 0; jCol[2] = 2;
= 0; jCol[3] = 3;
= 1; jCol[4] = 0;
= 1; jCol[5] = 1;
= 1; jCol[6] = 2;
= 1; jCol[7] = 3;

}
else {
// return the values of the Jacobian of the constraints
values[0]
values[1]
values[2]
values[3]

=
=
=
=

x[1]*x[2]*x[3];
x[0]*x[2]*x[3];
x[0]*x[1]*x[3];
x[0]*x[1]*x[2];

values[4]
values[5]
values[6]
values[7]

=
=
=
=

2*x[0];
2*x[1];
2*x[2];
2*x[3];

//
//
//
//

0,0
0,1
0,2
0,3

1,0
1,1
1,2
1,3

}
return true;
}

Method eval h

with prototype

virtual bool eval_h(Index n, const Number* x, bool new_x,
Number obj_factor, Index m, const Number* lambda,
bool new_lambda, Index nele_hess, Index* iRow,
Index* jCol, Number* values)
Return either the sparsity structure of the Hessian of the Lagrangian, or the values of the Hessian of the
Lagrangian (9) for the given values for x, σf , and λ.
• n: (in), the number of variables in the problem (dimension of x).
• x: (in), the values for the primal variables, x, at which the Hessian is to be evaluated.
• new x: (in), false if any evaluation method was previously called with the same values in x, true
otherwise.
• obj factor: (in), factor in front of the objective term in the Hessian, σf .
• m: (in), the number of constraints in the problem (dimension of g(x)).
• lambda: (in), the values for the constraint multipliers, λ, at which the Hessian is to be evaluated.
• new lambda: (in), false if any evaluation method was previously called with the same values in
lambda, true otherwise.
• nele hess: (in), the number of nonzero elements in the Hessian (dimension of iRow, jCol, and
values).
• iRow: (out), the row indices of entries in the Hessian.
• jCol: (out), the column indices of entries in the Hessian.
34

• values: (out), the values of the entries in the Hessian.
The Hessian matrix that Ipopt uses is defined in (9). See Appendix A for a discussion of the sparse
symmetric matrix format used in this method.
If the iRow and jCol arguments are not NULL, then Ipopt wants you to fill in the sparsity structure
of the Hessian (the row and column indices for the lower or upper triangular part only). In this case, the
x, lambda, and values arrays will be NULL.
If the x, lambda, and values arrays are not NULL, then Ipopt wants you to fill in the values of the
Hessian as calculated using x and lambda (using the same order as you used when specifying the sparsity
structure). In this case, the iRow and jCol arguments will be NULL.
The boolean variables new x and new lambda will both be false if the last call to any of the evaluation
methods (eval *) used the same values. This can be helpful when users have efficient implementations
that calculate multiple outputs at once. Ipopt internally caches results from the TNLP and generally, this
flag can be ignored.
The variables n, m, and nele hess are passed in for your convenience. These arguments will have the
same values you specified in get nlp info.
In our example, the Hessian is dense, but we still specify it using the sparse matrix format. Because
the Hessian is symmetric, we only need to specify the lower left corner.
bool HS071_NLP::eval_h(Index n, const Number* x, bool new_x,
Number obj_factor, Index m, const Number* lambda,
bool new_lambda, Index nele_hess, Index* iRow,
Index* jCol, Number* values)
{
if (values == NULL) {
// return the structure. This is a symmetric matrix, fill the lower left
// triangle only.
// the Hessian for this problem is actually dense
Index idx=0;
for (Index row = 0; row < 4; row++) {
for (Index col = 0; col <= row; col++) {
iRow[idx] = row;
jCol[idx] = col;
idx++;
}
}
assert(idx == nele_hess);
}
else {
// return the values. This is a symmetric matrix, fill the lower left
// triangle only
// fill the objective portion
values[0] = obj_factor * (2*x[3]); // 0,0
values[1] = obj_factor * (x[3]);
values[2] = 0;

// 1,0
// 1,1

values[3] = obj_factor * (x[3]);
values[4] = 0;
values[5] = 0;

// 2,0
// 2,1
// 2,2

values[6]
values[7]
values[8]
values[9]

=
=
=
=

obj_factor * (2*x[0] + x[1] + x[2]); // 3,0
obj_factor * (x[0]);
// 3,1
obj_factor * (x[0]);
// 3,2
0;
// 3,3

// add the portion for the first constraint

values[1] += lambda[0] * (x[2] * x[3]); // 1,0
values[3] += lambda[0] * (x[1] * x[3]); // 2,0
values[4] += lambda[0] * (x[0] * x[3]); // 2,1
values[6] += lambda[0] * (x[1] * x[2]); // 3,0
values[7] += lambda[0] * (x[0] * x[2]); // 3,1
values[8] += lambda[0] * (x[0] * x[1]); // 3,2
// add the portion for the second constraint
values[0] += lambda[1] * 2; // 0,0
values[2] += lambda[1] * 2; // 1,1
values[5] += lambda[1] * 2; // 2,2
values[9] += lambda[1] * 2; // 3,3
}
return true;
}

Method finalize solution with prototype
virtual void finalize_solution(SolverReturn status, Index n,
const Number* x, const Number* z_L,
const Number* z_U, Index m, const Number* g,
const Number* lambda, Number obj_value,
const IpoptData* ip_data,
IpoptCalculatedQuantities* ip_cq)
This is the only method that is not mentioned in Section 3.2. This method is called by Ipopt after the
algorithm has finished (successfully or even with most errors).
• status: (in), gives the status of the algorithm as specified in IpAlgTypes.hpp,
– SUCCESS: Algorithm terminated successfully at a locally optimal point, satisfying the convergence tolerances (can be specified by options).
– MAXITER EXCEEDED: Maximum number of iterations exceeded (can be specified by an option).
– CPUTIME EXCEEDED: Maximum number of CPU seconds exceeded (can be specified by an option).
– STOP AT TINY STEP: Algorithm proceeds with very little progress.
– STOP AT ACCEPTABLE POINT: Algorithm stopped at a point that was converged, not to “desired” tolerances, but to “acceptable” tolerances (see the acceptable-... options).
– LOCAL INFEASIBILITY: Algorithm converged to a point of local infeasibility. Problem may be
infeasible.
– USER REQUESTED STOP: The user call-back function intermediate callback (see Section 3.3.4)
returned false, i.e., the user code requested a premature termination of the optimization.
– DIVERGING ITERATES: It seems that the iterates diverge.
– RESTORATION FAILURE: Restoration phase failed, algorithm doesn’t know how to proceed.
– ERROR IN STEP COMPUTATION: An unrecoverable error occurred while Ipopt tried to compute
the search direction.
– INVALID NUMBER DETECTED: Algorithm received an invalid number (such as NaN or Inf) from
the NLP; see also option check derivatives for naninf.
36

– INTERNAL ERROR: An unknown internal error occurred. Please contact the Ipopt authors
through the mailing list.
• n: (in), the number of variables in the problem (dimension of x).
• x: (in), the final values for the primal variables, x∗ .
• z L: (in), the final values for the lower bound multipliers, z∗L .
• z U: (in), the final values for the upper bound multipliers, z∗U .
• m: (in), the number of constraints in the problem (dimension of g(x)).
• g: (in), the final value of the constraint function values, g(x∗ ).
• lambda: (in), the final values of the constraint multipliers, λ∗ .
• obj value: (in), the final value of the objective, f (x∗ ).
• ip data and ip cq are provided for expert users.
This method gives you the return status of the algorithm (SolverReturn), and the values of the
variables, the objective and constraint function values when the algorithm exited.
In our example, we will print the values of some of the variables to the screen.
void HS071_NLP::finalize_solution(SolverReturn status,
Index n, const Number* x, const Number* z_L,
const Number* z_U, Index m, const Number* g,
const Number* lambda, Number obj_value,
const IpoptData* ip_data, IpoptCalculatedQuantities* ip_cq)
{
// here is where we would store the solution to variables, or write to a file, etc
// so we could use the solution.
// For this example, we write the solution to the console
printf("\n\nSolution of the primal variables, x\n");
for (Index i=0; i mynlp = new HS071_NLP();
// Create a new instance of IpoptApplication
// (use a SmartPtr, not raw)
// We are using the factory, since this allows us to compile this
// example with an Ipopt Windows DLL
SmartPtr app = IpoptApplicationFactory();
// Change some options
// Note: The following choices are only examples, they might not be
//
suitable for your optimization problem.
app->Options()->SetNumericValue("tol", 1e-9);
app->Options()->SetStringValue("mu_strategy", "adaptive");
app->Options()->SetStringValue("output_file", "ipopt.out");
// Intialize the IpoptApplication and process the options
ApplicationReturnStatus status;
status = app->Initialize();
if (status != Solve_Succeeded) {
printf("\n\n*** Error during initialization!\n");
return (int) status;
}
// Ask Ipopt to solve the problem
status = app->OptimizeTNLP(mynlp);
if (status == Solve_Succeeded) {
printf("\n\n*** The problem solved!\n");
}
else {
printf("\n\n*** The problem FAILED!\n");
}
// As the SmartPtrs go out of scope, the reference count
// will be decremented and the objects will automatically
// be deleted.
return (int) status;
}

The first line of code in main creates an instance of HS071 NLP. We then create an instance of the
Ipopt solver, IpoptApplication. You could use new to create a new application object, but if you want
to make sure that your code would also work with a Windows DLL, you need to use the factory, as done
in the example above. The call to app->Initialize(...) will initialize that object and process the
options (particularly the output related options). The call to app->OptimizeTNLP(...) will run Ipopt
and try to solve the problem. By default, Ipopt will write its progress to the console, and return the
SolverReturn status.

3.3.3

Compiling and Testing the Example

Our next task is to compile and test the code. If you are familiar with the compiler and linker used
on your system, you can build the code, telling the linker about the necessary libraries, as listed in the
ipopt addlibs cpp.txt file. If you are using the autotools based build system, then a sample makefile
created by configure already exists. Copy Ipopt/examples/hs071 cpp/Makefile into your MyExample
directory. This makefile was created for the hs071 cpp code, but it can be easily modified for your
example problem. Edit the file, making the following changes,
• change the EXE variable
EXE = my example
• change the OBJS variable
OBJS = HS071 NLP.o MyExample.o
and the problem should compile easily with,
$ make
Now run the executable,
$ ./my example
and you should see output resembling the following,
******************************************************************************
This program contains Ipopt, a library for large-scale nonlinear optimization.
Ipopt is released as open source code under the Eclipse Public License (EPL).
For more information visit http://projects.coin-or.org/Ipopt
******************************************************************************
Number of nonzeros in equality constraint Jacobian...:
Number of nonzeros in inequality constraint Jacobian.:
Number of nonzeros in Lagrangian Hessian.............:

4
4
10

Total number of variables............................:
variables with only lower bounds:
variables with lower and upper bounds:
variables with only upper bounds:
Total number of equality constraints.................:
Total number of inequality constraints...............:
inequality constraints with only lower bounds:
inequality constraints with lower and upper bounds:
inequality constraints with only upper bounds:

4
0
4
0
1
1
1
0
0

iter
0
1
2
3
4
5
6
7
8

objective
1.6109693e+01
1.7410406e+01
1.8001613e+01
1.7199482e+01
1.6940955e+01
1.7003411e+01
1.7013974e+01
1.7014017e+01
1.7014017e+01

inf_pr
1.12e+01
8.38e-01
1.06e-02
9.04e-02
2.09e-01
2.29e-02
2.59e-04
2.26e-07
4.62e-14

inf_du lg(mu)
5.28e-01
0.0
2.25e+01 -0.3
4.96e+00 -0.3
4.24e-01 -1.0
4.58e-02 -1.4
8.42e-03 -2.9
8.65e-05 -4.5
5.71e-08 -8.0
9.09e-14 -8.0

||d|| lg(rg)
0.00e+00
7.97e-01
5.60e-02
2.0
9.91e-01
2.88e-01
7.03e-02
6.22e-03
1.43e-04
6.95e-08
-

alpha_du
0.00e+00
3.19e-01
9.97e-01
9.98e-01
9.66e-01
9.68e-01
1.00e+00
1.00e-00
1.00e+00

Number of Iterations....: 8
Number of
Number of
Number of
Number of
Number of
Number of
Number of
Total CPU

objective function evaluations
objective gradient evaluations
equality constraint evaluations
inequality constraint evaluations
equality constraint Jacobian evaluations
inequality constraint Jacobian evaluations
Lagrangian Hessian evaluations
secs in IPOPT (w/o function evaluations)

=
=
=
=
=
=
=
=

9
9
9
9
9
9
8
0.220

alpha_pr ls
0.00e+00
0
1.00e+00f 1
1.00e+00h 1
1.00e+00f 1
1.00e+00h 1
1.00e+00h 1
1.00e+00h 1
1.00e+00h 1
1.00e+00h 1

Total CPU secs in NLP function evaluations

0.000

EXIT: Optimal Solution Found.

Solution of the primal variables, x
x[0] = 1.000000e+00
x[1] = 4.743000e+00
x[2] = 3.821150e+00
x[3] = 1.379408e+00

Solution
z_L[0] =
z_L[1] =
z_L[2] =
z_L[3] =
z_U[0] =
z_U[1] =
z_U[2] =
z_U[3] =

of the bound multipliers, z_L and z_U
1.087871e+00
2.428776e-09
3.222413e-09
2.396076e-08
2.272727e-09
3.537314e-08
7.711676e-09
2.510890e-09

Objective value
f(x*) = 1.701402e+01

*** The problem solved!

This completes the basic C++ tutorial, but see Section 6 which explains the standard console output
of Ipopt and Section 5 for information about the use of options to customize the behavior of Ipopt.
The Ipopt/examples/ScalableProblems directory contains other NLP problems coded in C++.
3.3.4

Additional methods in TNLP

The following methods are available to additional features that are not explained in the example. Default
implementations for those methods are provided, so that a user can safely ignore them, unless she wants
to make use of those features. From these features, only the intermediate callback is already available in
the C and Fortran interfaces.
Method intermediate callback

with prototype

virtual bool intermediate_callback(AlgorithmMode mode,
Index iter, Number obj_value,
Number inf_pr, Number inf_du,
Number mu, Number d_norm,
Number regularization_size,
Number alpha_du, Number alpha_pr,
Index ls_trials,
const IpoptData* ip_data,
IpoptCalculatedQuantities* ip_cq)
It is not required to implement (overload) this method. This method is called once per iteration (during
the convergence check), and can be used to obtain information about the optimization status while Ipopt
solves the problem, and also to request a premature termination.
The information provided by the entities in the argument list correspond to what Ipopt prints in
the iteration summary (see also Section 6). Further information can be obtained from the ip data and
ip cq objects (in the C++ interface and for experts only :)).

You you let this method return false, Ipopt will terminate with the User Requested Stop status. If
you do not implement this method (as we do in this example), the default implementation always returns
true.
A frequently asked question is how to access the values of the primal and dual variables in this callback.
The values are stored in the ip cq object for the internal representation of the problem. To access the
values in a form that corresponds to those used in the evaluation routines, the user has to request Ipopt’s
TNLPAdapter object to “resort” the data vectors and to fill in information about possibly filtered out
fixed variables. The TNLPAdapter can be accessed as follows. First, add the following includes to your
TNLP implementation:
#include
#include
#include
#include

"IpIpoptCalculatedQuantities.hpp"
"IpIpoptData.hpp"
"IpTNLPAdapter.hpp"
"IpOrigIpoptNLP.hpp"

Next, add the following code to your implementation of the intermediate callback:
Ipopt::TNLPAdapter* tnlp_adapter = NULL;
if( ip_cq != NULL )
{
Ipopt::OrigIpoptNLP* orignlp;
orignlp = dynamic_cast(GetRawPtr(ip_cq->GetIpoptNLP()));
if( orignlp != NULL )
tnlp_adapter = dynamic_cast(GetRawPtr(orignlp->nlp()));
}
Note, that retrieving the TNLPAdapter will fail (i.e., orignlp will be NULL) if Ipopt is currently in
restoration mode. If, however, tnlp adapter is not NULL, then it can be used to obtain primal variable
values x and the dual values for the constraints 2 and the variable bounds 3 as follows.
double* primals = new double[n];
double* dualeqs = new double[m];
double* duallbs = new double[n];
double* dualubs = new double[n];
tnlp_adapter->ResortX(*ip_data->curr()->x(), primals);
tnlp_adapter->ResortG(*ip_data->curr()->y_c(), *ip_data->curr()->y_d(), dualeqs);
tnlp_adapter->ResortBnds(*ip_data->curr()->z_L(), duallbs,
*ip_data->curr()->z_U(), dualubs);
Additionally, information about scaled violation of constraint 2 and violation of complementarity constraints can be obtained via
tnlp_adapter->ResortG(*ip_data->curr_c(), *ip_data->curr_d_minus_s(), ...)
tnlp_adapter->ResortBnds(*ip_data->curr_compl_x_L(), ...,
*ip_data->curr_compl_x_U(), ...)
tnlp_adapter->ResortG(*ip_data->curr_compl_s_L(), ...)
tnlp_adapter->ResortG(*ip_data->curr_compl_s_U(), ...)
Method get scaling parameters with prototype
virtual bool get_scaling_parameters(Number& obj_scaling,
bool& use_x_scaling, Index n,
Number* x_scaling,
bool& use_g_scaling, Index m,
Number* g_scaling)
41

This method is called if the nlp scaling method is chosen as user-scaling. The user has to provide
scaling factors for the objective function as well as for the optimization variables and/or constraints. The
return value should be true, unless an error occurred, and the program is to be aborted.
The value returned in obj scaling determines, how Ipopt should internally scale the objective
function. For example, if this number is chosen to be 10, then Ipopt solves internally an optimization
problem that has 10 times the value of the original objective function provided by the TNLP. In particular,
if this value is negative, then Ipopt will maximize the objective function instead of minimizing it.
The scaling factors for the variables can be returned in x scaling, which has the same length as x in
the other TNLP methods, and the factors are ordered like x. You need to set use x scaling to true, if you
want Ipopt so scale the variables. If it is false, no internal scaling of the variables is done. Similarly, the
scaling factors for the constraints can be returned in g scaling, and this scaling is activated by setting
use g scaling to true.
As a guideline, we suggest to scale the optimization problem (either directly in the original formulation,
or after using scaling factors) so that all sensitivities, i.e., all non-zero first partial derivatives, are typically
of the order 0.1 − 10.
Method get number of nonlinear variables with prototype
virtual Index get_number_of_nonlinear_variables()
This method is only important if the limited-memory quasi-Newton options is used, see Section 4.2.
It is used to return the number of variables that appear nonlinearly in the objective function or in at least
one constraint function. If a negative number is returned, Ipopt assumes that all variables are nonlinear.
If the user doesn’t overload this method in her implementation of the class derived from TNLP, the
default implementation returns -1, i.e., all variables are assumed to be nonlinear.
Method get list of nonlinear variables

with prototype

virtual bool get_list_of_nonlinear_variables(Index num_nonlin_vars,
Index* pos_nonlin_vars)
This method is called by Ipopt only if the limited-memory quasi-Newton options is used and if the
get number of nonlinear variables method returns a positive number; this number is then identical
with num nonlin vars and the length of the array pos nonlin vars. In this call, you need to list
the indices of all nonlinear variables in pos nonlin vars, where the numbering starts with 0 order 1,
depending on the numbering style determined in get nlp info.
Method get variables linearity

with prototype

virtual bool get_variables_linearity(Index n,
LinearityType* var_types)
This method is never called by Ipopt, but is used by Bonmin to get information about which variables
occur only in linear terms. Ipopt passes a var types array of size n, which the user should fill with the
appropriate linearity type of the variables (TNLP::LINEAR or TNLP::NON LINEAR).
If the user doesn’t overload this method in her implementation of the class derived from TNLP, the
default implementation returns false.
Method get constraints linearity

with prototype

virtual bool get_constraints_linearity(Index m,
LinearityType* const_types)

This method is never called by Ipopt, but is used by Bonmin to get information about which
constraints are linear. Ipopt passes a const types array of size m, which the user should fill with the
appropriate linearity type of the constraints (TNLP::LINEAR or TNLP::NON LINEAR).
If the user doesn’t overload this method in her implementation of the class derived from TNLP, the
default implementation returns false.
Method get var con metadata with prototype
virtual bool get_var_con_metadata(Index n,
StringMetaDataMapType& var_string_md,
IntegerMetaDataMapType& var_integer_md,
NumericMetaDataMapType& var_numeric_md,
Index m,
StringMetaDataMapType& con_string_md,
IntegerMetaDataMapType& con_integer_md,
NumericMetaDataMapType& con_numeric_md)
This method is used to pass meta data about variables or constraints to Ipopt. The data can be either
of integer, numeric, or string type. Ipopt passes this data on to its internal problem representation. The
meta data type is a std::map with std::string as key type and a std::vector as value type. So far,
Ipopt itself makes only use of string meta data under the key idx names. With this key, variable and
constraint names can be passed to Ipopt, which are shown when printing internal vector or matrix data
structures if Ipopt is run with a high value for the print level option. This allows a user to identify
the original variables and constraints corresponding to Ipopt’s internal problem representation.
If the user doesn’t overload this method in her implementation of the class derived from TNLP, the
default implementation does not set any meta data and returns false.
Method finalize metadata with prototype
virtual void finalize_metadata(Index
const
const
const
Index
const
const
const

n,
StringMetaDataMapType& var_string_md,
IntegerMetaDataMapType& var_integer_md,
NumericMetaDataMapType& var_numeric_md,
m,
StringMetaDataMapType& con_string_md,
IntegerMetaDataMapType& con_integer_md,
NumericMetaDataMapType& con_numeric_md)

This method is called just before finalize solution and is used to return any meta data collected
during the algorithms run, including the meta data provided by the user with the get var con metadata
method.
If the user doesn’t overload this method in her implementation of the class derived from TNLP, the
default implementation does nothing.
Method get warm start iterate with prototype
virtual bool get_warm_start_iterate(IteratesVector& warm_start_iterate)
Overload this method to provide an Ipopt iterate which is already in the form Ipopt requires it
internally for a warm starts. This method is only for expert users.
If the user doesn’t overload this method in her implementation of the class derived from TNLP, the
default implementation does not provide a warm start iterate and returns false.

3.4

The C Interface

The C interface for Ipopt is declared in the header file IpStdCInterface.h, which is found in
$IPOPTDIR/include/coin (or in $PREFIX/include/coin if the switch --prefix=$PREFIX was used for
configure); while reading this section, it will be helpful to have a look at this file.
In order to solve an optimization problem with the C interface, one has to create an IpoptProblem17
with the function CreateIpoptProblem, which later has to be passed to the IpoptSolve function.
The IpoptProblem created by CreateIpoptProblem contains the problem dimensions, the variable
and constraint bounds, and the function pointers for callbacks that will be used to evaluate the NLP
problem functions and their derivatives (see also the discussion of the C++ methods get nlp info and
get bounds info in Section 3.3.1 for information about the arguments of CreateIpoptProblem).
The prototypes for the callback functions, Eval F CB, Eval Grad F CB, etc., are defined in the header
file IpStdCInterface.h. Their arguments correspond one-to-one to the arguments for the C++ methods
discussed in Section 3.3.1; for example, for the meaning of n, x, new x, obj value in the declaration of
Eval F CB see the discussion of “eval f”. The callback functions should return TRUE, unless there was
a problem doing the requested function/derivative evaluation at the given point x (then it should return
FALSE).
Note the additional argument of type UserDataPtr in the callback functions. This pointer argument
is available for you to communicate information between the main program that calls IpoptSolve and
any of the callback functions. This pointer is simply passed unmodified by Ipopt among those functions.
For example, you can use this to pass constants that define the optimization problem and are computed
before the optimization in the main C program to the callback functions.
After an IpoptProblem has been created, you can set algorithmic options for Ipopt (see Section 5)
using the AddIpopt...Option functions. Finally, the Ipopt algorithm is called with IpoptSolve, giving
Ipopt the IpoptProblem, the starting point, and arrays to store the solution values (primal and dual
variables), if desired. Finally, after everything is done, you should call FreeIpoptProblem to release
internal memory that is still allocated inside Ipopt.
In the remainder of this section we discuss how the example problem (4)–(7) can be solved using the
C interface. A completed version of this example can be found in Ipopt/examples/hs071 c.
In order to implement the example problem on your own, create a new directory MyCExample and
create a new file, hs071 c.c. Here, include the interface header file IpStdCInterface.h, along with
other necessary header files, such as stdlib.h and assert.h. Add the prototypes and implementations for the five callback functions. Have a look at the C++ implementation for eval f, eval g,
eval grad f, eval jac g, and eval h in Section 3.3.1. The C implementations have somewhat different
prototypes, but are implemented almost identically to the C++ code. See the completed example in
Ipopt/examples/hs071 c/hs071 c.c if you are not sure how to do this.
We now need to implement the main function, create the IpoptProblem, set options, and call
IpoptSolve. The CreateIpoptProblem function requires the problem dimensions, the variable and
constraint bounds, and the function pointers to the callback routines. The IpoptSolve function requires
the IpoptProblem, the starting point, and allocated arrays for the solution. The main function from the
example is shown next and discussed below.
int main()
{
Index n=-1;
Index m=-1;
Number* x_L = NULL;
Number* x_U = NULL;
Number* g_L = NULL;

/*
/*
/*
/*
/*

number of variables */
number of constraints */
lower bounds on x */
upper bounds on x */
lower bounds on g */

17 IpoptProblem is a pointer to a C structure; you should not access this structure directly, only through the functions
provided in the C interface.

Number* g_U = NULL;
/* upper bounds on g */
IpoptProblem nlp = NULL;
/* IpoptProblem */
enum ApplicationReturnStatus status; /* Solve return code */
Number* x = NULL;
/* starting point and solution vector */
Number* mult_x_L = NULL;
/* lower bound multipliers at the solution */
Number* mult_x_U = NULL;
/* upper bound multipliers at the solution */
Number obj;
/* objective value */
Index i;
/* generic counter */
/* set the number of variables and allocate space for the bounds */
n=4;
x_L = (Number*)malloc(sizeof(Number)*n);
x_U = (Number*)malloc(sizeof(Number)*n);
/* set the values for the variable bounds */
for (i=0; i print( ipoptr( x0 = x0,
eval_f = eval_f,
eval_grad_f = eval_grad_f,
lb = lb,
ub = ub,
eval_g = eval_g,
eval_jac_g = eval_jac_g,
constraint_lb = constraint_lb,
constraint_ub = constraint_ub,
eval_jac_g_structure = eval_jac_g_structure,
eval_h = eval_h,
eval_h_structure = eval_h_structure,
opts = opts) )
53

Call:
ipoptr(x0 = x0, eval_f = eval_f, eval_grad_f = eval_grad_f, lb = lb,
ub = ub, eval_g = eval_g, eval_jac_g = eval_jac_g,
eval_jac_g_structure = eval_jac_g_structure, constraint_lb = constraint_lb,
constraint_ub = constraint_ub, eval_h = eval_h, eval_h_structure = eval_h_structure,
opts = opts)
Ipopt solver status: 0 ( SUCCESS: Algorithm terminated
successfully at a locally optimal point, satisfying the
convergence tolerances (can be specified by options). )
Number of Iterations....: 8
Optimal value of objective function: 17.0140171451792
Optimal value of controls: 1 4.743 3.82115 1.379408
To pass additional data to the evaluation routines, one can either supply additional arguments to the
user defined functions and ipoptr or define an environment that holds the data and pass this environment
to ipoptr. Both methods are shown in the file tests/parameters.R that comes with ipoptr.
As a very simple example, suppose we want to find the minimum of
f (x) = a1 x2 + a2 x + a3
for different values of the parameters a1 , a2 and a3 .
First, we define the objective function and its gradient using, assuming that there is some variable
params that contains the values of the parameters.
> eval_f_ex1 <- function(x, params) {
return( params[1]*x^2 + params[2]*x + params[3] )
}
> eval_grad_f_ex1 <- function(x, params) {
return( 2*params[1]*x + params[2] )
}
Note, that the first parameter should always be the control variable. All of the user-defined functions
should contain the same set of additional parameters. You have to supply them as input argument to all
functions, even if you’re not using them in some of the functions.
Then we can solve the problem for a specific set of parameters, in this case a1 = 1, a2 = 2 and a3 = 3,
from initial value x0 = 0, with the following command
> # solve using ipoptr
> ipoptr(x0
=
eval_f
=
eval_grad_f =
opts
=
params
=

with additional parameters
0,
eval_f_ex1,
eval_grad_f_ex1,
list("print_level"=0),
c(1,2,3) )

Call:
ipoptr(x0 = 0, eval_f = eval_f_ex1, eval_grad_f = eval_grad_f_ex1,
opts = list(print_level = 0), params = c(1, 2, 3))

Ipopt solver status: 0 ( SUCCESS: Algorithm terminated
successfully at a locally optimal point, satisfying the
convergence tolerances (can be specified by options). )
Number of Iterations....: 1
Optimal value of objective function:
Optimal value of controls: -1

For the second method, we don’t have to supply the parameters as additional arguments to the
function.
> eval_f_ex2 <- function(x) {
return( params[1]*x^2 + params[2]*x + params[3] )
}
> eval_grad_f_ex2 <- function(x) {
return( 2*params[1]*x + params[2] )
}
Instead, we define an environment that contains specific values of params
> # define a new environment that contains params
> auxdata
<- new.env()
> auxdata$params <- c(1,2,3)
To solve this we supply auxdata as an argument to ipoptr, which will take care of evaluating the
functions in the correct environment, so that the auxiliary data is available.
> # pass the environment that
> ipoptr(x0
=
eval_f
=
eval_grad_f
=
ipoptr_environment =
opts
=

should be used to evaluate functions to ipoptr
0,
eval_f_ex2,
eval_grad_f_ex2,
auxdata,
list("print_level"=0) )

Call:
ipoptr(x0 = 0, eval_f = eval_f_ex2, eval_grad_f = eval_grad_f_ex2,
opts = list(print_level = 0), ipoptr_environment = auxdata)

3.8

The MATLAB Interface
based on documentation by Peter Carbonetto20

See Section 2.8 for instructions on how to build a mex file of the MATLAB interface for Ipopt and
how to make it known to MATLAB.
20 University

of British Columbia

The $IPOPTDIR/contrib/MatlabInterface/examples directory contains several illustrative examples on how to use the MATLAB interface. The best way to understand how to use the interface is to
carefully go over these examples.
For more information, type help ipopt in the MATLAB prompt.
Further, Jonas Asprion assembled information about the MATLAB ipopt function and its arguments:
http://www.idsc.ethz.ch/Downloads/IPOPT_InstallMatlab/IPOPT_MatlabInterface_V0p1.pdf
Note, that this document refers to Ipopt versions before 3.11. With 3.11, the auxdata option has
been removed from the mex code. The new wrapper function ipopt auxdata implements the same
functionality as the previous ipopt function, but uses MATLAB function handles to do so.

Special Features

4.1

Derivative Checker

When writing code for the evaluation of derivatives it is very easy to make mistakes (much easier than
writing it correctly the first time :)). As a convenient feature, Ipopt provides the option to run a simple
derivative checker, based on finite differences, before the optimization is started.
To use the derivative checker, you need to use the option derivative test. By default, this option is
set to none, i.e., no finite difference test is performed, It is set to first-order, then the first derivatives
of the objective function and the constraints are verified, and for the setting second-order, the second
derivatives are tested as well.
The verification is done by a simple finite differences approximation, where each component of the userprovided starting point is perturbed one of the other. The relative size of the perturbation is determined
by the option derivative test perturbation. The default value (10−8 , about the square root of the
machine precision) is probably fine in most cases, but if you believe that you see wrong warnings, you
might want to play with this parameter. When the test is performed, Ipopt prints out a line for every
partial derivative, for which the user-provided derivative value deviates too much from the finite difference
approximation. The relative tolerance for deciding when a warning should be issued, is determined by the
option derivative test tol. If you want to see the user-provided and estimated derivative values with
the relative deviation for each single partial derivative, you can switch the derivative test print all
option to yes.
A typical output is:
Starting derivative checker.
*
*
*
*

grad_f[
jac_g [
jac_g [
jac_g [

4,
4,
6,

2]
4]
5]
7]

= -6.5159999999999991e+02
= 0.0000000000000000e+00
= 1.3798494268463347e+01 v
= 1.4776333636790881e+01 v

~ -6.5559997134793468e+02
~ 2.2160643690464592e-02
~ 1.3776333629422766e+01
~ 1.3776333629422766e+01

[
[
[
[

6.101e-03]
2.216e-02]
1.609e-03]
7.259e-02]

Derivative checker detected 4 error(s).

The star (“*”) in the first column indicates that this line corresponds to some partial derivative for
which the error tolerance was exceeded. Next, we see which partial derivative is concerned in this output
line. For example, in the first line, it is the second component of the objective function gradient (or the
third, if the C STYLE numbering is used, i.e., when counting of indices starts with 0 instead of 1). The
first floating point number is the value given by the user code, and the second number (after “~”) is the
finite differences estimation. Finally, the number in square brackets is the relative difference between
these two numbers.
For constraints, the first index after jac g is the index of the constraint, and the second one corresponds to the variable index (again, the choice of the numbering style matters).
Since also the sparsity structure of the constraint Jacobian has to be provided by the user, it can be
faulty as well. For this, the “v” after a user-provided derivative value indicates that this component of
the Jacobian is part of the user provided sparsity structure. If there is no “v”, it means that the user
56

did not include this partial derivative in the list of non-zero elements. In the above output, the partial
derivative “jac g[4,4]” is non-zero (based on the finite difference approximation), but it is not included
in the list of non-zero elements (missing “v”), so that the user probably made a mistake in the sparsity
structure. The other two Jacobian entries are provided in the non-zero structure but their values seem
to be off.
For second derivatives, the output looks like:
*
*

obj_hess[
3-th constr_hess[

1,
2,

1] =
4] =

1.8810000000000000e+03 v
1.0000000000000000e+00 v

~
~

1.8820000036612328e+03
0.0000000000000000e+00

[ 5.314e-04]
[ 1.000e+00]

There, the first line shows the deviation of the user-provided partial second derivative in the Hessian
for the objective function, and the second line show an error in a partial derivative for the Hessian of the
third constraint (again, the numbering style matters).
Since the second derivatives are approximates by finite differences of the first derivatives, you should
first correct errors for the first derivatives. Also, since the finite difference approximations are quite
expensive, you should try to debug a small instance of your problem if you can.
Another useful option is derivative test first index which allows your to start the derivative test
with variables with a larger index. Finally, it is of course always a good idea to run your code through
some memory checker, such as valgrind on Linux.

4.2

Quasi-Newton Approximation of Second Derivatives

Ipopt has an option to approximate the Hessian of the Lagrangian by a limited-memory quasi-Newton
method (L-BFGS). You can use this feature by setting the hessian approximation option to the value
limited-memory. In this case, it is not necessary to implement the Hessian computation method eval h
in TNLP. If you are using the C or Fortran interface, you still need to implement these functions, but they
should return false or IERR=1, respectively, and don’t need to do anything else.
In general, when second derivatives can be computed with reasonable computational effort, it is usually
a good idea to use them, since then Ipopt normally converges in fewer iterations and is more robust. An
exception might be in cases, where your optimization problem has a dense Hessian, i.e., a large percentage
of non-zero entries in the Hessian. In such a case, using the quasi-Newton approximation might be better,
even if it increases the number of iterations, since with exact second derivatives the computation time
per iteration might be significantly higher due to the very large number of non-zero elements in the linear
systems that Ipopt solve in order to compute the search direction.
Since the Hessian of the Lagrangian is zero for all variables that appear only linearly in the objective and constraint functions, the Hessian approximation should only take place in the space of
all nonlinear variables. By default, it is assumed that all variables are nonlinear, but you can tell
Ipopt explicitly which variables are nonlinear, using the get number of nonlinear variables and
get list of nonlinear variables method of the TNLP class, see Section 3.3.4. (Those methods have
been implemented for the AMPL interface, so you would automatically only approximate the Hessian
in the space of the nonlinear variables, if you are using the quasi-Newton option for AMPL models.)
Currently, those two methods are not available through the C or Fortran interface.

4.3

Warm-Starting Capabilities via AMPL
based on documentation by Victor M. Zavala 21

Warm-starting an interior-point algorithm is an important issue. One of the main difficulties arises
from the fact that full-space variable information is required to generate the warm-starting point. While
Ipopt is currently equipped to retrieve and receive this type of information through the TNLP interface,
there exist some communication barriers in the AMPL interface. When the user solves the problem (1)–
(3), Ipopt will only return the optimal values of the primal variables x and of the constraint multipliers
21 Department

of Chemical Engineering, Carnegie Mellon University

corresponding to the active bounds of g(x) (see (2)). The constraint multiplier values can be accessed
through the .dual suffix or through the .sol file. If this information is used to solve the same problem
again, you will notice that Ipopt will take some iterations in finding the same solution. The reason
for this is that we are missing the input information of the multipliers z L and z U corresponding to the
variable bounds (see (3)).
However, Ipopt also passes the values of the bound multipliers z L and z U to AMPL. This will be
communicated to the AMPL user through the suffixes ipopt zL out and ipopt zU out, respectively. The
user does not need to declare these suffixes, they will be generated automatically in the AMPL interface.
The user can use the suffix values to initialize the bound multipliers for subsequent calls. In order to pass
this information to Ipopt, the user will need to declare and assign values to the suffixes ipopt zL in
and ipopt zU in. For instance, for a given variable x[i], this can be done by setting:
let x[i].ipopt_zL_in := x[i].ipopt_zL_out;
let x[i].ipopt_zU_in := x[i].ipopt_zU_out;
If the user does not specify some of these values, Ipopt will set these multipliers to 1.0 (as before).
In order to make the warm-start effective, the user has control over the following options from AMPL:
warm start init point
warm start bound push
warm start mult bound push
Note, that the use of this feature is far from solving the complicated issue of warm- starting interiorpoint algorithms. As a general advice, this feature will be useful if the user observes that the solution
of subsequent problems (i.e., for different data instances) preserves the same set of active inequalities and bounds (monitor the values of z L and z U for subsequent solutions). In this case, initializing
the bound multipliers and setting warm start init point to yes and setting warm start bound push,
warm start mult bound push and mu init to a small value (10−6 or so) will reduce significantly the
number of iterations. This is particularly useful in setting up on-line applications and high-level optimization strategies in AMPL. If active-set changes are observed between subsequent solutions, then this
strategy might not decrease the number of iterations (in some cases, it might even tend to increase the
number of iterations).
You might also want to try the adaptive barrier update (instead of the default monotone one where
above we chose the initial value 10−6 ) when doing the warm start. This can be activated by setting the
mu strategy option to adaptive. Also the option mu oracle gives some alternative choices. In general,
the adaptive choice often leads to less iterations, but the computational cost per iteration might be higher.
The file $IPOPTDIR/Ipopt/doc/hs071 warmstart.mod illustrates the use of the warm-start feature
on the HS071 problem, see also Section 3.1.

4.4

sIpopt: Optimal Sensitivity Based on Ipopt
based on documentation by Hans Pirnay22 and Rodrigo López-Negrete23

The sIpopt project provides a toolbox that uses NLP sensitivity theory to generate fast approximations to solutions when parameters in the problem change. It has been developed primarily by Hans
Pirnay (RWTH-Aachen), Rodrigo López-Negrete (CMU), and Lorenz Biegler (CMU).
Sensitivity of nonlinear programming problems is a key step in any optimization study. Sensitivity
provides information on regularity and curvature conditions at KKT points, assesses which variables play
dominant roles in the optimization, and provides first order estimates for parametric nonlinear programs.
Moreover, for NLP algorithms that use exact second derivatives, sensitivity can be implemented very
efficiently within NLP solvers and provide valuable information with very little added computation. This
implementation provides Ipopt with the capabilities to calculate sensitivities, and approximate perturbed
solutions with them.
22 RWTH

Aachen, hans.pirnay@avt.rwth-aachen.de
Mellon University, rln@cmu.edu

23 Carnegie

The basic sensitivity strategy implemented here is based on the application of the Implicit Function
Theorem (IFT) to the KKT conditions of the NLP. As shown by Fiacco (1983), sensitivities can be
obtained from a solution with suitable regularity conditions merely by solving a linearization of the KKT
conditions. More details can be found in [7]. If you are using sIpopt for your research, please cite [7].
The sIpopt project is available in the Ipopt repository under $IPOPTDIR/Ipopt/contrib/sIPOPT.
After having installed Ipopt successfully, sIpopt can be build and installed by changing to the directory $IPOPTDIR/build/Ipopt/contrib/sIPOPT and executing make install. This should copy the
generated libraries libsipopt.* to $IPOPTDIR/build/lib and an AMPL executable ipopt sens to
$IPOPTDIR/build/bin.
The files $IPOPTDIR/Ipopt/contrib/sIPOPT/examples/parametric ampl/parametric.{mod,run}
are an example that shows how to use sIpopt to solve the NLP
min x21 + x22 + x23 ,
such that

(10)

6x1 + 3x2 + 2x3 = p1 ,

(11)

p2 x1 + x2 − x3 = 1,

(12)

x1 , x2 , x3 ≥ 0,

(13)

where we perturb the parameters p1 and p2 from pa = (p1 , p2 ) = (5, 1) to pb = (4.5, 1).
Note, that sIpopt has been developed under the constraint that it must work with the regular Ipopt
code. Due to this constraint, some compromises had to be made. However, there is an ongoing effort to
develop sIpopt 2, which is a fork of the Ipopt code that allows for the explicit definition of parametric
NLPs. This code can be found at https://github.com/athrpf/sipopt2. If you have questions about
sIpopt 2, please contact Hans Pirnay.

Ipopt Options

Ipopt has many (maybe too many) options that can be adjusted for the algorithm. Options are all
identified by a string name, and their values can be of one of three types: Number (real), Integer,
or String. Number options are used for things like tolerances, integer options are used for things like
maximum number of iterations, and string options are used for setting algorithm details, like the NLP
scaling method. Options can be set through code, through the AMPL interface if you are using AMPL,
or by creating a ipopt.opt file in the directory you are executing Ipopt.
The ipopt.opt file is read line by line and each line should contain the option name, followed by
whitespace, and then the value. Comments can be included with the # symbol. For example,
# This is a comment
# Turn off the NLP scaling
nlp_scaling_method none
# Change the initial barrier parameter
mu_init 1e-2
# Set the max number of iterations
max_iter 500
is a valid ipopt.opt file.
Options can also be set in code. Have a look at the examples to see how this is done.
A subset of Ipopt options are available through AMPL. To set options through AMPL, use the
internal AMPL command options. For example,
options ipopt options "nlp scaling method=none mu init=1e-2 max iter=500"
is a valid options command in AMPL. The most important options are referenced in Appendix C. To
59

see which options are available through AMPL, you can run the AMPL solver executable with the “-=”
flag from the command prompt. To specify other options when using AMPL, you can always create
ipopt.opt. Note, the ipopt.opt file is given preference when setting options. This way, you can easily
override any options set in a particular executable or AMPL model by specifying new values in ipopt.opt.
For a list of the most important valid options, see the Appendix C. You can print the documentation
for all Ipopt options by using the option
print options documentation yes
and running Ipopt (like the AMPL solver executable, for instance). This will output the documentation of almost all options to the console.

Ipopt Output

This section describes the standard Ipopt console output with the default setting for print level. The
output is designed to provide a quick summary of each iteration as Ipopt solves the problem.
Before Ipopt starts to solve the problem, it displays the problem statistics (number of nonzeroelements in the matrices, number of variables, etc.). Note that if you have fixed variables (both upper
and lower bounds are equal), Ipopt may remove these variables from the problem internally and not
include them in the problem statistics.
Following the problem statistics, Ipopt will begin to solve the problem and you will see output
resembling the following,
iter
0
1
2

objective
inf_pr
inf_du lg(mu) ||d|| lg(rg)
1.6109693e+01 1.12e+01 5.28e-01
0.0 0.00e+00
1.8029749e+01 9.90e-01 6.62e+01
0.1 2.05e+00
1.8719906e+01 1.25e-02 9.04e+00 -2.2 5.94e-02
2.0

alpha_du
0.00e+00
2.14e-01
8.04e-01

alpha_pr ls
0.00e+00
0
1.00e+00f 1
1.00e+00h 1

and the columns of output are defined as,
iter: The current iteration count. This includes regular iterations and iterations during the restoration
phase. If the algorithm is in the restoration phase, the letter r’ will be appended to the iteration
number.
objective: The unscaled objective value at the current point. During the restoration phase, this value
remains the unscaled objective value for the original problem.
inf pr: The unscaled constraint violation at the current point. This quantity is the infinity-norm (max)
of the (unscaled) constraints (2). During the restoration phase, this value remains the constraint
violation of the original problem at the current point. The option inf pr output can be used to
switch to the printing of a different quantity.
inf du: The scaled dual infeasibility at the current point. This quantity measure the infinity-norm (max)
of the internal dual infeasibility, Eq. (4a) in the implementation paper [13], including inequality
constraints reformulated using slack variables and problem scaling. During the restoration phase,
this is the value of the dual infeasibility for the restoration phase problem.
lg(mu): log10 of the value of the barrier parameter µ.
||d||: The infinity norm (max) of the primal step (for the original variables x and the internal slack
variables s). During the restoration phase, this value includes the values of additional variables, p
and n (see Eq. (30) in [13]).
lg(rg): log10 of the value of the regularization term for the Hessian of the Lagrangian in the augmented
system (δw in Eq. (26) and Section 3.1 in [13]). A dash (“-”) indicates that no regularization was
done.
60

Tag
f
F
h
H
k
K
n
N
R
w
s
t/T
r

Description
f-type iteration in the filter method w/o second order correction
f-type iteration in the filter method w/ second order correction
h-type iteration in the filter method w/o second order correction
h-type iteration in the filter method w/ second order correction
penalty value unchanged in merit function method w/o second order correction
penalty value unchanged in merit function method w/ second order correction
penalty value updated in merit function method w/o second order correction
penalty value updated in merit function method w/ second order correction
Restoration phase just started
in watchdog procedure
step accepted in soft restoration phase
tiny step accepted without line search
some previous iterate restored

Table 1: Diagnostic output in alpha pr column.
alpha du: The stepsize for the dual variables (αkz in Eq. (14c) in [13]).
alpha pr: The stepsize for the primal variables (αk in Eq. (14a) in [13]). The number is usually followed
by a character for additional diagnostic information regarding the step acceptance criterion, see
Table 1.
ls: The number of backtracking line search steps (does not include second-order correction steps).
Note that the step acceptance mechanisms in Ipopt consider the barrier objective function (Eq (3a)
in [13]) which is usually different from the value reported in the objective column. Similarly, for the
purposes of the step acceptance, the constraint violation is measured for the internal problem formulation,
which includes slack variables for inequality constraints and potentially scaling of the constraint functions.
This value, too, is usually different from the value reported in inf pr. As a consequence, a new iterate
might have worse values both for the objective function and the constraint violation as reported in the
iteration output, seemingly contradicting globalization procedure.
When the algorithm terminates, Ipopt will output a message to the screen based on the return status
of the call to Optimize. The following is a list of the possible return codes, their corresponding output
message to the console, and a brief description.
Solve Succeeded:
Console Message: EXIT: Optimal Solution Found.
This message indicates that Ipopt found a (locally) optimal point within the desired tolerances.
Solved To Acceptable Level:
Console Message: EXIT: Solved To Acceptable Level.
This indicates that the algorithm did not converge to the “desired” tolerances, but that it was
able to obtain a point satisfying the “acceptable” tolerance level as specified by the acceptable *
options. This may happen if the desired tolerances are too small for the current problem.
Feasible Point Found:
Console Message: EXIT: Feasible point for square problem found.
This message is printed if the problem is “square” (i.e., it has as many equality constraints as free
variables) and Ipopt found a feasible point.
Infeasible Problem Detected:
Console Message: EXIT: Converged to a point of local infeasibility.

Problem may be

infeasible.
The restoration phase converged to a point that is a minimizer for the constraint violation (in the
`1 -norm), but is not feasible for the original problem. This indicates that the problem may be
infeasible (or at least that the algorithm is stuck at a locally infeasible point). The returned point
(the minimizer of the constraint violation) might help you to find which constraint is causing the
problem. If you believe that the NLP is feasible, it might help to start the optimization from a
different point.
Search Direction Becomes Too Small:
Console Message: EXIT: Search Direction is becoming Too Small.
This indicates that Ipopt is calculating very small step sizes and is making very little progress.
This could happen if the problem has been solved to the best numerical accuracy possible given the
current scaling.
Diverging Iterates:
Console Message: EXIT: Iterates divering; problem might be unbounded.
This message is printed if the max-norm of the iterates becomes larger than the value of the option
diverging iterates tol. This can happen if the problem is unbounded below and the iterates
are diverging.
User Requested Stop:
Console Message: EXIT: Stopping optimization at current point as requested by user.
This message is printed if the user call-back method intermediate callback returned false (see
Section 3.3.4).
Maximum Iterations Exceeded:
Console Message: EXIT: Maximum Number of Iterations Exceeded.
This indicates that Ipopt has exceeded the maximum number of iterations as specified by the
option max iter.
Maximum CpuTime Exceeded:
Console Message: EXIT: Maximum CPU time exceeded.
This indicates that Ipopt has exceeded the maximum number of CPU seconds as specified by the
option max cpu time.
Restoration Failed:
Console Message: EXIT: Restoration Failed!
This indicates that the restoration phase failed to find a feasible point that was acceptable to the
filter line search for the original problem. This could happen if the problem is highly degenerate,
does not satisfy the constraint qualification, or if your NLP code provides incorrect derivative
information.
Error In Step Computation:
Console Output: EXIT: Error in step computation (regularization becomes too large?)!
This messages is printed if Ipopt is unable to compute a search direction, despite several attempts
to modify the iteration matrix. Usually, the value of the regularization parameter then becomes
too large. One situation where this can happen is when values in the Hessian are invalid (NaN or
Inf). You can check whether this is true by using the check derivatives for naninf option.
Invalid Option:
Console Message: (details about the particular error will be output to the console)
This indicates that there was some problem specifying the options. See the specific message for
details.

Not Enough Degrees Of Freedom:
Console Message: EXIT: Problem has too few degrees of freedom.
This indicates that your problem, as specified, has too few degrees of freedom. This can happen
if you have too many equality constraints, or if you fix too many variables (Ipopt removes fixed
variables by default, see also the fixed variable treatment option).
Invalid Problem Definition:
Console Message: (no console message, this is a return code for the C and Fortran interfaces only.)
This indicates that there was an exception of some sort when building the IpoptProblem structure
in the C or Fortran interface. Likely there is an error in your model or the main routine.
Unrecoverable Exception:
Console Message: (details about the particular error will be output to the console)
This indicates that Ipopt has thrown an exception that does not have an internal return code. See
the specific message for details.
NonIpopt Exception Thrown:
Console Message: Unknown Exception caught in Ipopt
An unknown exception was caught in Ipopt. This exception could have originated from your model
or any linked in third party code.
Insufficient Memory:
Console Message: EXIT: Not enough memory.
An error occurred while trying to allocate memory. The problem may be too large for your current
memory and swap configuration.
Internal Error:
Console: EXIT: INTERNAL ERROR: Unknown SolverReturn value - Notify IPOPT Authors.
An unknown internal error has occurred. Please notify the authors of Ipopt via the mailing list.

6.1

Diagnostic Tags for Ipopt

To print additional diagnostic tags for each iteration of Ipopt, set the options print info string to
yes. With this, a tag will appear at the end of an iteration line with the following diagnostic meaning
that are useful to flag difficulties for a particular Ipopt run. A list of possible strings is given in Table 2.

Triplet Format for Sparse Matrices

Ipopt was designed for optimizing large sparse nonlinear programs. Because of problem sparsity, the
required matrices (like the constraints Jacobian or Lagrangian Hessian) are not stored as dense matrices,
but rather in a sparse matrix format. For the tutorials in this document, we use the triplet format.
Consider the matrix


1.1 0
0
0
0
0 0.5
 0 1.9 0
0
0
0 0.5 


 0
0 2.6 0
0
0 0.5 




(14)
0 7.8 0.6 0
0
0 
 0


 0

0
0 1.5 2.7 0
0


 1.6 0
0
0 0.4 0
0 
0
0
0
0
0 0.9 1.7
A standard dense matrix representation would need to store 7 · 7=49 floating point numbers, where
many entries would be zero. In triplet format, however, only the nonzero entries are stored. The triplet
format records the row number, the column number, and the value of all nonzero entries in the matrix.
For the matrix above, this means storing 14 integers for the rows, 14 integers for the columns, and 14

Tag
!
A
a
C
Dh
Dhj
Dj
dx
e
FF+
L
l
M
Nh
Nhj
Nj
NW
q
R
S
s
s
Tmax
W
w
Wb
We
Wp
Wr
Ws
WS
y
z

Description
Tighten resto tolerance if only slightly infeasible
Current iteration is acceptable
Perturbation for PD singularity impossible, assume singular
Second Order Correction taken
Hessian degenerate based on multiple iterations
Hessian/Jacobian degenerate based on multiple iterations
Jacobian degenerate based on multiple iterations
δx perturbation too large
Cutting back α due to evaluation error
Filter should be reset, but maximal resets exceeded
Resetting filter due to last few rejections of filter
Degenerate Jacobian, δc already perturbed
Degenerate Jacobian, δc perturbed
Magic step taken for slack variables
Hessian not yet degenerate
Hessian/Jacobian not yet degenerate
Jacobian not yet degenerate
Warm start initialization failed
PD system possibly singular, attempt improving sol. quality
Solution of restoration phase
PD system possibly singular, accept current solution
PD system singular
Square Problem. Set multipliers to zero
Trial θ is larger than θmax
Watchdog line search procedure successful
Watchdog line search procedure unsuccessful, stopped
Undoing most recent SR1 update
Skip Limited-Memory Update in restoration phase
Safeguard B 0 = σI for Limited-Memory Update
Resetting Limited-Memory Update
Skip Limited-Memory Update since sT y is not positive
Skip Limited-Memory Update since ∆x is too small
Dual infeasibility, use least square multiplier update
Apply correction to bound multiplier if too large

Reference
Section 3.3 in [13]
Alternate termination
Section 3.1 in [13]
Section 2.4 in [13]
Section 3.1 in [13]
Section 3.1 in [13]
Section 3.1 in [13]
Section 3.1 in [13]
in backtracking line search
Section 2.3 in [13]
Section 2.3 in [13]
Section 3.1 in [13]
Section 3.1 in [13]
in backtracking line search
Section 3.1 in [13]
Section 3.1 in [13]
Section 3.1 in [13]
in Warm Start Initialization
Section 3.1 in [13]
Section 3.3 in [13]
Section 3.1 in [13]
Section 3.1 in [13]
Default initialization routine
filter parameter, see (21) in [13]
Section 3.2 in [13]
Section 3.2 in [13]
Section 5.4.1 in [1]
Section 5.4.1 in [1]
Section 5.4.1 in [1]
Section 5.4.1 in [1]
Section 5.4.1 in [1]
Section 5.4.1 in [1]
during ipopt algorithm
during ipopt algorithm

Table 2: Diagnostic output appended using print info sting.

floating point numbers for the values. While this does not seem like a huge space saving over the 49
floating point numbers stored in the dense representation, for larger matrices, the space savings are very
dramatic24 .
The parameter index style in get nlp info tells Ipopt if you prefer to use C style indexing (0based, i.e., starting the counting at 0) for the row and column indices or Fortran style (1-based). Tables
3 and 4 below show the triplet format for both indexing styles, using the example matrix (14).
row
iRow[0] = 1
iRow[1] = 1
iRow[2] = 2
iRow[3] = 2
iRow[4] = 3
iRow[5] = 3
iRow[6] = 4
iRow[7] = 4
iRow[8] = 5
iRow[9] = 5
iRow[10] = 6
iRow[11] = 6
iRow[12] = 7
iRow[13] = 7

col
jCol[0] = 1
jCol[1] = 7
jCol[2] = 2
jCol[3] = 7
jCol[4] = 3
jCol[5] = 7
jCol[6] = 3
jCol[7] = 4
jCol[8] = 4
jCol[9] = 5
jCol[10] = 1
jCol[11] = 5
jCol[12] = 6
jCol[13] = 7

value
values[0] = 1.1
values[1] = 0.5
values[2] = 1.9
values[3] = 0.5
values[4] = 2.6
values[5] = 0.5
values[6] = 7.8
values[7] = 0.6
values[8] = 1.5
values[9] = 2.7
values[10] = 1.6
values[11] = 0.4
values[12] = 0.9
values[13] = 1.7

Table 3: Triplet Format of Matrix (14) with index style=FORTRAN STYLE

row
iRow[0] = 0
iRow[1] = 0
iRow[2] = 1
iRow[3] = 1
iRow[4] = 2
iRow[5] = 2
iRow[6] = 3
iRow[7] = 3
iRow[8] = 4
iRow[9] = 4
iRow[10] = 5
iRow[11] = 5
iRow[12] = 6
iRow[13] = 6

col
jCol[0] = 0
jCol[1] = 6
jCol[2] = 1
jCol[3] = 6
jCol[4] = 2
jCol[5] = 6
jCol[6] = 2
jCol[7] = 3
jCol[8] = 3
jCol[9] = 4
jCol[10] = 0
jCol[11] = 4
jCol[12] = 5
jCol[13] = 6

Table 4: Triplet Format of Matrix (14) with index style=C STYLE

The individual elements of the matrix can be listed in any order, and if there are multiple items for
the same nonzero position, the values provided for those positions are added.
The Hessian of the Lagrangian is a symmetric matrix. In the case of a symmetric matrix, you only
need to specify the lower left triangular part (or, alternatively, the upper right triangular part). For
24 For an n × n matrix, the dense representation grows with the the square of n, while the sparse representation grows
linearly in the number of nonzeros.

example, given the matrix,








1.0 0
0 1.1
3.0 0
0
0
2.0 5.0

3.0 0
0
0
1.2 6.0
6.0 1.3
0 9.0

2.0
5.0
0
9.0
1.4









(15)

the triplet format is shown in Tables 5 and 6.
row
iRow[0]
iRow[1]
iRow[2]
iRow[3]
iRow[4]
iRow[5]
iRow[6]
iRow[7]
iRow[8]
iRow[9]

=
=
=
=
=
=
=
=
=
=

1
2
3
3
4
4
5
5
5
5

col
jCol[0]
jCol[1]
jCol[2]
jCol[3]
jCol[4]
jCol[5]
jCol[6]
jCol[7]
jCol[8]
jCol[9]

=
=
=
=
=
=
=
=
=
=

1
1
1
3
3
4
1
2
4
5

value
values[0] =
values[1] =
values[2] =
values[3] =
values[4] =
values[5] =
values[6] =
values[7] =
values[8] =
values[9] =

1.0
1.1
3.0
1.2
6.0
1.3
2.0
5.0
9.0
1.4

Table 5: Triplet Format of Matrix (15) with index style=FORTRAN STYLE

row
iRow[0]
iRow[1]
iRow[2]
iRow[3]
iRow[4]
iRow[5]
iRow[6]
iRow[7]
iRow[8]
iRow[9]

=
=
=
=
=
=
=
=
=
=

0
1
2
2
3
3
4
4
4
4

col
jCol[0]
jCol[1]
jCol[2]
jCol[3]
jCol[4]
jCol[5]
jCol[6]
jCol[7]
jCol[8]
jCol[9]

=
=
=
=
=
=
=
=
=
=

0
0
0
2
2
3
0
1
3
4

value
values[0] =
values[1] =
values[2] =
values[3] =
values[4] =
values[5] =
values[6] =
values[7] =
values[8] =
values[9] =

1.0
1.1
3.0
1.2
6.0
1.3
2.0
5.0
9.0
1.4

Table 6: Triplet Format of Matrix (15) with index style=C STYLE

The Smart Pointer Implementation: SmartPtr

The SmartPtr class is described in IpSmartPtr.hpp. It is a template class that takes care of counting
references to objects and deleting them when not references anymore. Instead of pointing to an object
with a raw C++ pointer (e.g. HS071 NLP*), we use a SmartPtr. Every time a SmartPtr is set to
reference an object, it increments a counter in that object (see the ReferencedObject base class if you
are interested). If a SmartPtr is done with the object, either by leaving scope or being set to point to
another object, the counter is decremented. When the count of the object goes to zero, the object is
automatically deleted. SmartPtr’s are very simple, just use them as you would use a standard pointer.
It is very important to use SmartPtr’s instead of raw pointers when passing objects to Ipopt. Internally, Ipopt uses smart pointers for referencing objects. If you use a raw pointer in your executable, the
object’s counter will NOT get incremented. Then, when Ipopt uses smart pointers inside its own code,
the counter will get incremented. However, before Ipopt returns control to your code, it will decrement
66

as many times as it incremented earlier, and the counter will return to zero. Therefore, Ipopt will delete
the object. When control returns to you, you now have a raw pointer that points to a deleted object.
This might sound difficult to anyone not familiar with the use of smart pointers, but just follow one
simple rule; always use a SmartPtr when creating or passing an Ipopt object.

Options Reference

Options can be set using ipopt.opt, through your own code, or through the AMPL ipopt options
command. See Section 5 for an explanation of how to use these commands. Shown here is a list
of the most important options for Ipopt. To view the full list of options, you can set the option
print options documentation to yes or simply run the Ipopt AMPL solver executable as
ipopt --print-options
Usually, option values are identical for the regular mode of Ipopt and the restoration phase. However, to set an option value specifically for the restoration phase, the prefix “resto.” should be
appended. For example, to set the acceptable tolerance for the restoration phase, use the keyword
“resto.acceptable tol”.
The most common options are:

C.1

Output

print level: Output verbosity level.
Sets the default verbosity level for console output. The larger this value the more detailed is the output.
The valid range for this integer option is 0 ≤ print level ≤ 12 and its default value is 5.
print user options: Print all options set by the user.
If selected, the algorithm will print the list of all options set by the user including their values and whether
they have been used. In some cases this information might be incorrect, due to the internal program
flow. The default value for this string option is ”no”.
Possible values:
• no: don’t print options
• yes: print options
print options documentation: Switch to print all algorithmic options.
If selected, the algorithm will print the list of all available algorithmic options with some documentation
before solving the optimization problem. The default value for this string option is ”no”.
Possible values:
• no: don’t print list
• yes: print list
print frequency iter: Determines at which iteration frequency the summarizing iteration output line
should be printed.
Summarizing iteration output is printed every print frequency iter iterations, if at least print frequency time
seconds have passed since last output. The valid range for this integer option is 1 ≤ print frequency iter <
+inf and its default value is 1.

print frequency time: Determines at which time frequency the summarizing iteration output line
should be printed.
Summarizing iteration output is printed if at least print frequency time seconds have passed since last
output and the iteration number is a multiple of print frequency iter. The valid range for this real option
is 0 ≤ print frequency time < +inf and its default value is 0.
output file: File name of desired output file (leave unset for no file output).
NOTE: This option only works when read from the ipopt.opt options file! An output file with this name
will be written (leave unset for no file output). The verbosity level is by default set to ”print level”, but
can be overridden with ”file print level”. The file name is changed to use only small letters. The default
value for this string option is ””.
Possible values:
• *: Any acceptable standard file name
file print level: Verbosity level for output file.
NOTE: This option only works when read from the ipopt.opt options file! Determines the verbosity level
for the file specified by ”output file”. By default it is the same as ”print level”. The valid range for this
integer option is 0 ≤ file print level ≤ 12 and its default value is 5.
option file name: File name of options file.
By default, the name of the Ipopt options file is ”ipopt.opt” - or something else if specified in the
IpoptApplication::Initialize call. If this option is set by SetStringValue BEFORE the options file is read,
it specifies the name of the options file. It does not make any sense to specify this option within the
options file. Setting this option to an empty string disables reading of an options file. The default value
for this string option is ”ipopt.opt”.
Possible values:
• *: Any acceptable standard file name
print info string: Enables printing of additional info string at end of iteration output.
This string contains some insider information about the current iteration. For details, look for ”Diagnostic
Tags” in the Ipopt documentation. The default value for this string option is ”no”.
Possible values:
• no: don’t print string
• yes: print string at end of each iteration output
inf pr output: Determines what value is printed in the ”inf pr” output column.
Ipopt works with a reformulation of the original problem, where slacks are introduced and the problem
might have been scaled. The choice ”internal” prints out the constraint violation of this formulation.
With ”original” the true constraint violation in the original NLP is printed. The default value for this
string option is ”original”.
Possible values:
• internal: max-norm of violation of internal equality constraints
• original: maximal constraint violation in original NLP

print timing statistics: Switch to print timing statistics.
If selected, the program will print the CPU usage (user time) for selected tasks. The default value for
this string option is ”no”.
Possible values:
• no: don’t print statistics
• yes: print all timing statistics

C.2

Termination

tol: Desired convergence tolerance (relative).
Determines the convergence tolerance for the algorithm. The algorithm terminates successfully, if the
(scaled) NLP error becomes smaller than this value, and if the (absolute) criteria according to ”dual inf tol”,
”primal inf tol”, and ”compl inf tol” are met. (This is epsilon tol in Eqn. (6) in implementation paper).
See also ”acceptable tol” as a second termination criterion. Note, some other algorithmic features also
use this quantity to determine thresholds etc. The valid range for this real option is 0 < tol < +inf
and its default value is 1 · 10−08 .
max iter: Maximum number of iterations.
The algorithm terminates with an error message if the number of iterations exceeded this number. The
valid range for this integer option is 0 ≤ max iter < +inf and its default value is 3000.
max cpu time: Maximum number of CPU seconds.
A limit on CPU seconds that Ipopt can use to solve one problem. If during the convergence check this
limit is exceeded, Ipopt will terminate with a corresponding error message. The valid range for this real
option is 0 < max cpu time < +inf and its default value is 1 · 10+06 .
dual inf tol: Desired threshold for the dual infeasibility.
Absolute tolerance on the dual infeasibility. Successful termination requires that the max-norm of the
(unscaled) dual infeasibility is less than this threshold. The valid range for this real option is 0 <
dual inf tol < +inf and its default value is 1.
constr viol tol: Desired threshold for the constraint violation.
Absolute tolerance on the constraint violation. Successful termination requires that the max-norm of
the (unscaled) constraint violation is less than this threshold. The valid range for this real option is
0 < constr viol tol < +inf and its default value is 0.0001.
compl inf tol: Desired threshold for the complementarity conditions.
Absolute tolerance on the complementarity. Successful termination requires that the max-norm of the
(unscaled) complementarity is less than this threshold. The valid range for this real option is 0 <
compl inf tol < +inf and its default value is 0.0001.
acceptable tol: ”Acceptable” convergence tolerance (relative).
Determines which (scaled) overall optimality error is considered to be ”acceptable.” There are two levels
of termination criteria. If the usual ”desired” tolerances (see tol, dual inf tol etc) are satisfied at an
iteration, the algorithm immediately terminates with a success message. On the other hand, if the
algorithm encounters ”acceptable iter” many iterations in a row that are considered ”acceptable”, it will
terminate before the desired convergence tolerance is met. This is useful in cases where the algorithm
might not be able to achieve the ”desired” level of accuracy. The valid range for this real option is
0 < acceptable tol < +inf and its default value is 1 · 10−06 .

acceptable iter: Number of ”acceptable” iterates before triggering termination.
If the algorithm encounters this many successive ”acceptable” iterates (see ”acceptable tol”), it terminates, assuming that the problem has been solved to best possible accuracy given round-off. If it is set to
zero, this heuristic is disabled. The valid range for this integer option is 0 ≤ acceptable iter < +inf
and its default value is 15.
acceptable constr viol tol: ”Acceptance” threshold for the constraint violation.
Absolute tolerance on the constraint violation. ”Acceptable” termination requires that the max-norm of
the (unscaled) constraint violation is less than this threshold; see also acceptable tol. The valid range for
this real option is 0 < acceptable constr viol tol < +inf and its default value is 0.01.
acceptable dual inf tol: ”Acceptance” threshold for the dual infeasibility.
Absolute tolerance on the dual infeasibility. ”Acceptable” termination requires that the (max-norm of
the unscaled) dual infeasibility is less than this threshold; see also acceptable tol. The valid range for
this real option is 0 < acceptable dual inf tol < +inf and its default value is 1 · 10+10 .
acceptable compl inf tol: ”Acceptance” threshold for the complementarity conditions.
Absolute tolerance on the complementarity. ”Acceptable” termination requires that the max-norm of the
(unscaled) complementarity is less than this threshold; see also acceptable tol. The valid range for this
real option is 0 < acceptable compl inf tol < +inf and its default value is 0.01.
acceptable obj change tol: ”Acceptance” stopping criterion based on objective function change.
If the relative change of the objective function (scaled by Max(1,—f(x)—)) is less than this value, this
part of the acceptable tolerance termination is satisfied; see also acceptable tol. This is useful for the
quasi-Newton option, which has trouble to bring down the dual infeasibility. The valid range for this real
option is 0 ≤ acceptable obj change tol < +inf and its default value is 1 · 10+20 .
diverging iterates tol: Threshold for maximal value of primal iterates.
If any component of the primal iterates exceeded this value (in absolute terms), the optimization is
aborted with the exit message that the iterates seem to be diverging. The valid range for this real option
is 0 < diverging iterates tol < +inf and its default value is 1 · 10+20 .

C.3

NLP Scaling

obj scaling factor: Scaling factor for the objective function.
This option sets a scaling factor for the objective function. The scaling is seen internally by Ipopt but
the unscaled objective is reported in the console output. If additional scaling parameters are computed
(e.g. user-scaling or gradient-based), both factors are multiplied. If this value is chosen to be negative,
Ipopt will maximize the objective function instead of minimizing it. The valid range for this real option
is −inf < obj scaling factor < +inf and its default value is 1.
nlp scaling method: Select the technique used for scaling the NLP.
Selects the technique used for scaling the problem internally before it is solved. For user-scaling, the
parameters come from the NLP. If you are using AMPL, they can be specified through suffixes (”scaling factor”) The default value for this string option is ”gradient-based”.
Possible values:
• none: no problem scaling will be performed
• user-scaling: scaling parameters will come from the user
• gradient-based: scale the problem so the maximum gradient at the starting point is scaling max gradient

• equilibration-based: scale the problem so that first derivatives are of order 1 at random points (only
available with MC19)
nlp scaling max gradient: Maximum gradient after NLP scaling.
This is the gradient scaling cut-off. If the maximum gradient is above this value, then gradient based
scaling will be performed. Scaling parameters are calculated to scale the maximum gradient back to
this value. (This is g max in Section 3.8 of the implementation paper.) Note: This option is only
used if ”nlp scaling method” is chosen as ”gradient-based”. The valid range for this real option is 0 <
nlp scaling max gradient < +inf and its default value is 100.
nlp scaling min value: Minimum value of gradient-based scaling values.
This is the lower bound for the scaling factors computed by gradient-based scaling method. If some
derivatives of some functions are huge, the scaling factors will otherwise become very small, and the
(unscaled) final constraint violation, for example, might then be significant. Note: This option is only
used if ”nlp scaling method” is chosen as ”gradient-based”. The valid range for this real option is 0 ≤
nlp scaling min value < +inf and its default value is 1 · 10−08 .

C.4

NLP

bound relax factor: Factor for initial relaxation of the bounds.
Before start of the optimization, the bounds given by the user are relaxed. This option sets the factor for
this relaxation. If it is set to zero, then then bounds relaxation is disabled. (See Eqn.(35) in implementation paper.) The valid range for this real option is 0 ≤ bound relax factor < +inf and its default
value is 1 · 10−08 .
honor original bounds: Indicates whether final points should be projected into original bounds.
Ipopt might relax the bounds during the optimization (see, e.g., option ”bound relax factor”). This
option determines whether the final point should be projected back into the user-provide original bounds
after the optimization. The default value for this string option is ”yes”.
Possible values:
• no: Leave final point unchanged
• yes: Project final point back into original bounds
check derivatives for naninf: Indicates whether it is desired to check for Nan/Inf in derivative matrices
Activating this option will cause an error if an invalid number is detected in the constraint Jacobians or
the Lagrangian Hessian. If this is not activated, the test is skipped, and the algorithm might proceed with
invalid numbers and fail. If test is activated and an invalid number is detected, the matrix is written to
output with print level corresponding to J MORE DETAILED; so beware of large output! The default
value for this string option is ”no”.
Possible values:
• no: Don’t check (faster).
• yes: Check Jacobians and Hessian for Nan and Inf.
nlp lower bound inf: any bound less or equal this value will be considered -inf (i.e. not lower
bounded).
The valid range for this real option is −inf < nlp lower bound inf < +inf and its default value is
−1 · 10+19 .

nlp upper bound inf: any bound greater or this value will be considered +inf (i.e. not upper bounded).
The valid range for this real option is −inf < nlp upper bound inf < +inf and its default value is
1 · 10+19 .
fixed variable treatment: Determines how fixed variables should be handled.
The main difference between those options is that the starting point in the ”make constraint” case
still has the fixed variables at their given values, whereas in the case ”make parameter” the functions
are always evaluated with the fixed values for those variables. Also, for ”relax bounds”, the fixing
bound constraints are relaxed (according to” bound relax factor”). For both ”make constraints” and
”relax bounds”, bound multipliers are computed for the fixed variables. The default value for this string
option is ”make parameter”.
Possible values:
• make parameter: Remove fixed variable from optimization variables
• make constraint: Add equality constraints fixing variables
• relax bounds: Relax fixing bound constraints
jac c constant: Indicates whether all equality constraints are linear
Activating this option will cause Ipopt to ask for the Jacobian of the equality constraints only once from
the NLP and reuse this information later. The default value for this string option is ”no”.
Possible values:
• no: Don’t assume that all equality constraints are linear
• yes: Assume that equality constraints Jacobian are constant
jac d constant: Indicates whether all inequality constraints are linear
Activating this option will cause Ipopt to ask for the Jacobian of the inequality constraints only once
from the NLP and reuse this information later. The default value for this string option is ”no”.
Possible values:
• no: Don’t assume that all inequality constraints are linear
• yes: Assume that equality constraints Jacobian are constant
hessian constant: Indicates whether the problem is a quadratic problem
Activating this option will cause Ipopt to ask for the Hessian of the Lagrangian function only once from
the NLP and reuse this information later. The default value for this string option is ”no”.
Possible values:
• no: Assume that Hessian changes
• yes: Assume that Hessian is constant

C.5

Initialization

bound frac: Desired minimum relative distance from the initial point to bound.
Determines how much the initial point might have to be modified in order to be sufficiently inside the
bounds (together with ”bound push”). (This is kappa 2 in Section 3.6 of implementation paper.) The
valid range for this real option is 0 < bound frac ≤ 0.5 and its default value is 0.01.

bound push: Desired minimum absolute distance from the initial point to bound.
Determines how much the initial point might have to be modified in order to be sufficiently inside the
bounds (together with ”bound frac”). (This is kappa 1 in Section 3.6 of implementation paper.) The
valid range for this real option is 0 < bound push < +inf and its default value is 0.01.
slack bound frac: Desired minimum relative distance from the initial slack to bound.
Determines how much the initial slack variables might have to be modified in order to be sufficiently
inside the inequality bounds (together with ”slack bound push”). (This is kappa 2 in Section 3.6 of
implementation paper.) The valid range for this real option is 0 < slack bound frac ≤ 0.5 and its
default value is 0.01.
slack bound push: Desired minimum absolute distance from the initial slack to bound.
Determines how much the initial slack variables might have to be modified in order to be sufficiently
inside the inequality bounds (together with ”slack bound frac”). (This is kappa 1 in Section 3.6 of
implementation paper.) The valid range for this real option is 0 < slack bound push < +inf and its
default value is 0.01.
bound mult init val: Initial value for the bound multipliers.
All dual variables corresponding to bound constraints are initialized to this value. The valid range for
this real option is 0 < bound mult init val < +inf and its default value is 1.
constr mult init max: Maximum allowed least-square guess of constraint multipliers.
Determines how large the initial least-square guesses of the constraint multipliers are allowed to be
(in max-norm). If the guess is larger than this value, it is discarded and all constraint multipliers are set to zero. This options is also used when initializing the restoration phase. By default,
”resto.constr mult init max” (the one used in RestoIterateInitializer) is set to zero. The valid range
for this real option is 0 ≤ constr mult init max < +inf and its default value is 1000.
bound mult init method: Initialization method for bound multipliers
This option defines how the iterates for the bound multipliers are initialized. If ”constant” is chosen,
then all bound multipliers are initialized to the value of ”bound mult init val”. If ”mu-based” is chosen,
the each value is initialized to the the value of ”mu init” divided by the corresponding slack variable.
This latter option might be useful if the starting point is close to the optimal solution. The default value
for this string option is ”constant”.
Possible values:
• constant: set all bound multipliers to the value of bound mult init val
• mu-based: initialize to mu init/x slack

C.6

Barrier Parameter

mehrotra algorithm: Indicates if we want to do Mehrotra’s algorithm.
If set to yes, Ipopt runs as Mehrotra’s predictor-corrector algorithm. This works usually very well for
LPs and convex QPs. This automatically disables the line search, and chooses the (unglobalized) adaptive mu strategy with the ”probing” oracle, and uses ”corrector type=affine” without any safeguards;
you should not set any of those options explicitly in addition. Also, unless otherwise specified, the
values of ”bound push”, ”bound frac”, and ”bound mult init val” are set more aggressive, and sets ”alpha for y=bound mult”. The default value for this string option is ”no”.
Possible values:
• no: Do the usual Ipopt algorithm.
• yes: Do Mehrotra’s predictor-corrector algorithm.
73

mu strategy: Update strategy for barrier parameter.
Determines which barrier parameter update strategy is to be used. The default value for this string
option is ”monotone”.
Possible values:
• monotone: use the monotone (Fiacco-McCormick) strategy
• adaptive: use the adaptive update strategy
mu oracle: Oracle for a new barrier parameter in the adaptive strategy.
Determines how a new barrier parameter is computed in each ”free-mode” iteration of the adaptive barrier
parameter strategy. (Only considered if ”adaptive” is selected for option ”mu strategy”). The default
value for this string option is ”quality-function”.
Possible values:
• probing: Mehrotra’s probing heuristic
• loqo: LOQO’s centrality rule
• quality-function: minimize a quality function

quality function max section steps: Maximum number of search steps during direct search procedure determining the optimal centering parameter.
The golden section search is performed for the quality function based mu oracle. (Only used if option
”mu oracle” is set to ”quality-function”.) The valid range for this integer option is 0 ≤ quality function max section step
+inf and its default value is 8.
fixed mu oracle: Oracle for the barrier parameter when switching to fixed mode.
Determines how the first value of the barrier parameter should be computed when switching to the ”monotone mode” in the adaptive strategy. (Only considered if ”adaptive” is selected for option ”mu strategy”.)
The default value for this string option is ”average compl”.
Possible values:
• probing: Mehrotra’s probing heuristic
• loqo: LOQO’s centrality rule
• quality-function: minimize a quality function
• average compl: base on current average complementarity
adaptive mu globalization: Globalization strategy for the adaptive mu selection mode.
To achieve global convergence of the adaptive version, the algorithm has to switch to the monotone mode
(Fiacco-McCormick approach) when convergence does not seem to appear. This option sets the criterion
used to decide when to do this switch. (Only used if option ”mu strategy” is chosen as ”adaptive”.) The
default value for this string option is ”obj-constr-filter”.
Possible values:
• kkt-error: nonmonotone decrease of kkt-error
• obj-constr-filter: 2-dim filter for objective and constraint violation
• never-monotone-mode: disables globalization

mu init: Initial value for the barrier parameter.
This option determines the initial value for the barrier parameter (mu). It is only relevant in the monotone,
Fiacco-McCormick version of the algorithm. (i.e., if ”mu strategy” is chosen as ”monotone”) The valid
range for this real option is 0 < mu init < +inf and its default value is 0.1.
mu max fact: Factor for initialization of maximum value for barrier parameter.
This option determines the upper bound on the barrier parameter. This upper bound is computed as
the average complementarity at the initial point times the value of this option. (Only used if option
”mu strategy” is chosen as ”adaptive”.) The valid range for this real option is 0 < mu max fact < +inf
and its default value is 1000.
mu max: Maximum value for barrier parameter.
This option specifies an upper bound on the barrier parameter in the adaptive mu selection mode. If this
option is set, it overwrites the effect of mu max fact. (Only used if option ”mu strategy” is chosen as
”adaptive”.) The valid range for this real option is 0 < mu max < +inf and its default value is 100000.
mu min: Minimum value for barrier parameter.
This option specifies the lower bound on the barrier parameter in the adaptive mu selection mode. By
default, it is set to the minimum of 1e-11 and min(”tol”,”compl inf tol”)/(”barrier tol factor”+1), which
should be a reasonable value. (Only used if option ”mu strategy” is chosen as ”adaptive”.) The valid
range for this real option is 0 < mu min < +inf and its default value is 1 · 10−11 .
mu target: Desired value of complementarity.
Usually, the barrier parameter is driven to zero and the termination test for complementarity is measured
with respect to zero complementarity. However, in some cases it might be desired to have Ipopt solve
barrier problem for strictly positive value of the barrier parameter. In this case, the value of ”mu target”
specifies the final value of the barrier parameter, and the termination tests are then defined with respect
to the barrier problem for this value of the barrier parameter. The valid range for this real option is
0 ≤ mu target < +inf and its default value is 0.
barrier tol factor: Factor for mu in barrier stop test.
The convergence tolerance for each barrier problem in the monotone mode is the value of the barrier
parameter times ”barrier tol factor”. This option is also used in the adaptive mu strategy during the
monotone mode. (This is kappa epsilon in implementation paper). The valid range for this real option
is 0 < barrier tol factor < +inf and its default value is 10.
mu linear decrease factor: Determines linear decrease rate of barrier parameter.
For the Fiacco-McCormick update procedure the new barrier parameter mu is obtained by taking the
minimum of mu*”mu linear decrease factor” and muˆ”superlinear decrease power”. (This is kappa mu in
implementation paper.) This option is also used in the adaptive mu strategy during the monotone mode.
The valid range for this real option is 0 < mu linear decrease factor < 1 and its default value is 0.2.
mu superlinear decrease power: Determines superlinear decrease rate of barrier parameter.
For the Fiacco-McCormick update procedure the new barrier parameter mu is obtained by taking the
minimum of mu*”mu linear decrease factor” and muˆ”superlinear decrease power”. (This is theta mu in
implementation paper.) This option is also used in the adaptive mu strategy during the monotone mode.
The valid range for this real option is 1 < mu superlinear decrease power < 2 and its default value is
1.5.

C.7

Multiplier Updates

alpha for y: Method to determine the step size for constraint multipliers.
This option determines how the step size (alpha y) will be calculated when updating the constraint
multipliers. The default value for this string option is ”primal”.
Possible values:
• primal: use primal step size
• bound-mult: use step size for the bound multipliers (good for LPs)
• min: use the min of primal and bound multipliers
• max: use the max of primal and bound multipliers
• full: take a full step of size one
• min-dual-infeas: choose step size minimizing new dual infeasibility
• safer-min-dual-infeas: like ”min dual infeas”, but safeguarded by ”min” and ”max”
• primal-and-full: use the primal step size, and full step if delta x ¡= alpha for y tol
• dual-and-full: use the dual step size, and full step if delta x ¡= alpha for y tol
• acceptor: Call LSAcceptor to get step size for y
alpha for y tol: Tolerance for switching to full equality multiplier steps.
This is only relevant if ”alpha for y” is chosen ”primal-and-full” or ”dual-and-full”. The step size for
the equality constraint multipliers is taken to be one if the max-norm of the primal step is less than this
tolerance. The valid range for this real option is 0 ≤ alpha for y tol < +inf and its default value is
10.
recalc y: Tells the algorithm to recalculate the equality and inequality multipliers as least square
estimates.
This asks the algorithm to recompute the multipliers, whenever the current infeasibility is less than
recalc y feas tol. Choosing yes might be helpful in the quasi-Newton option. However, each recalculation
requires an extra factorization of the linear system. If a limited memory quasi-Newton option is chosen,
this is used by default. The default value for this string option is ”no”.
Possible values:
• no: use the Newton step to update the multipliers
• yes: use least-square multiplier estimates
recalc y feas tol: Feasibility threshold for recomputation of multipliers.
If recalc y is chosen and the current infeasibility is less than this value, then the multipliers are recomputed. The valid range for this real option is 0 < recalc y feas tol < +inf and its default value is
1 · 10−06 .

C.8

Line Search

max soc: Maximum number of second order correction trial steps at each iteration.
Choosing 0 disables the second order corrections. (This is pm̂ax of Step A-5.9 of Algorithm A in the
implementation paper.) The valid range for this integer option is 0 ≤ max soc < +inf and its default
value is 4.

watchdog shortened iter trigger: Number of shortened iterations that trigger the watchdog.
If the number of successive iterations in which the backtracking line search did not accept the first trial
point exceeds this number, the watchdog procedure is activated. Choosing ”0” here disables the watchdog
procedure. The valid range for this integer option is 0 ≤ watchdog shortened iter trigger < +inf
and its default value is 10.
watchdog trial iter max: Maximum number of watchdog iterations.
This option determines the number of trial iterations allowed before the watchdog procedure is aborted
and the algorithm returns to the stored point. The valid range for this integer option is 1 ≤ watchdog trial iter max <
+inf and its default value is 3.
accept every trial step: Always accept the first trial step.
Setting this option to ”yes” essentially disables the line search and makes the algorithm take aggressive
steps, without global convergence guarantees. The default value for this string option is ”no”.
Possible values:
• no: don’t arbitrarily accept the full step
• yes: always accept the full step
corrector type: The type of corrector steps that should be taken (unsupported!).
If ”mu strategy” is ”adaptive”, this option determines what kind of corrector steps should be tried. The
default value for this string option is ”none”.
Possible values:
• none: no corrector
• affine: corrector step towards mu=0
• primal-dual: corrector step towards current mu

C.9

Warm Start

warm start init point: Warm-start for initial point
Indicates whether this optimization should use a warm start initialization, where values of primal and
dual variables are given (e.g., from a previous optimization of a related problem.) The default value for
this string option is ”no”.
Possible values:
• no: do not use the warm start initialization
• yes: use the warm start initialization
warm start bound push: same as bound push for the regular initializer.
The valid range for this real option is 0 < warm start bound push < +inf and its default value is 0.001.
warm start bound frac: same as bound frac for the regular initializer.
The valid range for this real option is 0 < warm start bound frac ≤ 0.5 and its default value is 0.001.
warm start slack bound frac: same as slack bound frac for the regular initializer.
The valid range for this real option is 0 < warm start slack bound frac ≤ 0.5 and its default value is
0.001.

warm start slack bound push: same as slack bound push for the regular initializer.
The valid range for this real option is 0 < warm start slack bound push < +inf and its default value
is 0.001.
warm start mult bound push: same as mult bound push for the regular initializer.
The valid range for this real option is 0 < warm start mult bound push < +inf and its default value is
0.001.
warm start mult init max: Maximum initial value for the equality multipliers.
The valid range for this real option is −inf < warm start mult init max < +inf and its default value
is 1 · 10+06 .

C.10

Restoration Phase

expect infeasible problem: Enable heuristics to quickly detect an infeasible problem.
This options is meant to activate heuristics that may speed up the infeasibility determination if you
expect that there is a good chance for the problem to be infeasible. In the filter line search procedure, the
restoration phase is called more quickly than usually, and more reduction in the constraint violation is
enforced before the restoration phase is left. If the problem is square, this option is enabled automatically.
The default value for this string option is ”no”.
Possible values:
• no: the problem probably be feasible
• yes: the problem has a good chance to be infeasible
expect infeasible problem ctol: Threshold for disabling ”expect infeasible problem” option.
If the constraint violation becomes smaller than this threshold, the ”expect infeasible problem” heuristics
in the filter line search are disabled. If the problem is square, this options is set to 0. The valid range for
this real option is 0 ≤ expect infeasible problem ctol < +inf and its default value is 0.001.
expect infeasible problem ytol: Multiplier threshold for activating ”expect infeasible problem” option.
If the max norm of the constraint multipliers becomes larger than this value and ”expect infeasible problem”
is chosen, then the restoration phase is entered. The valid range for this real option is 0 < expect infeasible problem ytol
+inf and its default value is 1 · 10+08 .
start with resto: Tells algorithm to switch to restoration phase in first iteration.
Setting this option to ”yes” forces the algorithm to switch to the feasibility restoration phase in the first
iteration. If the initial point is feasible, the algorithm will abort with a failure. The default value for this
string option is ”no”.
Possible values:
• no: don’t force start in restoration phase
• yes: force start in restoration phase
soft resto pderror reduction factor: Required reduction in primal-dual error in the soft restoration
phase.
The soft restoration phase attempts to reduce the primal-dual error with regular steps. If the damped
primal-dual step (damped only to satisfy the fraction-to-the-boundary rule) is not decreasing the primaldual error by at least this factor, then the regular restoration phase is called. Choosing ”0” here disables
the soft restoration phase. The valid range for this real option is 0 ≤ soft resto pderror reduction factor <
+inf and its default value is 0.9999.
78

required infeasibility reduction: Required reduction of infeasibility before leaving restoration phase.
The restoration phase algorithm is performed, until a point is found that is acceptable to the filter and
the infeasibility has been reduced by at least the fraction given by this option. The valid range for this
real option is 0 ≤ required infeasibility reduction < 1 and its default value is 0.9.
bound mult reset threshold: Threshold for resetting bound multipliers after the restoration phase.
After returning from the restoration phase, the bound multipliers are updated with a Newton step for
complementarity. Here, the change in the primal variables during the entire restoration phase is taken
to be the corresponding primal Newton step. However, if after the update the largest bound multiplier
exceeds the threshold specified by this option, the multipliers are all reset to 1. The valid range for this
real option is 0 ≤ bound mult reset threshold < +inf and its default value is 1000.
constr mult reset threshold: Threshold for resetting equality and inequality multipliers after restoration phase.
After returning from the restoration phase, the constraint multipliers are recomputed by a least square
estimate. This option triggers when those least-square estimates should be ignored. The valid range for
this real option is 0 ≤ constr mult reset threshold < +inf and its default value is 0.
evaluate orig obj at resto trial: Determines if the original objective function should be evaluated
at restoration phase trial points.
Setting this option to ”yes” makes the restoration phase algorithm evaluate the objective function of the
original problem at every trial point encountered during the restoration phase, even if this value is not
required. In this way, it is guaranteed that the original objective function can be evaluated without error
at all accepted iterates; otherwise the algorithm might fail at a point where the restoration phase accepts
an iterate that is good for the restoration phase problem, but not the original problem. On the other
hand, if the evaluation of the original objective is expensive, this might be costly. The default value for
this string option is ”yes”.
Possible values:
• no: skip evaluation
• yes: evaluate at every trial point

C.11

Linear Solver

linear solver: Linear solver used for step computations.
Determines which linear algebra package is to be used for the solution of the augmented linear system
(for obtaining the search directions). Note, the code must have been compiled with the linear solver you
want to choose. Depending on your Ipopt installation, not all options are available. The default value
for this string option is ”ma27”.
Possible values:
• ma27: use the Harwell routine MA27
• ma57: use the Harwell routine MA57
• ma77: use the Harwell routine HSL MA77
• ma86: use the Harwell routine HSL MA86
• ma97: use the Harwell routine HSL MA97
• pardiso: use the Pardiso package
• wsmp: use WSMP package

• mumps: use MUMPS package
• custom: use custom linear solver
linear system scaling: Method for scaling the linear system.
Determines the method used to compute symmetric scaling factors for the augmented system (see also
the ”linear scaling on demand” option). This scaling is independent of the NLP problem scaling. By
default, MC19 is only used if MA27 or MA57 are selected as linear solvers. This value is only available
if Ipopt has been compiled with MC19. The default value for this string option is ”mc19”.
Possible values:
• none: no scaling will be performed
• mc19: use the Harwell routine MC19
• slack-based: use the slack values
linear scaling on demand: Flag indicating that linear scaling is only done if it seems required.
This option is only important if a linear scaling method (e.g., mc19) is used. If you choose ”no”, then
the scaling factors are computed for every linear system from the start. This can be quite expensive.
Choosing ”yes” means that the algorithm will start the scaling method only when the solutions to the
linear system seem not good, and then use it until the end. The default value for this string option is
”yes”.
Possible values:
• no: Always scale the linear system.
• yes: Start using linear system scaling if solutions seem not good.
max refinement steps: Maximum number of iterative refinement steps per linear system solve.
Iterative refinement (on the full unsymmetric system) is performed for each right hand side. This option
determines the maximum number of iterative refinement steps. The valid range for this integer option is
0 ≤ max refinement steps < +inf and its default value is 10.
min refinement steps: Minimum number of iterative refinement steps per linear system solve.
Iterative refinement (on the full unsymmetric system) is performed for each right hand side. This option determines the minimum number of iterative refinements (i.e. at least ”min refinement steps” iterative refinement steps are enforced per right hand side.) The valid range for this integer option is
0 ≤ min refinement steps < +inf and its default value is 1.

C.12

Hessian Perturbation

max hessian perturbation: Maximum value of regularization parameter for handling negative curvature.
In order to guarantee that the search directions are indeed proper descent directions, Ipopt requires that
the inertia of the (augmented) linear system for the step computation has the correct number of negative
and positive eigenvalues. The idea is that this guides the algorithm away from maximizers and makes
Ipopt more likely converge to first order optimal points that are minimizers. If the inertia is not correct, a
multiple of the identity matrix is added to the Hessian of the Lagrangian in the augmented system. This
parameter gives the maximum value of the regularization parameter. If a regularization of that size is not
enough, the algorithm skips this iteration and goes to the restoration phase. (This is delta wm̂ax in the
implementation paper.) The valid range for this real option is 0 < max hessian perturbation < +inf
and its default value is 1 · 10+20 .

min hessian perturbation: Smallest perturbation of the Hessian block.
The size of the perturbation of the Hessian block is never selected smaller than this value, unless no
perturbation is necessary. (This is delta wm̂in in implementation paper.) The valid range for this real
option is 0 ≤ min hessian perturbation < +inf and its default value is 1 · 10−20 .
first hessian perturbation: Size of first x-s perturbation tried.
The first value tried for the x-s perturbation in the inertia correction scheme.(This is delta 0 in the
implementation paper.) The valid range for this real option is 0 < first hessian perturbation < +inf
and its default value is 0.0001.
perturb inc fact first: Increase factor for x-s perturbation for very first perturbation.
The factor by which the perturbation is increased when a trial value was not sufficient - this value is used
for the computation of the very first perturbation and allows a different value for for the first perturbation
than that used for the remaining perturbations. (This is bar kappa w+̂ in the implementation paper.)
The valid range for this real option is 1 < perturb inc fact first < +inf and its default value is 100.
perturb inc fact: Increase factor for x-s perturbation.
The factor by which the perturbation is increased when a trial value was not sufficient - this value is used
for the computation of all perturbations except for the first. (This is kappa w+̂ in the implementation
paper.) The valid range for this real option is 1 < perturb inc fact < +inf and its default value is 8.
perturb dec fact: Decrease factor for x-s perturbation.
The factor by which the perturbation is decreased when a trial value is deduced from the size of the most
recent successful perturbation. (This is kappa w-̂ in the implementation paper.) The valid range for this
real option is 0 < perturb dec fact < 1 and its default value is 0.333333.

jacobian regularization value: Size of the regularization for rank-deficient constraint Jacobians.
(This is bar delta c in the implementation paper.) The valid range for this real option is 0 ≤ jacobian regularization valu
+inf and its default value is 1 · 10−08 .

C.13

Quasi-Newton

hessian approximation: Indicates what Hessian information is to be used.
This determines which kind of information for the Hessian of the Lagrangian function is used by the
algorithm. The default value for this string option is ”exact”.
Possible values:
• exact: Use second derivatives provided by the NLP.
• limited-memory: Perform a limited-memory quasi-Newton approximation
limited memory update type: Quasi-Newton update formula for the limited memory approximation.
Determines which update formula is to be used for the limited-memory quasi-Newton approximation.
The default value for this string option is ”bfgs”.
Possible values:
• bfgs: BFGS update (with skipping)
• sr1: SR1 (not working well)

limited memory max history: Maximum size of the history for the limited quasi-Newton Hessian
approximation.
This option determines the number of most recent iterations that are taken into account for the limitedmemory quasi-Newton approximation. The valid range for this integer option is 0 ≤ limited memory max history <
+inf and its default value is 6.
limited memory max skipping: Threshold for successive iterations where update is skipped.
If the update is skipped more than this number of successive iterations, we quasi-Newton approximation
is reset. The valid range for this integer option is 1 ≤ limited memory max skipping < +inf and its
default value is 2.
limited memory initialization: Initialization strategy for the limited memory quasi-Newton approximation.
Determines how the diagonal Matrix B 0 as the first term in the limited memory approximation should
be computed. The default value for this string option is ”scalar1”.
Possible values:
• scalar1: sigma = sT̂y/sT̂s
• scalar2: sigma = yT̂y/sT̂y
• scalar3: arithmetic average of scalar1 and scalar2
• scalar4: geometric average of scalar1 and scalar2
• constant: sigma = limited memory init val
limited memory init val: Value for B0 in low-rank update.
The starting matrix in the low rank update, B0, is chosen to be this multiple of the identity in the first
iteration (when no updates have been performed yet), and is constantly chosen as this value, if ”limited memory initialization” is ”constant”. The valid range for this real option is 0 < limited memory init val <
+inf and its default value is 1.
limited memory init val max: Upper bound on value for B0 in low-rank update.
The starting matrix in the low rank update, B0, is chosen to be this multiple of the identity in the first
iteration (when no updates have been performed yet), and is constantly chosen as this value, if ”limited memory initialization” is ”constant”. The valid range for this real option is 0 < limited memory init val max <
+inf and its default value is 1 · 10+08 .
limited memory init val min: Lower bound on value for B0 in low-rank update.
The starting matrix in the low rank update, B0, is chosen to be this multiple of the identity in the first
iteration (when no updates have been performed yet), and is constantly chosen as this value, if ”limited memory initialization” is ”constant”. The valid range for this real option is 0 < limited memory init val min <
+inf and its default value is 1 · 10−08 .
limited memory special for resto: Determines if the quasi-Newton updates should be special during
the restoration phase.
Until Nov 2010, Ipopt used a special update during the restoration phase, but it turned out that this
does not work well. The new default uses the regular update procedure and it improves results. If for
some reason you want to get back to the original update, set this option to ”yes”. The default value for
this string option is ”no”.
Possible values:
• no: use the same update as in regular iterations
82

• yes: use the a special update during restoration phase

C.14

Derivative Test

derivative test: Enable derivative checker
If this option is enabled, a (slow!) derivative test will be performed before the optimization. The test
is performed at the user provided starting point and marks derivative values that seem suspicious The
default value for this string option is ”none”.
Possible values:
• none: do not perform derivative test
• first-order: perform test of first derivatives at starting point
• second-order: perform test of first and second derivatives at starting point
• only-second-order: perform test of second derivatives at starting point
derivative test perturbation: Size of the finite difference perturbation in derivative test.
This determines the relative perturbation of the variable entries. The valid range for this real option is
0 < derivative test perturbation < +inf and its default value is 1 · 10−08 .
derivative test tol: Threshold for indicating wrong derivative.
If the relative deviation of the estimated derivative from the given one is larger than this value, the corresponding derivative is marked as wrong. The valid range for this real option is 0 < derivative test tol <
+inf and its default value is 0.0001.
derivative test print all: Indicates whether information for all estimated derivatives should be printed.
Determines verbosity of derivative checker. The default value for this string option is ”no”.
Possible values:
• no: Print only suspect derivatives
• yes: Print all derivatives
derivative test first index: Index of first quantity to be checked by derivative checker
If this is set to -2, then all derivatives are checked. Otherwise, for the first derivative test it specifies the
first variable for which the test is done (counting starts at 0). For second derivatives, it specifies the first
constraint for which the test is done; counting of constraint indices starts at 0, and -1 refers to the objective
function Hessian. The valid range for this integer option is −2 ≤ derivative test first index < +inf
and its default value is −2.
point perturbation radius: Maximal perturbation of an evaluation point.
If a random perturbation of a points is required, this number indicates the maximal perturbation. This
is for example used when determining the center point at which the finite difference derivative test is
executed. The valid range for this real option is 0 ≤ point perturbation radius < +inf and its default
value is 10.

C.15

MA27 Linear Solver

ma27 pivtol: Pivot tolerance for the linear solver MA27.
A smaller number pivots for sparsity, a larger number pivots for stability. This option is only available
if Ipopt has been compiled with MA27. The valid range for this real option is 0 < ma27 pivtol < 1 and
its default value is 1 · 10−08 .

ma27 pivtolmax: Maximum pivot tolerance for the linear solver MA27.
Ipopt may increase pivtol as high as pivtolmax to get a more accurate solution to the linear system. This
option is only available if Ipopt has been compiled with MA27. The valid range for this real option is
0 < ma27 pivtolmax < 1 and its default value is 0.0001.
ma27 liw init factor: Integer workspace memory for MA27.
The initial integer workspace memory = liw init factor * memory required by unfactored system. Ipopt
will increase the workspace size by meminc factor if required. This option is only available if Ipopt has
been compiled with MA27. The valid range for this real option is 1 ≤ ma27 liw init factor < +inf
and its default value is 5.
ma27 la init factor: Real workspace memory for MA27.
The initial real workspace memory = la init factor * memory required by unfactored system. Ipopt will
increase the workspace size by meminc factor if required. This option is only available if Ipopt has been
compiled with MA27. The valid range for this real option is 1 ≤ ma27 la init factor < +inf and its
default value is 5.
ma27 meminc factor: Increment factor for workspace size for MA27.
If the integer or real workspace is not large enough, Ipopt will increase its size by this factor. This
option is only available if Ipopt has been compiled with MA27. The valid range for this real option is
1 ≤ ma27 meminc factor < +inf and its default value is 2.

C.16

MA57 Linear Solver

ma57 pivtol: Pivot tolerance for the linear solver MA57.
A smaller number pivots for sparsity, a larger number pivots for stability. This option is only available
if Ipopt has been compiled with MA57. The valid range for this real option is 0 < ma57 pivtol < 1 and
its default value is 1 · 10−08 .
ma57 pivtolmax: Maximum pivot tolerance for the linear solver MA57.
Ipopt may increase pivtol as high as ma57 pivtolmax to get a more accurate solution to the linear system.
This option is only available if Ipopt has been compiled with MA57. The valid range for this real option
is 0 < ma57 pivtolmax < 1 and its default value is 0.0001.
ma57 pre alloc: Safety factor for work space memory allocation for the linear solver MA57.
If 1 is chosen, the suggested amount of work space is used. However, choosing a larger number might
avoid reallocation if the suggest values do not suffice. This option is only available if Ipopt has been
compiled with MA57. The valid range for this real option is 1 ≤ ma57 pre alloc < +inf and its default
value is 1.05.
ma57 pivot order: Controls pivot order in MA57
This is ICNTL(6) in MA57. The valid range for this integer option is 0 ≤ ma57 pivot order ≤ 5 and its
default value is 5.
ma57 automatic scaling: Controls MA57 automatic scaling
This option controls the internal scaling option of MA57. For higher reliability of the MA57 solver, you
may want to set this option to yes. This is ICNTL(15) in MA57. The default value for this string option
is ”no”.
Possible values:
• no: Do not scale the linear system matrix
• yes: Scale the linear system matrix
84

ma57 block size: Controls block size used by Level 3 BLAS in MA57BD
This is ICNTL(11) in MA57. The valid range for this integer option is 1 ≤ ma57 block size < +inf
and its default value is 16.
ma57 node amalgamation: Node amalgamation parameter
This is ICNTL(12) in MA57. The valid range for this integer option is 1 ≤ ma57 node amalgamation <
+inf and its default value is 16.
ma57 small pivot flag: If set to 1, then when small entries defined by CNTL(2) are detected they
are removed and the corresponding pivots placed at the end of the factorization. This can be particularly
efficient if the matrix is highly rank deficient.
This is ICNTL(16) in MA57. The valid range for this integer option is 0 ≤ ma57 small pivot flag ≤ 1
and its default value is 0.

C.17

MA77 Linear Solver

ma77 print level: Debug printing level for the linear solver MA77
The valid range for this integer option is −inf < ma77 print level < +inf and its default value is −1.
ma77 buffer lpage: Number of scalars per MA77 buffer page
Number of scalars per an in-core buffer in the out-of-core solver MA77. Must be at most ma77 file size.
The valid range for this integer option is 1 ≤ ma77 buffer lpage < +inf and its default value is 4096.
ma77 buffer npage: Number of pages that make up MA77 buffer
Number of pages of size buffer lpage that exist in-core for the out-of-core solver MA77. The valid range
for this integer option is 1 ≤ ma77 buffer npage < +inf and its default value is 1600.
ma77 file size: Target size of each temporary file for MA77, scalars per type
MA77 uses many temporary files, this option controls the size of each one. It is measured in the number of
entries (int or double), NOT bytes. The valid range for this integer option is 1 ≤ ma77 file size < +inf
and its default value is 2097152.
ma77 maxstore: Maximum storage size for MA77 in-core mode
If greater than zero, the maximum size of factors stored in core before out-of-core mode is invoked. The
valid range for this integer option is 0 ≤ ma77 maxstore < +inf and its default value is 0.
ma77 nemin: Node Amalgamation parameter
Two nodes in elimination tree are merged if result has fewer than ma77 nemin variables. The valid range
for this integer option is 1 ≤ ma77 nemin < +inf and its default value is 8.
ma77 order: Controls type of ordering used by HSL MA77
This option controls ordering for the solver HSL MA77. The default value for this string option is ”metis”.
Possible values:
• amd: Use the HSL MC68 approximate minimum degree algorithm
• metis: Use the MeTiS nested dissection algorithm (if available)
ma77 small: Zero Pivot Threshold
Any pivot less than ma77 small is treated as zero. The valid range for this real option is 0 ≤ ma77 small <
+inf and its default value is 1 · 10−20 .

ma77 static: Static Pivoting Threshold
See MA77 documentation. Either ma77 static=0.0 or ma77 static¿ma77 small. ma77 static=0.0 disables
static pivoting. The valid range for this real option is 0 ≤ ma77 static < +inf and its default value is
0.
ma77 u: Pivoting Threshold
See MA77 documentation. The valid range for this real option is 0 ≤ ma77 u ≤ 0.5 and its default value
is 1 · 10−08 .
ma77 umax: Maximum Pivoting Threshold
Maximum value to which u will be increased to improve quality. The valid range for this real option is
0 ≤ ma77 umax ≤ 0.5 and its default value is 0.0001.

C.18

MA86 Linear Solver

ma86 print level: Debug printing level for the linear solver MA86
The valid range for this integer option is −inf < ma86 print level < +inf and its default value is −1.
ma86 nemin: Node Amalgamation parameter
Two nodes in elimination tree are merged if result has fewer than ma86 nemin variables. The valid range
for this integer option is 1 ≤ ma86 nemin < +inf and its default value is 32.
ma86 order: Controls type of ordering used by HSL MA86
This option controls ordering for the solver HSL MA86. The default value for this string option is ”auto”.
Possible values:
• auto: Try both AMD and MeTiS, pick best
• amd: Use the HSL MC68 approximate minimum degree algorithm
• metis: Use the MeTiS nested dissection algorithm (if available)
ma86 scaling: Controls scaling of matrix
This option controls scaling for the solver HSL MA86. The default value for this string option is ”mc64”.
Possible values:
• none: Do not scale the linear system matrix
• mc64: Scale linear system matrix using MC64
• mc77: Scale linear system matrix using MC77 [1,3,0]
ma86 small: Zero Pivot Threshold
Any pivot less than ma86 small is treated as zero. The valid range for this real option is 0 ≤ ma86 small <
+inf and its default value is 1 · 10−20 .
ma86 static: Static Pivoting Threshold
See MA86 documentation. Either ma86 static=0.0 or ma86 static¿ma86 small. ma86 static=0.0 disables
static pivoting. The valid range for this real option is 0 ≤ ma86 static < +inf and its default value is
0.
ma86 u: Pivoting Threshold
See MA86 documentation. The valid range for this real option is 0 ≤ ma86 u ≤ 0.5 and its default value
is 1 · 10−08 .
86

ma86 umax: Maximum Pivoting Threshold
Maximum value to which u will be increased to improve quality. The valid range for this real option is
0 ≤ ma86 umax ≤ 0.5 and its default value is 0.0001.

C.19

MA97 Linear Solver

ma97 print level: Debug printing level for the linear solver MA97
The valid range for this integer option is −inf < ma97 print level < +inf and its default value is 0.
ma97 nemin: Node Amalgamation parameter
Two nodes in elimination tree are merged if result has fewer than ma97 nemin variables. The valid range
for this integer option is 1 ≤ ma97 nemin < +inf and its default value is 8.
ma97 order: Controls type of ordering used by HSL MA97
The default value for this string option is ”auto”.
Possible values:
• auto: Use HSL MA97 heuristic to guess best of AMD and METIS
• best: Try both AMD and MeTiS, pick best
• amd: Use the HSL MC68 approximate minimum degree algorithm
• metis: Use the MeTiS nested dissection algorithm
• matched-auto: Use the HSL MC80 matching with heuristic choice of AMD or METIS
• matched-metis: Use the HSL MC80 matching based ordering with METIS
• matched-amd: Use the HSL MC80 matching based ordering with AMD
ma97 scaling: Specifies strategy for scaling in HSL MA97 linear solver
The default value for this string option is ”dynamic”.
Possible values:
• none: Do not scale the linear system matrix
• mc30: Scale all linear system matrices using MC30
• mc64: Scale all linear system matrices using MC64
• mc77: Scale all linear system matrices using MC77 [1,3,0]
• dynamic: Dynamically select scaling according to rules specified by ma97 scalingX and ma97 switchX
options.
ma97 scaling1: First scaling.
If ma97 scaling=dynamic, this scaling is used according to the trigger ma97 switch1. If ma97 switch2 is
triggered it is disabled. The default value for this string option is ”mc64”.
Possible values:
• none: No scaling
• mc30: Scale linear system matrix using MC30
• mc64: Scale linear system matrix using MC64
• mc77: Scale linear system matrix using MC77 [1,3,0]
87

ma97 scaling2: Second scaling.
If ma97 scaling=dynamic, this scaling is used according to the trigger ma97 switch2. If ma97 switch3 is
triggered it is disabled. The default value for this string option is ”mc64”.
Possible values:
• none: No scaling
• mc30: Scale linear system matrix using MC30
• mc64: Scale linear system matrix using MC64
• mc77: Scale linear system matrix using MC77 [1,3,0]
ma97 scaling3: Third scaling.
If ma97 scaling=dynamic, this scaling is used according to the trigger ma97 switch3. The default value
for this string option is ”mc64”.
Possible values:
• none: No scaling
• mc30: Scale linear system matrix using MC30
• mc64: Scale linear system matrix using MC64
• mc77: Scale linear system matrix using MC77 [1,3,0]
ma97 small: Zero Pivot Threshold
Any pivot less than ma97 small is treated as zero. The valid range for this real option is 0 ≤ ma97 small <
+inf and its default value is 1 · 10−20 .
ma97 solve blas3: Controls if blas2 or blas3 routines are used for solve
The default value for this string option is ”no”.
Possible values:
• no: Use BLAS2 (faster, some implementations bit incompatible)
• yes: Use BLAS3 (slower)
ma97 switch1: First switch, determine when ma97 scaling1 is enabled.
If ma97 scaling=dynamic, ma97 scaling1 is enabled according to this condition. If ma97 switch2 occurs
this option is henceforth ignored. The default value for this string option is ”od hd reuse”.
Possible values:
• never: Scaling is never enabled.
• at start: Scaling to be used from the very start.
• at start reuse: Scaling to be used on first iteration, then reused thereafter.
• on demand: Scaling to be used after Ipopt request improved solution (i.e. iterative refinement has
failed).
• on demand reuse: As on demand, but reuse scaling from previous itr
• high delay: Scaling to be used after more than 0.05*n delays are present
• high delay reuse: Scaling to be used only when previous itr created more that 0.05*n additional
delays, otherwise reuse scaling from previous itr
• od hd: Combination of on demand and high delay
• od hd reuse: Combination of on demand reuse and high delay reuse
88

ma97 switch2: Second switch, determine when ma97 scaling2 is enabled.
If ma97 scaling=dynamic, ma97 scaling2 is enabled according to this condition. If ma97 switch3 occurs
this option is henceforth ignored. The default value for this string option is ”never”.
Possible values:
• never: Scaling is never enabled.
• at start: Scaling to be used from the very start.
• at start reuse: Scaling to be used on first iteration, then reused thereafter.
• on demand: Scaling to be used after Ipopt request improved solution (i.e. iterative refinement has
failed).
• on demand reuse: As on demand, but reuse scaling from previous itr
• high delay: Scaling to be used after more than 0.05*n delays are present
• high delay reuse: Scaling to be used only when previous itr created more that 0.05*n additional
delays, otherwise reuse scaling from previous itr
• od hd: Combination of on demand and high delay
• od hd reuse: Combination of on demand reuse and high delay reuse
ma97 switch3: Third switch, determine when ma97 scaling3 is enabled.
If ma97 scaling=dynamic, ma97 scaling3 is enabled according to this condition. The default value for
this string option is ”never”.
Possible values:
• never: Scaling is never enabled.
• at start: Scaling to be used from the very start.
• at start reuse: Scaling to be used on first iteration, then reused thereafter.
• on demand: Scaling to be used after Ipopt request improved solution (i.e. iterative refinement has
failed).
• on demand reuse: As on demand, but reuse scaling from previous itr
• high delay: Scaling to be used after more than 0.05*n delays are present
• high delay reuse: Scaling to be used only when previous itr created more that 0.05*n additional
delays, otherwise reuse scaling from previous itr
• od hd: Combination of on demand and high delay
• od hd reuse: Combination of on demand reuse and high delay reuse
ma97 u: Pivoting Threshold
See MA97 documentation. The valid range for this real option is 0 ≤ ma97 u ≤ 0.5 and its default value
is 1 · 10−08 .
ma97 umax: Maximum Pivoting Threshold
See MA97 documentation. The valid range for this real option is 0 ≤ ma97 umax ≤ 0.5 and its default
value is 0.0001.

C.20

MUMPS Linear Solver

mumps pivtol: Pivot tolerance for the linear solver MUMPS.
A smaller number pivots for sparsity, a larger number pivots for stability. This option is only available
if Ipopt has been compiled with MUMPS. The valid range for this real option is 0 ≤ mumps pivtol ≤ 1
and its default value is 1 · 10−06 .
mumps pivtolmax: Maximum pivot tolerance for the linear solver MUMPS.
Ipopt may increase pivtol as high as pivtolmax to get a more accurate solution to the linear system. This
option is only available if Ipopt has been compiled with MUMPS. The valid range for this real option is
0 ≤ mumps pivtolmax ≤ 1 and its default value is 0.1.
mumps mem percent: Percentage increase in the estimated working space for MUMPS.
In MUMPS when significant extra fill-in is caused by numerical pivoting, larger values of mumps mem percent
may help use the workspace more efficiently. On the other hand, if memory requirement are too large at
the very beginning of the optimization, choosing a much smaller value for this option, such as 5, might
reduce memory requirements. The valid range for this integer option is 0 ≤ mumps mem percent < +inf
and its default value is 1000.
mumps permuting scaling: Controls permuting and scaling in MUMPS
This is ICNTL(6) in MUMPS. The valid range for this integer option is 0 ≤ mumps permuting scaling ≤
7 and its default value is 7.
mumps pivot order: Controls pivot order in MUMPS
This is ICNTL(7) in MUMPS. The valid range for this integer option is 0 ≤ mumps pivot order ≤ 7 and
its default value is 7.
mumps scaling: Controls scaling in MUMPS
This is ICNTL(8) in MUMPS. The valid range for this integer option is −2 ≤ mumps scaling ≤ 77 and
its default value is 77.

C.21

Pardiso Linear Solver

pardiso matching strategy: Matching strategy to be used by Pardiso
This is IPAR(13) in Pardiso manual. The default value for this string option is ”complete+2x2”.
Possible values:
• complete: Match complete (IPAR(13)=1)
• complete+2x2: Match complete+2x2 (IPAR(13)=2)
• constraints: Match constraints (IPAR(13)=3)
pardiso max iterative refinement steps: Limit on number of iterative refinement steps.
The solver does not perform more than the absolute value of this value steps of iterative refinement and
stops the process if a satisfactory level of accuracy of the solution in terms of backward error is achieved. If
negative, the accumulation of the residue uses extended precision real and complex data types. Perturbed
pivots result in iterative refinement. The solver automatically performs two steps of iterative refinements
when perturbed pivots are obtained during the numerical factorization and this option is set to 0. The
valid range for this integer option is −inf < pardiso max iterative refinement steps < +inf and
its default value is 0.

pardiso msglvl: Pardiso message level
This determines the amount of analysis output from the Pardiso solver. This is MSGLVL in the Pardiso
manual. The valid range for this integer option is 0 ≤ pardiso msglvl < +inf and its default value is
0.
pardiso order: Controls the fill-in reduction ordering algorithm for the input matrix.
The default value for this string option is ”five”.
Possible values:
• amd: minimum degree algorithm
• one: undocumented
• metis: MeTiS nested dissection algorithm
• pmetis: parallel (OpenMP) version of MeTiS nested dissection algorithm
• four: undocumented
• five: undocumented

C.22

WSMP Linear Solver

wsmp num threads: Number of threads to be used in WSMP
This determines on how many processors WSMP is running on. This option is only available if Ipopt has
been compiled with WSMP. The valid range for this integer option is −inf < wsmp num threads < +inf
and its default value is 1.
wsmp ordering option: Determines how ordering is done in WSMP (IPARM(16)
This corresponds to the value of WSSMP’s IPARM(16). This option is only available if Ipopt has been
compiled with WSMP. The valid range for this integer option is −2 ≤ wsmp ordering option ≤ 3 and
its default value is 1.
wsmp pivtol: Pivot tolerance for the linear solver WSMP.
A smaller number pivots for sparsity, a larger number pivots for stability. This option is only available if
Ipopt has been compiled with WSMP. The valid range for this real option is 0 < wsmp pivtol < 1 and
its default value is 0.0001.
wsmp pivtolmax: Maximum pivot tolerance for the linear solver WSMP.
Ipopt may increase pivtol as high as pivtolmax to get a more accurate solution to the linear system. This
option is only available if Ipopt has been compiled with WSMP. The valid range for this real option is
0 < wsmp pivtolmax < 1 and its default value is 0.1.
wsmp scaling: Determines how the matrix is scaled by WSMP.
This corresponds to the value of WSSMP’s IPARM(10). This option is only available if Ipopt has been
compiled with WSMP. The valid range for this integer option is 0 ≤ wsmp scaling ≤ 3 and its default
value is 0.
wsmp singularity threshold: WSMP’s singularity threshold.
WSMP’s DPARM(10) parameter. The smaller this value the less likely a matrix is declared singular.
This option is only available if Ipopt has been compiled with WSMP. The valid range for this real option
is 0 < wsmp singularity threshold < 1 and its default value is 1 · 10−18 .

Options available via the AMPL Interface

The following is a list of options that is available via AMPL:
acceptable compl inf tol Acceptance threshold for the complementarity conditions
acceptable constr viol tol Acceptance threshold for the constraint violation
acceptable dual inf tol Acceptance threshold for the dual infeasibility
acceptable tol Acceptable convergence tolerance (relative)
alpha for y Step size for constraint multipliers
bound frac Desired minimal relative distance of initial point to bound
bound mult init val Initial value for the bound multipliers
bound push Desired minimal absolute distance of initial point to bound
bound relax factor Factor for initial relaxation of the bounds
compl inf tol Acceptance threshold for the complementarity conditions
constr mult init max Maximal allowed least-square guess of constraint multipliers
constr viol tol Desired threshold for the constraint violation
diverging iterates tol Threshold for maximal value of primal iterates
dual inf tol Desired threshold for the dual infeasibility
expect infeasible problem Enable heuristics to quickly detect an infeasible problem
file print level Verbosity level for output file
halt on ampl error Exit with message on evaluation error
hessian approximation Can enable Quasi-Newton approximation of hessian
honor original bounds If no, solution might slightly violate bounds
linear scaling on demand Enables heuristic for scaling only when seems required
linear solver Linear solver to be used for step calculation
linear system scaling Method for scaling the linear systems
ma27 pivtol Pivot tolerance for the linear solver MA27
ma27 pivtolmax Maximal pivot tolerance for the linear solver MA27
ma57 pivot order Controls pivot order in MA57
ma57 pivtol Pivot tolerance for the linear solver MA57
ma57 pivtolmax Maximal pivot tolerance for the linear solver MA57
max cpu time CPU time limit
max iter Maximum number of iterations
max refinement steps Maximal number of iterative refinement steps per linear system solve
max soc Maximal number of second order correction trial steps
92

maxit (Ipopt name: max iter) Maximum number of iterations
min refinement steps Minimum number of iterative refinement steps per linear system solve
mu init Initial value for the barrier parameter
mu max Maximal value for barrier parameter for adaptive strategy
mu oracle Oracle for a new barrier parameter in the adaptive strategy
mu strategy Update strategy for barrier parameter
nlp scaling max gradient Maximum gradient after scaling
nlp scaling method Select the technique used for scaling the NLP
obj scaling factor Scaling factor for the objective function
option file name File name of options file (default: ipopt.opt)
outlev (Ipopt name: print level) Verbosity level
output file File name of an output file (leave unset for no file output)
pardiso matching strategy Matching strategy for linear solver Pardiso
print level Verbosity level
print options documentation Print all available options (for ipopt.opt)
print user options Toggle printing of user options
required infeasibility reduction Required infeasibility reduction in restoration phase
slack bound frac Desired minimal relative distance of initial slack to bound
slack bound push Desired minimal absolute distance of initial slack to bound
tol Desired convergence tolerance (relative)
wantsol solution report without -AMPL: sum of, 1 == write .sol file, 2 == print primal variable values,
4 == print dual variable values, 8 == do not print solution message
warm start bound push Enables to specify how much should variables should be pushed inside the
feasible region
warm start init point Enables to specify bound multiplier values
warm start mult bound push Enables to specify how much should bound multipliers should be
pushed inside the feasible region
watchdog shortened iter trigger Trigger counter for watchdog procedure

References
[1] L. T. Biegler. Nonlinear Programming: Concepts, Algorithms and Applications to Chemical Processes
SIAM, Philadelphia (2010)
[2] F. E. Curtis, J. Huber, O. Schenk, A. Wächter. A note on the implementation of an interiorpoint algorithm for nonlinear optimization with inexact step computations. Mathematical Programming, 136(1):209–227, 2012 doi: 10.1007/s10107-012-0557-4. preprint at http://www.
optimization-online.org/DB_HTML/2011/04/2992.html
[3] R. Fourer, D. M. Gay, and B. W. Kernighan. AMPL: A Modeling Language For Mathematical
Programming. Thomson Publishing Company, Danvers, MA, USA, 1993.
[4] O. Schenk, A. Wächter, and M. Hagemann. Matching-based preprocessing algorithms to the solution of saddle-point problems in large-scale nonconvex interior-point optimization. Computational
Optimization Applications, 36(2-3):321–341, 2007 doi: 10.1007/s10589-006-9003-y.
[5] W. Hock and K. Schittkowski. Test examples for nonlinear programming codes. Lecture Notes in
Economics and Mathematical Systems, 187, 1981. doi: 10.1007/978-3-642-48320-2.
[6] J. Nocedal, A. Wächter, and R. A. Waltz. Adaptive barrier strategies for nonlinear interior methods.
SIAM Journal on Optimization, 19(4):1674–1693, 2008. doi: 10.1137/060649513. preprint at http:
//www.optimization-online.org/DB_HTML/2005/03/1089.html
[7] H. Pirnay, R. López-Negrete, and L. T. Biegler. Optimal Sensitivity based on Ipopt. Mathematical
Programming Computations, 4(4):307–331, 2012. doi: 10.1007/s12532-012-0043-2. preprint at http:
//www.optimization-online.org/DB_HTML/2011/04/3008.html
[8] A. Wächter. An Interior Point Algorithm for Large-Scale Nonlinear Optimization with Applications
in Process Engineering. PhD thesis, Carnegie Mellon University, Pittsburgh, PA, USA, January 2002.
available at http://researcher.watson.ibm.com/researcher/files/us-andreasw/thesis.pdf
[9] A. Wächter. Short Tutorial: Getting Started With Ipopt in 90 Minutes. In Combinatorial Scientific
Computing (U. Naumann, O. Schenk, H. D. Simon, eds.), 2009. urn: nbn:de:0030-drops-20890
[10] A. Wächter and L. T. Biegler. Line search filter methods for nonlinear programming: Local convergence. SIAM Journal on Optimization, 16(1):32–48, 2005. doi: 10.1137/S1052623403426544
[11] A. Wächter and L. T. Biegler. Line search filter methods for nonlinear programming: Motivation and global convergence.
SIAM Journal on Optimization, 16(1):1–31, 2005.
doi:
10.1137/S1052623403426556
[12] A. Wächter and L. T. Biegler. Global and Local Convergence of Line Search Filter Methods for
Nonlinear Programming. Optimization Online, 2001. http://www.optimization-online.org/DB_
HTML/2001/08/367.html
[13] A. Wächter and L. T. Biegler. On the implementation of a primal-dual interior point filter line search
algorithm for large-scale nonlinear programming. Mathematical Programming, 106(1):25–57, 2006. doi:
10.1007/s10107-004-0559-y. preprint at http://www.optimization-online.org/DB_HTML/2004/03/
836.html

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.5
Linearized                      : No
Page Count                      : 94
Page Mode                       : UseOutlines
Language                        : en
Has XFA                         : No
Author                          : 
Title                           : Ipopt documentation
Subject                         : 
Creator                         : LaTeX with hyperref package
Producer                        : pdfTeX-1.40.15
Create Date                     : 2015:04:16 12:05:14+02:00
Modify Date                     : 2016:10:27 23:36:04-03:00
Trapped                         : False
PTEX Fullbanner                 : This is pdfTeX, Version 3.14159265-2.6-1.40.15 (TeX Live 2014/Arch Linux) kpathsea version 6.2.0

EXIF Metadata provided by EXIF.tools

Ipopt Ation Manual

manual_IPOPT

Navigation menu

Versions of this User Manual:

Views

Navigation