USING SEDUMI 1.02, A MATLAB TOOLBOX FOR

y
OPTIMIZATION OVER SYMMETRIC CONES

(Updated for Version 1.05)

JOS F. STURM

Department of E onometri s, Tilburg University,

Tilburg, The Netherlands.
http://few al.kub.nl/sturm, E-mail: j.f.sturmkub.nl.
August 1998 { O tober 2001

Abstra t

SeDuMi is an add-on for MATLAB, whi h lets you solve optimization problems with
linear, quadrati and semideniteness onstraints. It is possible to have omplex valued data
and variables in SeDuMi. Moreover, large s ale optimization problems are solved e iently,
by exploiting sparsity. This paper des ribes how to work with this toolbox.
Keywords: Symmetri one, semidenite programming, se ond order one programming,
self{duality, MATLAB, SeDuMi

SeDuMi stands for Self-Dual-Minimization: it implements the self-dual embedding te hnique

for optimization over self-dual homogeneous ones. The self-dual embedding te hnique as proposed by Ye, Todd and Mizuno [31, essentially makes it possible to solve ertain optimization
problems in a single phase, leading either to an optimal solution, or a erti ate of infeasibility.
Optimization over self{dual homogeneous ones, or more on isely, optimization over symmetri
ones, was rst studied by Nesterov and Todd [21, and is urrently an a tive area of resear h.
Semidenite programming is a spe ial ase of optimization over symmetri ones. The popular pa kage SP by Vandenberghe and Boyd [29 is one of the rst software tools that was
developed for semidenite programming. Some ontrol theorists use SP indire tly via LMITOOL, by El Ghaoui, Nikoukhah and Delebe que [8, or MRCT, by Dussy and El Ghaoui [6,
 MATLAB is a registered trademark of The MathWorks, In .
y
Appeared in: Optimization Methods and Software 11{12 (1999) 625{653
z Resear h up to version 1.02 performed at Communi ations Resear h Laboratory,

M Master University, Hamilton, Canada. Supported by Netherlands Organization for S ienti Resear h (NWO).

whi h are user-friendly front-ends for SP. More re ently, the user-friendly front-end SeDuMi
Interfa e[22 is available from
http://www.laas.fr/~peau ell/SeDuMiInt.html

A more re ent and faster solver for semidenite programming is SDPA, by Fujisawa, Kojima and
Nakata [11. Other solvers for semidenite programming are CSDP by Bor hers [2, SDPHA by
Brixius and Potra [3 and SDPT3 by Toh, Todd and Tutun u [28. See Mittelmann [19 for a
omparison of the performan e of various solvers (in luding SeDuMi) on semidenite programming problems of the SDPLIB test set.
For optimization over symmetri ones, there were until re ently only two software tools
available, viz. SDPPa k, by Alizadeh et al. [1, and SeDuMi. (SDPT3 has been extended to this
lass of problems in 2001.) Both operate under the MATLAB environment, so that they an
easily be used within spe i appli ations. SeDuMi has some features that are not available in
SDPPa k, namely it
 allows the use of omplex valued data,
 generates Farkas-dual solutions for infeasible problems,
 takes full advantage of sparsity, leading to signi ant speed benets,
 has a theoreti ally proven O(pn log(1=)) worst- ase iteration bound,
 promotes sparsity by handling dense olumns separately (sin e version 1.03), using a te hnique proposed by Goldfarb and S heinberg [13,
 an import linear programs in MPS format (either using a link to LIPSOL [32 or by using
the loadmps add-on by the author), and semidenite programs in SDPA [11 format.
It is also possible to onvert optimization problems from SDPPa k [1 format into SeDuMi. Noti e
that earlier versions of the semidenite programming solver SDPT3 were also able to handle
omplex valued data [28, but this option has been removed re ently. The issue of expoiting
sparsity in semidenite programming was studied by Fujisawa, Kojima and Nakata [10. Unlike
the approa h of [10, SeDuMi uses always the same sparsity expoiting pro edure to form the
normal equations; this pro edure is e ient regardless of the degree of sparsity. See Ross [24
for a omparison between SDPpa k and SeDuMi.
The remainder of this do ument is a step-by-step tutorial for SeDuMi. The on-line help pages
serve as a referen e to the toolbox. In addition, the Appendix to this do ument has the details
of the alling sequen e for the main fun tion, sedumi.

1 introdu tion to sedumi

Throughout this do ument, we assume that SeDuMi is orre tly installed, and that you are
working under MATLAB Version 5 or later. Entering the MATLAB ommand `help' should
produ e a list of all installed MATLAB Toolboxes, in luding the following lines:
>> help
HELP topi s:
matlab/SeDuMi105
SeDuMi105/ onversion

- SeDuMi 1.05
(OCT2001)
- Conversion to SeDuMi.

For more help on dire tory/topi , type "help topi ".

The ommand help onversion produ es a list of fun tions for importing data into SeDuMi.
This in ludes an `umbrella' s ript, getproblem1, whi h works as follows:
>> pname = 'truss2'; getproblem, who
Your variables are:
At
K

MATNAME
PROBDIR

pname

This imports problem `truss2', and pla es it in the variables At, b, and K. To do this, SeDuMi
must be able to nd the requested problem somewhere on your disk. It an lo ate sparse SDPA
problems, if you have assigned a UNIX or DOS environment variable `SDPLIB' to the dire tory
path where SDPA problems are stored. (SDPLIB and SeDuMi have a dierent anoni al form; if
y is a dual optimal solution al ulated by SeDuMi, then y is an optimal solution in the SDPLIB
anoni al form.) If LIPSOL is installed, it uses LIPSOL's fun tion findprob to lo ate linear
programming problems in MPS format. Finally, if SDPPa k is installed, and the environment
variable SDPPACK points to the SDPPa k dire tory, then SDPPa k problems are sear hed for in
the dire tory `SDPPACK/problems'.
Typing `help SeDuMi105' produ es a list of the fun tions that you an use to build and
solve optimization models over symmetri ones. They are: sedumi, eigK, ve , mat and eyeK.
Online help is provided by help sedumi, help eigK, and so on. The following se tions give a
more detailed explanation of these fun tions, with some illustrating examples.
1

This fun tion has not been updated sin e Version 1.00, and may not be ompatible with the latest software

2 linear programming
It is possible to formulate your linear programming model in either the primal standard form,
minimize
T x
su h that
Ax = b
(1)
x  0 for i = 1; 2; : : : ; n;
or the dual standard form,
maximize
bT y
(2)
su h that aTy  0 for i = 1; 2; : : : ; n:
Mixed form and symmetri form linear programming models may be formulated using the `free
variable' denition in SeDuMi, as outlined in Se tion 2.1.
Suppose that you want to solve the following linear programming problem:
minimize
x1 x2
su h that 10x1 7x2  5
(3)
x1 + x2 =2  3
x1  0; x2  0:
In order to formulate this LP problem in the primal standard form, we have to add sla k
variables, say x3 and x4 . In MATLAB, we an then enter the b and ve tors, and the A matrix
as follows:
i

>> = [1; -1; 0; 0;

>> A = [10, -7, -1, 0; 1, 1/2, 0, 1;
>> b = [5; 3;

We an now solve problem (3) in the primal form (1) by invoking the fun tion sedumi.
Remark that MATLAB is ase sensitive, and it is therefore essential to write sedumi in lower
ase.
>> sedumi(A,b, )
SeDuMi by Jos F. Sturm, 1998.
Alg = 2: xz- orre tor, theta = 0.250, beta = 0.500
eqs m = 2, order n = 5, dim = 5, blo ks = 1
it :
x
gap delta rate t/maxt feas
0 :
5.00E+00 0.000
1 : 7.81E-01 9.79E-01 0.000 0.1959 0.9000 0.77
2 : -5.52E-02 9.40E-02 0.000 0.0959 0.9900 0.93

* 3 : -1.25E-01 5.70E-04 0.000 0.0061 0.9990 1.08

iter se onds digits
*x
b*y
3
0.1 15.1 -1.2500000000e-01 -1.2500000000e-01
|Ax-b| = 1.8e-15, |x|= 2.9e+00, |y|= 2.8e-01
ans =
1.9583
2.0833
0
0

This shows that the optimal value is 0:125, as listed under *x. The fun tion sedumi
returns an optimal solution, whi h in this ase is x1 = 1:9583 and x2 = 2:0833. Noti e that
x is indeed feasible for (1), be ause all its omponents are nonnegative, and Ax = b, as an
be he ked by the ommands min(x) and norm(Ax-b), respe tively. Of ourse, some round-o
errors may o ur, as an be seen from the following MATLAB output:
>> A=sparse(A);norm(A*x-b)
ans =
1.7764e-15
>> norm(A*(24*x)-24*b)
ans =
0

The rst quantity is listed as jAx bj = 1:8e 15 in the output of SeDuMi. The se ond line
shows that the reported value of |Ax-b| does not only ontain the residual, but also errors in
omputing the residual. The meaning of the other parts in the output of SeDuMi is explained in
Se tion 5.
Using dual solutions, it is possible to he k also optimality. Namely, if we let z := AT y,
then if x and y are feasible to (1) and (2) respe tively, we have
0  zTx = T x yTAx = Tx bTy:
Thus, if T x bTy = 0, then T x must be minimal, and bTy must be maximal, over all feasible
solutions. The dual solution y to (2) is assigned to the se ond output argument of sedumi, as in
5

>> [x,y = sedumi(A,b, )

In this example, we have y1 = 0:125 and y2 = 0:25. Issueing the ommand

>> z = - A'*y
z =
0
0
0.1250
0.2500

we see that z x = 0 for all i, proving optimality. However, due to some round-o errors,
bT y is positive in this ase. The quantity digits = 15:1 in the output of SeDuMi, is
dened as follows:
log10 (( T x bTy)=(jbT yj + 10 10 )) if Tx bT y > 0
digits =
(4)
1 otherwise:
As is well known, y is a subgradient of the optimal value fun tion in terms of hanges in b. If
the optimal value fun tion is lo ally not dierentiable in b, i.e. if there are multiple dual optimal
solutions, then it is said to be primal degenerate. SeDuMi usually generates a solution y in
the relative interior of the subgradient set, be ause it uses a Mehrotra-Ye [18 type termination
pro edure for linear programs. For a detailed treatment of sensitivity analysis based on su h
solutions, we refer to Monteiro and Mehrotra [20 and the book of Roos, Terlaky and Vial [23.
For large problems, it is usually not feasible to store A as a full matrix, due to memory
limitations. In this ase, A should be stored in sparse format; type help sparfun for details.
Internally, SeDuMi always onverts A to sparse format. The b and ve tors an also be in sparse
format, if desired.
In the pre eding, we dened b and in MATLAB as olumn ve tors, but this is not essential;
SeDuMi produ es the same output if b and/or are dened as row ve tors. Similarly, SeDuMi is
not pi ky about the orientation of A: it will dete t the orre t orientation based on the b and
ve tors (ex ept in the unrealisti ase that A is square). In fa t, it is good pra ti e to store A
in su h a way that it has more rows than olumns, whi h is the transpose orientation of the A
matri es that we have seen so far. Namely, if A is stored in sparse format, then it is stored as a
set of sparse olumn ve tors. Hen e, if there are fewer olumns, it will o upy less spa e.
There is a third output argument of SeDuMi, alled info. In our example,
T x

8
<
:

>> [x,y,info=sedumi(A,b, ); info

info =

 puse :
iter:
feasratio:
numerr:
pinf:
dinf:

0.1100
3
1
0
0
0

This is a ompound output argument, or stru ture, with a eld puse for the solution time,
iter for the number of iterations, a eld numerr whi h is nonzero in ase of numeri al problems
(1 means premature termination: results are ina urate, 2 means failure), a eld feasratio
for the nal value of the feasibility indi ator, and two elds, pinf and dinf, for the dete ted
feasibility status of the optimization problem. If pinf = 1, then the primal problem (1) is
infeasible, and y is a Farkas dual solution.
For instan e, if we hange the b ve tor in the pre eding example to b = 5; 0:4 , then
SeDuMi yields info:pinf = 1, bT y = 0:0955 > 0, max =1 2 3 4 aT y = 0:1866  0. Noti e that
for any x with Ax = b, we have yTAx = bTy = 0:0955 > 0, whereas yTAx  0 for nonnegative
x, be ause all omponents of AT y are nonpositive. A Farkas dual solution thus provides a
erti ate of infeasibility. In this example, there appears to be a Farkas dual solution for whi h
all entries of ATy are stri tly negative. In general though, they are merely nonpositive. For
numeri al reasons, ATy an then ontain some small positive omponents, and in this ase we
have an approximate Farkas dual solution. Loosely speaking, su h solutions demonstrate that
there annot be any reasonably sized primal feasible solution; see Todd and Ye [27 for details.
Suppose now that we want to solve a problem in the dual standard form (2). In this ase,
y with bT y > 0 and AT y  0 has the interpretation of an improving dire tion. Namely, if
there exists a feasible solution y, i.e. if ATy  0, then y + ty is feasible for all t  0, and
lim !1 bTy = 1. In this ase, we say that the problem is unbounded. The other possibility is
that there does not exist any feasible solution y, i.e. the problem is infeasible. To distinguish
between an infeasible and an unbounded problem, we have to go through a se ond stage, by
solving a feasibility problem:
h

; ; ;

>> [x,y,info = sedumi(A,zeros(length(b),1), )

whi h may be entered in simplied form (sin e version 1.05) as

>> [x,y,info = sedumi(A,0, )

In our previous example, the dual problem is feasible, and the above ommand yields a feasible
solution y. The need for this se ond stage feasibility problem is typi al for interior point methods
with the self{dual embedding te hnique of Ye, Todd and Mizuno [31.
7

The interpretation of info:dinf is analogous to that of info:pinf. Namely, if info:dinf = 1

then the dual problem (2) is infeasible, and this laim is ertied by a Farkas solution x with
T x < 0; Ax = 0; x  0 for i = 1; 2; : : : ; n:
To distinguish between primal unboundedness and primal infeasibility, we would then solve the
feasibility problem
i

>> [x,y,info = sedumi(A,b,zeros(length( ),1))

an also generate Gordan-Stiemke dual solutions. For instan e, if we restore the

ve tor b to bT = 5; 3 , and solve the feasibility problem x = sedumi(A,b,zeros(4,1)), we
obtain a stri tly positive ve tor x. This is be ause interior point methods try to nd a solution
in the relative interior of the solution set. To see what happens if feasible solutions an merely
be nonnegative, onsider the following example:
SeDuMi

>> b = [5, 1/2;

>> [x,y,info=sedumi(A,b,zeros(4,1));
>> [x -A'*y
ans =
0.5000
0
0
0

0
1.1875
0.0990
0.9895

>> b*y
ans =
1.3878e-17

In this example, the primal does not have an interior solution, i.e. it is weakly feasible, and this
is demonstrated by a Gordan-Stiemke dual solution y. Namely, 0 6= ATy  0, and bTy = 0,
whi h learly implies that there annot be any x > 0 su h that A  x = b.
SeDuMi treats the primal and dual in a symmetri way, i.e. it does not favor one over the
other. From a modeling point of view however, the primal standard form and the dual standard
form are quite dierent, and it depends on the appli ation whi h one is more favorable. The
primal form has the advantage of expli it equality onstraints. In prin iple, equality onstraints
an be onstru ted in the dual form also, simply by means of two inequality onstraints, su h as
8

and aTy  . However, this te hnique is not re ommended, sin e su h onstraint pairs
tend to get a pair of very large primal multipliers x , hen e leading to numeri al di ulties. It
may be better to enfor e an equality onstraint by eliminating a y variable. However, the latter
te hnique may destroy the sparsity stru ture of the A-matrix, thus leading to longer solution
times.
Exa tly the same problems arise in modeling a free (i.e. unresistri ted in sign) variable in
the primal standard form. Splitting su h a variable into two, its positive part and its negative
part, often results in numeri al di ulties. One may also try to eliminate su h a variable by
removing an equality onstraint, but this usually auses an in rease in the number of nonzeros in
the A-matrix. An alternative is to model all free variables in a quadrati one. Quadrati ones
are dis ussed in Se tion 3. To prevent numeri al di ulties with this te hnique, it is desirable
to x a { possibly large { upper bound on the norm of the ve tor of free variables, whi h is
easily done in a quadrati one.
Sin e Version 1.05, the user does not need to worry about these issues, sin e free variables
are allowed. This is the topi of the se tion below.
aT y 
i

2.1 Free variables

It is possible to formulate your linear programming model in a primal form with free variables
as follows:
minimize
T x
su h that
Ax = b
(5)
x 2 < for i = 1; 2; : : : ; K:f;
x  0 for j = K:f + 1; K:f + 2; : : : ; n;
where K.f is the number of free variables. The asso iated sla k variables in the dual problem
are then restri ted to be zero, thus allowing equality onstraints in the dual:
maximize
bT y
T
su h that
a y = 0 for i = 1; 2; : : : ; K:f
(6)
T
a y  0 for j = K:f + 1; K:f + 2; : : : ; n
i

3 quadrati and semidenite onstraints

In SeDuMi, it is possible to impose quadrati or semidenite onstraints, by restri ting variables
to a quadrati one or the one of positive semidenite matri es, respe tively. Su h a restri tion
then repla es the nonnegativity restri tion in linear programming. Thus, instead of requiring
x 2 <+ as in (1), we will now require x 2 K, where K is a so- alled symmetri one. A
symmetri one is a Cartesian produ t of a nonnegative orthant, quadrati ones and ones of
n

positive semidenite matri es. The standard primal form for su h optimization problems is
minimize T x
su h that Ax = b
(7)
x2K

and the dual standard form is

maximize
su h that

bT y
AT y 2 K :

(8)

3.1 The quadrati one

A quadrati one is by denition a one of the form

Q one := f(x1 ; x2 ) 2 <  < 1 j x1  kx2 kg;
(9)
where k
k denotes the Eu lidean norm (the fun tion norm in MATLAB). The quadrati one
is also known as the se ond order one or Lorentz one. As an example, onsider the following
optimization problem:
N

min

y1 + y2 y1  kq

P y3 k; y2 

1 + ky3 k2 ;

(10)

where P is a given matrix, and q a given ve tor. The above is a robust least squares problem,
see El Ghaoui and Lebret [7. The de ision variables are the s alars y1 and y2, and the ve tor
y3 . This problem has two quadrati onstraints, viz.
(y1 ; q P y3) 2 Q one; y2; 1 2 Q one:
(11)
"

#!

y3

Given P and q, the following MATLAB fun tion (rls.m) onstru ts problem (10) in the
standard dual form (8). The A matrix will be in transposed orientation, and is hen e denoted
as At.
1
2
3
4
5
6
7
8

% [At,b, ,K = rls(P,q)
% Creates dual standard form for robust least squares problem "Pu=q".
fun tion [At,b, ,K = rls(P,q)
[m, n = size(P);
% ---------- minimize y_1 + y_2 ---------b = -sparse([1; 1; zeros(n,1));
% ---------- (y_1, q - P y_3) in Q one ----------

10

9
10
11
12
13
14
15
16
17
18

At = sparse([-1, zeros(1,1+n); ...

zeros(m,2), P);
= [0;q;
K.q = [1+m;
% ---------- (y_2, (1,y_3)) in Q one ---------At = [At; 0, -1, zeros(1,n); ...
zeros(1,2+n); ...
zeros(n,2), -eye(n);
= sparse([ ; 0;1;zeros(n,1));
K.q = [K.q, 2+n;

Noti e rst that the above fun tion uses sparse data types, in order to save memory. Furthermore, a stru ture K is dened, with a eld K.q that lists the dimensions of the quadrati
ones. (The `q' in K.q stands for `quadrati '.) The K{stru ture will be used to tell SeDuMi
that the omponents of ATy are not restri ted to be nonnegative as they would be in linear
programming. Instead, the rst K:q(1) entries are restri ted to a quadrati one, and the last
K.q(2) entries are restri ted to another quadrati one. This is the way in whi h we model the
symmetri one K in (7) and (8), and hen e onstru t the two quadrati onstraints in (11).
As a numeri al example, we solve a 4  3 robust least squares problem with dependent
olumns in P . The example is from [7.
>> P = [3 1 4;0 1 1;-2 5 3; 1 4 5; q = [0;2;1;3;
>> [At,b, ,K = rls(P,q);
>> [x,y,info = sedumi(At,b, ,K);
SeDuMi by Jos F. Sturm, 1998.
Alg = 1: v- orre tor, theta = 0.250, beta = 0.500
eqs m = 5, order n = 5, dim = 11, blo ks = 3
it :
x
gap delta rate t/maxt feas
0 :
5.00E+00 0.000
1 : -1.23E+01 1.30E+00 0.000 0.2605 0.9000 -0.18
2 : -5.94E+00 3.34E-01 0.000 0.2568 0.9000 0.54
3 : -3.60E+00 6.14E-02 0.116 0.1838 0.9000 0.86
4 : -3.34E+00 1.80E-03 0.000 0.0293 0.9900 1.10
* 5 : -3.33E+00 4.00E-06 0.000 0.0022 0.9990 1.00
* 6 : -3.33E+00 9.59E-09 0.000 0.0024 0.9990 1.00
* 7 : -3.33E+00 6.06E-10 0.153 0.0632 0.9900 1.00
* 8 : -3.33E+00 1.24E-10 0.000 0.2037 0.9000 1.00
iter se onds digits
*x
b*y
8
0.1 11.3 -3.3329085968e+00 -3.3329085968e+00
|Ax-b| = 2.8e-16, |x|= 2.0e+00, |y|= 2.5e+00

11

In the above all to SeDuMi, we see a new input argument, viz. K. This argument makes
solve an optimization problem in the form (7){(8), where the symmetri one K is
des ribed by the stru ture K. Without the fourth input argument (K), SeDuMi would solve a
linear programming problem of the form (1){(2).
To he k that (11) is indeed satised by the solution y, it is in prin iple possible to verify
the inequality in denition (9) dire tly. However, it is more onvenient to use the fun tion
eigK, whi h is part of SeDuMi. This fun tion returns the eigenvalues (or spe tral values) of a
ve tor with respe t to a symmetri one. A symmetri one onsists of those ve tors whi h
have nonnegative eigenvalues, see e.g. the book by Faraut and Koranyi [9. For a quadrati
one (9), there are merelyp two eigenvalues, viz. given a ve tor
p (x1 ; x2 ) 2 <  < 1, we have
1 (x1 ; x2 ) = (x1 kx2 k)= 2 and 2 (x1 ; x2 ) = (x1 + kx2 k)= 2.
We an thus he k feasibility and optimality as follows:

SeDuMi

>> [eigK(x,K), eigK( -At*y,K)

ans =
0.0000
1.4142
0.0000
1.4142

-0.0000
3.2307
-0.0000
1.4827

>> x'*( -At*y)

ans =
1.5807e-11

For symmetri ones K, it holds that xTz  0 for all x 2 K and z 2 K. Therefore, x provides
an optimality erti ate for y just as in the ase of linear programming. The interpretation
of Farkas dual solutions extends in the same way. See the survey paper of Luo, Sturm and
Zhang [16 for the details. However, a paradoxal phenomenon an o ur, viz. that x and y are
almost feasible, whereas Tx bT y is onsiderably negative (kxk and/or kyk must then obviously
be very large). SeDuMi will then report an innite number of digits in a ura y, a ording to
formula (4). This phenomenon was explained by Luo, Sturm and Zhang [15 and Sturm [25.
It is possible that an optimization model has both nonnegativity and quadrati one onstraints. For instan e, we may extend the above example with the restri tion that y3[1  0:1,
where y3 [1 denotes the rst omponent in the ve tor y3. This restri tion an be added to the
model as follows:
>> a1 = zeros(1,length(y)); a1(3) = 1;

12

>>
>>
>>
>>

= [-0.1; ; At = [a1;At;
K.l = 1;
[x,y,info = sedumi(At,b, ,K);
eigK( -At*y,K)'

ans =
0.0000

-0.0000

3.2307

-0.0000

1.4904

The eld K:l is the number of nonnegative variables, whi h in this ase is one. (The `l' in K.l
stands for `linear'.) By onvention, the nonnegative variables are always the rst omponents,
so that K = <+  Q one  Q one in our ase. As an be seen from the output of eigK, there
are 5 eigenvalues for this one: 1 for ea h nonnegativity onstraint, and 2 for ea h quadrati
onstraint. We say that K is a symmetri one of order 5. (SeDuMi reports `order n = 6', be ause
of its internal self{dual reformulation.)
SeDuMi supports an alternative form of the quadrati one, viz.
R one := (x1 ; x2 ; x3) 2 <  <  < 2 x1 x2  12 kx3k2 ; x1 + x2  0 :
(12)
Geometri ally, R one is simply a rotation of Q one. The spe i form of R one is onvenient for
modeling onvex quadrati fun tions. Namely, by adding the linear equality onstraint `x1 = 1'
to the model, we obtain the restri tion
1
x2  kx3 k2 :
2
Throughout the model, we an then use x2 as a tight upper bound on kx3 k2 =2. Fra tions are
also onveniently modeled by R one onstraints. For instan e, we may minimize 1=x1 for x1 > 0
by solving the model
minfx2 j x1x2  1; x1 + x2  0g:
Noti e that this problem does not have a solution: the inmum of 1=x1 is zero, for x1 ! 1.


>>
>>
>>
>>

lear K;
= [0, 1, 0; b = sqrt(2); A = [0, 0, 1; K.r = 3;
[x,y,info = sedumi(A,b, ,K);
x(2), x(1)*x(2)

ans =
1.5360e-05

13

ans =
1.0147

You may nd that x2 is not yet lose enough to zero, and that x1 is not equal to 1 either.
However, the primal solution is feasible, the dual solution is almost feasible, and the duality
gap is even negative. This illustrates an error bound di ulty, whi h is usual for this type of
irregular problems. In Se tion 5, we will see how to obtain a more a urate solution, by setting
an optional parameter, pars.eps.
As illustrated by the above example, the eld K.r serves to list the dimensions of R one
onstraints, analogously to the denition of Q one onstraints by K.q. (The `r' in K.r stands
for `rotated quadrati one'.) Setting both K.l, K.q and K.r elds yields a symmetri one of
the form
K = <K+l  (Q one 


 Q one)  (R one 


 R one):
For instan e, we an add a bound `x1  107 ' to the model as follows:
:

>> = [0, 0, 1, 0; b = [sqrt(2); 1E7; A = [0, 0, 0, 1;1, 1, 0, 0;

>> K.l = 1; K.r = 3;
>> [x,y,info = sedumi(A,b, ,K);

Some appli ations of Q one and R one onstraints are dis ussed in Lobo et al. [14.
3.2 The positive semidenite one

Semideniteness onstraints are an important lass of restri tions that an be modeled with
SeDuMi. As an example, onsider the following problem:
min

(
m
X
i

=1

(m i)x

ii


m
Xk



i

=1

x +
i;i

= b for k = 0; : : : ; m 1; X is psd
k

(13)

Here, X is an m  m symmetri matrix, and x denotes the entry on row i and olumn j . The
length m ve tor b is given. The abbreviation `psd' stands for `positive semidenite'. The above
optimization problem yields a minimal phase spe tral fa torization of an auto orrelation ve tor
b, see Davidson, Luo and Sturm [4. Problem (13) is stated in terms of an m  m symmetri
matrix of de ision variables, whereas SeDuMi works with a ve tor of de ision variables, as in (7){
(8). This small issue is resolved by using the well known te hnique of ve torization. Ve torization
is implemented by the fun tions ve and mat, whi h are part of SeDuMi. The fun tion ve (X)
reates a long ve tor, by sta king the olumns of the matrix X , as in:
ij

14

>> x = ve ([1, 5, -3; 5, 2, -9; -3, -9, 4)'

x =
1

-3

-9

-3

-9

The inverse of ve is mat. Thus, if x is a ve tor of length n2, then mat(x) onstru ts an n  n
matrix, and lls it with the entries of the ve tor x, starting at the rst olumn.
>> mat(x)
ans =
1
5
-3

5
2
-9

-3
-9
4

The following MATLAB fun tion produ es a standard primal form for problem (13).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

% [At,b, ,K = spe fa (b)

% Creates primal standard form for minimal phase spe tral fa torization.
fun tion [At,b, ,K = spe fa (b)
m = length(b);
% ---------- minimize sum (m-i)*X(i,i) --------- = ve (spdiags((m-1:-1:0)',0,m,m));
% ----- Let e be all-1, and allo ate spa e for the A-matrix ----e = ones(m,1);
At = sparse([,[,[,m^2,m,m*(m+1)/2);
% ---------- sum(diag(X,k)) = b(k) ---------for k = 1:m
At(:,k) = ve (spdiags(e,k-1,m,m));
end
K.s = [m;

The eld K:s = [m tells SeDuMi that we want the m  m matrix mat(x) to be symmetri
positive semidenite. (The `s' in K.s stands for `semidenite'.) We an now solve problem (13)
as follows:
>> b = [2; 0.2; -0.3;
>> [At,b, ,K = spe fa (b);
>> [x,y,info = sedumi(At,b, ,K);

15

SeDuMi by Jos F. Sturm, 1998.

Alg = 1: v- orre tor, theta = 0.250, beta = 0.500
eqs m = 3, order n = 4, dim = 10, blo ks = 2
it :
x
gap delta rate t/maxt feas
0 :
4.00E+00 0.000
1 : 8.14E+00 1.40E+00 0.000 0.3497 0.9000 0.32
2 : 2.29E+00 4.68E-01 0.000 0.3346 0.9000 0.59
3 : 3.42E-01 1.12E-01 0.337 0.2391 0.9000 0.84
4 : 1.26E-01 1.92E-03 0.000 0.0172 0.9900 1.24
* 5 : 1.23E-01 3.97E-06 0.000 0.0021 0.9990 1.00
* 6 : 1.23E-01 8.88E-10 0.000 0.0002 0.9999 1.00
* 7 : 1.23E-01 2.27E-12 0.000 0.0026 0.9990 1.00
iter se onds digits
*x
b*y
7
0.1 10.7 1.2273256502e-01 1.2273256502e-01
|Ax-b| = 0.0e+00, |x|= 2.0e+00, |y|= 7.6e-01

To he k positive semideniteness, we an either use the fun tion eig that is part of MATLAB,
or the fun tion eigK, whi h omes with SeDuMi.

>> [eig(mat(x)), eigK(x,K)

ans =
0.0000
0.0000
2.0000

0.0000
0.0000
2.0000

The use of eigK is more onvenient, espe ially if there are multiple semideniteness onstraints, or if there are also nonnegativity or quadrati one onstraints. SeDuMi will always
produ e symmetri matrix variables, i.e. mat(x) is symmetri . Do not add symmetry onstraints expli itly, as in `x x = 0'. At best, su h onstraints will be removed by SeDuMi
from the A matrix.
However, the dual solution ATy need not be symmetri , as an be seen in the numeri al
example that we are dealing with:
ij

ji

>> mat( -At*y)

ans =
2.0727
0
0

-0.3130
1.0727
0

0.6849
-0.3130
0.0727

16

In this ase, the dual solution is upper triangular, be ause mat( ) is diagonal, and mat(At(:,k))
is upper triangular for all k = 1; 2; : : : ; m. Letting Z = mat( AT y), SeDuMi restri ts the
symmetri part of Z , whi h is (Z + Z T)=2, to be positive semidenite. The fun tion eigK yields
the eigenvalues of the symmetri part. Thus,
>> eigK( -At*y,K)
ans =
1.0583
2.1597
-0.0000

produ es the same result as

>> Z = mat( -At*y); eig(Z+Z')/2

Noti e that problem (13) is equivalent to

x + +x+
min (m i)x
= b for k = 0; : : : ; m 1; X is psd : (14)
2
=1
=1
Namely, x + = (x + + x + )=2, be ause X is symmetri . Thus, we may hange the A matrix
as follows:
(
m
X

ii

i;i

k;i

i;i


m
Xk




i;i

k;i

>> for k=1:size(At,2), Ak = mat(At(:,k)); At(:,k) = ve (Ak+Ak')/2; end

The solutions x and y, as produ ed by SeDuMi, will be exa tly the same. However, sin e
the onstraints in the A matrix have been symmetrized, we nd that mat( At  y) is now
symmetri ; it is the matrix (Z + Z 0)=2.
For SeDuMi, it does not make any dieren e whether the onstraints in A and the obje tive
are symmetrized or not. However, when modeling in the primal standard form, you will probably
nd it more natural to work with upper or lower triangular matri es in A and ; your model will
also use less memory like this. On the other hand, symmetri matri es are more natural when
modeling in the dual form.
There an be multiple positive semideniteness onstraints, in whi h ase K.s lists the orders
of the respe tive matri es. This is analogous to the denition of multiple quadrati onstraints
in K.q and/or K.r. The positive semidenite variables are always the last omponents of x and
AT y, i.e.
K = <K+l  (Q one 


 Q one)  (R one 


 R one)  (S one 


 S one);
where S one denotes the one of positive semidenite matri es. It is easy to remember the above
arrangement, by noting the alphabeti al order of `l', `q', `r' and `s'.
:

17

4 omplex values
In some appli ation areas, su h as signal pro essing, optimization problems may involve omplex
valued data. An example is the Toeplitz Hermitian ovarian e estimation problem, whi h is
dis ussed in Wu, Luo and Wong [30. Other stru tured ovarian e estimation problems, su h as
dis ussed in Deng and Hu [5, an be treated similarly. Given a Hermitian matrix P , the goal
is to nd a Hermitian positive denite matrix Z with a Toeplitz stru ture, su h that kP Z k
is minimal. Thus, the optimization problem is:
minimize =1 (z p )2 + 2 = +1 jz p j2
(15)
su h that Z is Toeplitz, i.e. z = z +1 +1 for all i; j = 1; 2; : : : ; m 1
Z is psd:
If the matrix P has omplex entries, then we will usually also see omplex entries in the optimal
solution Z . Noti e that the Toeplitz property is better modeled in the dual form, than in the
primal form. In fa t, mat(At*y) in (13) is an upper triangular real Toeplitz matrix, and in (14),
it is a symmetri Toeplitz matrix. The MATLAB formulation of (15) therefore resembles the
MATLAB formulation of (13).
F

Pm
i

ii

i;j

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

Pm

ii

ij

ij

;j

% [At,b, ,K = toepest(P)
% Creates dual standard form for Toeplitz- ovarian e estimation
fun tion [At,b, ,K = toepest(P)
m = size(P,1);
% ---------- maximize y(m+1) ---------b = [sparse(m,1); 1;
% ----- Let e be all-1, and allo ate spa e for the A-matrix ----e = ones(m,1);
K.q = [1 + m*(m+1)/2;
K.x omplex = 2:K.q(1);
%Norm-bound entries are omplex valued
At = sparse([,[,[,K.q(1) + m^2,m+1,1 + 2*m^2);
% ---------- onstraints ---------% -y(m+1) >= norm( ve (P) - sum(y_i * Ti) )
(Q one)
% sum(y_i * Ti) is psd
(S one)
% --------------------------------At(:,1) = [sparse(2:(m+1),1,1,K.q(1),1); -ve (speye(m));
= [0; diag(P);
firstk = m+2;
for k = 1:(m-1)
lastk = firstk + m-k-1;
Ti = spdiags(e,k,m,m);

18

23
24
25
26
27
28
29
30
31
32

At(:,k+1) = [sqrt(2) * sparse(firstk:lastk,1,1,K.q(1),1); -2*ve (Ti);

= [ ; sqrt(2) * diag(P,k);
firstk = lastk + 1;
end
At(:,m+1) = [1; sparse(K.q(1) + m^2-1,1); % "obje tive" variable y(m+1)
= [ ; zeros(m^2,1);
% all-0 in the psd-part
K.s = [m;
K.s omplex=1;
%Complex Hermitian PSD
% ---------- y(2:m) omplex, y(1) and y(m+1) real ---------K.y omplex = 2:m;

We have modeled the obje tive fun tion by means of an arti ial variable, y +1, and y2 +1
is bounded from below by the original quadrati obje tive fun tion, using a 1 + m(m + 1)=2dimensional quadrati one. The Toeplitz matrix is modeled as
m

y1 I + 2

X
i

(16)

y +1 T ;
i

=1

where T is all-1 along the k-th upper diagonal, and zero everywhere else. Re all from problem
(13) in Se tion 3.2, that in the real ase, SeDuMi restri ts the symmetri part of mat( AT y) to
be positive semidenite. In the omplex ase, SeDuMi restri ts the Hermitian part, i.e. mat(
AT y) + mat( AT y)0 , to be positive semidenite. Letting Z denote the Hermitian part of (16),
we have
1
Z = y1 I +
(y +1T + y +1T T);
i

X
i

=1

where y +1 denotes the omplex onjugate of y +1. Thus, we have indeed modeled Z as a
Hermitian Toeplitz matrix, and SeDuMi further restri ts it to be positive semidenite, be ause
of the eld K.s. Furthermore, we tell SeDuMi to allow omplex values for y2; y3; : : : ; y , by
setting K.y omplex = 2:m. Remark that unlike K.l, K.q, K.r and K.s, the eld K.y omplex is
not involved in the denition of the symmetri one K in (7){(8).
The following lines show how to solve problem (15), for a parti ular 3  3 Hermitian matrix
P , whi h is neither Toeplitz, nor positive semidenite.
i

>> i = sqrt(-1);
>> P = [4, 1+2*i, 3-i; 1-2*i, 3.5, 0.8+2.3*i; 3+i, 0.8-2.3*i, 4
P =
4.0000
1.0000 - 2.0000i

1.0000 + 2.0000i
3.5000

3.0000 - 1.0000i
0.8000 + 2.3000i

19

3.0000 + 1.0000i

0.8000 - 2.3000i

4.0000

>> [At,b, ,K = toepest(P);

>> [x,y,info = sedumi(At,b, ,K);
>> z = -At*y; Z = mat(z(K.q+1:length(z))); Z = (Z+Z')/2
Z =
4.2827
0.8079 - 1.7342i
2.5574 + 0.7938i

0.8079 + 1.7342i
4.2827
0.8079 - 1.7342i

2.5574 - 0.7938i
0.8079 + 1.7342i
4.2827

>> eigK(z,K)'
ans =
-0.0000

2.0517

0.0000

7.2810

5.5670

Instead of using the mat() fun tion, one may use the ellK() fun tion as follows:
>> z = ellK( -At*y,K); Z=z.s{1}; Z=(Z+Z')/2

We have found the optimal positive semidenite Toeplitz matrix Z , whi h has eigenvalues 0,
7:281 and 5:567. Che king the obje tive values reveals a new phenomenon:
>> [ '*x; b'*y
ans =
-1.4508 - 0.2428i
-1.4508

The value of Hx, where H means omplex onjugate transpose, may no longer be real, and the
same is true for bHy in general. Obviously, we annot minimize or maximize omplex valued
fun tions. Instead, SeDuMi minimizes Re Hx in the primal, and maximizes Re bHy in the dual.
Here, Re stands for real part. In the sequel, we will also use the notation Im , to denote the
imaginary part.
If we make K.y omplex = [, then all dual multipliers y are restri ted to be real.
i

>> K.y omplex = [;

>> [x2,y2,info2=sedumi(At,b, ,K);
>> [ '*x2; b'*y2

20

ans =
-4.5592 - 0.3816i
-4.5592

Clearly, by restri ting y to be real, the dual optimal value Re bHy = y +1 gets worse. Apparently, something has hanged in the primal problem as well, sin e the primal optimal value
has improved from 1:4508 to 4:5592. The dieren e is in the `Ax = b' restri tion, as the
following lines show:
m

>> [b-At'*x b-At'*x2

ans =
0.0000
-0.0000
0.0000 + 0.0000i -0.0000 + 1.8863i
-0.0000
-0.0000 - 0.4387i
0
0

The restri tion `Ax = b' is interpreted by SeDuMi as

aH x = b if i 2 K:y omplex
(17)
Re aHx = b otherwise:
By making K.y omplex = [, we therefore removed the restri tions on Im Ax, and impli itly
added the restri tion that Im y = 0. Complex y-variables in the dual form orrespond with
omplex equality onstraints in the primal form.
If size(A,2) = length(b), then the primal feasibility requirements are A'*x = b, using
the omplex onjugate transpose AH.
The eld K.s omplex ontains a list of the PSD matrix variables, of order K.s(K.s omplex),
whi h are restri ted to be Hermitian positive semidenite matri es. For the remaining matrix
variables, the primal x-variables are restri ted to be real symmetri positive semidenite, whereas
the dual sla k variables are restri ted to be positive semi-denite on the real part only (the dual
imaginary part is then unrestri ted).
The eld K.x omplex lists the primal x-variables that are allowed to have a nonzero imaginary part. For the free and nonnegative x- omponents, ths imaginary part is then unrestri ted
in sign. For example, the restri tion `x  0' is interpreted by SeDuMi as
x 2 <+ if i 62 K:x omplex
(18)
Re x  0 if i 2 K:x omplex:
(

21

A similar onvention holds for the rst entry in a q-se ond order one. The remaining entries
in a se ond order one that are listed in K.x omplex are simply omplex variables that appear
in the norm-bound restri tion. Only entries in the f,l,q,r- ones an be listed in K.x omplex;
the matrix variables are handled by the eld K.s omplex.
On the dual side, K.x omplex lists the equality and nonnegativity onstraints for whi h the
restri tion Im aTy = 0 must be imposed on the imaginary part. This interpretation also
works for the rst entry in a se ond order q- one. The remaining entries in a se ond order
one that are listed in K.x omplex are simply omplex variables that appear in the norm-bound
restri tion (this is ompletely symmetri to the primal).
For sensitivity analysis, it is interesting to note that Re (
)H x is a supergradient for the
optimal value fun tion, under perturbations of the form + t
, whereas Re (
b)H y is a subgradient of the optimal value fun tion under perturbation of b. For a dis ussion of sensitivity
analysis in (real symmetri ) semidenite programming, see Goldfarb and S heinberg [12.
i

5 optional settings
By default, SeDuMi lls your terminal s reen with some output on erning its iterative progress.
This an be an annoying feature, in parti ular if SeDuMi is merely used as a subroutine within
a larger program. To suppress the on-s reen output of SeDuMi, it su es to set an optional
parameter, pars.fid, to zero.
>> load truss1
>> pars.fid = 0;
>> [x,y,info = sedumi(At,b, ,K,pars);

The stru ture pars is not only used for suppressing iterative output of SeDuMi. It an ontain
a number of optional elds, whi h we will dis uss in this se tion.
The abbreviation `d' in pars.fid stands for `le identier': the output of SeDuMi will be
sent to the le whose le identier is pars.fid. The le identier for the null-devi e is 0, whi h
is useful for suppressing output, and for the terminal s reen it is 1. Output an also be redire ted
to a le, e.g.
>> pars.fid = fopen('truss1.out','w');
>> [x,y,info=sedumi(At,b, ,K,pars);
>> f lose(pars.fid); pars.fid = 1;

With the above lines, the output is redire ted to the le `truss1.out', as an be he ked with the
ommand dbtype truss1.out.
SeDuMi uses a variant of the primal{dual interior point method, whi h is known as the
entering{predi tor{ orre tor method [25. There are 3 variants of the entering{predi tor{
orre tor method implemented, whi h an be sele ted with the eld pars.alg. With pars.alg
22

Wide region vplot

normalized vvalues

1.5

0.5

6
iterations

10

12

Reduction rates
1

reduction rate

0.8
0.6
0.4
0.2
0

5
6
iterations

10

11

Figure 1: Plot produ ed by setting pars.vplot = 1.

= 0,

you sele t a longest-step algorithm, without any se ond order orre tor. To enhan e the
algorithm with a se ond order orre tor, you an either set pars.alg = 1 or pars.alg = 2.
With pars.alg = 1, the se ond order orre tor is derived by linearization of the so- alled vvalues, whereas pars.alg = 2 uses linearization of the squared v-values, whi h is also known as
xz -linearization. For linear programming, xz -linearization results in the well-known Mehrotra's
orre tor [17. In all three variants, the entering step is determined by the entral region parameter, pars.theta. This parameter an take any value in (0; 1. At one extreme, pars.theta
= 1 results in path{following, whi h typi ally involves relatively short step lengths. Setting
pars.theta to a smaller value, su h as 1=4, makes the algorithm work in the neighborhood of
a full dimensional entral region, and this typi ally allows for larger step lengths, see Sturm
and Zhang [26. The size of the neighborhood is ontroled by the parameter pars.beta, whi h
an be assigned any value in (0; 1). In the output of SeDuMi on the terminal s reen, there is
a olumn labeled `delta', whi h lists the a tual distan e to the entral region in ea h iteration.
The step length will always be su h that this is at most pars.beta. The ratio of the a tual step
length and the maximal steplength to the boundary of the one K is listed in the olumn labeled
`t/maxt'. For some iterations, an asterisk (`*') appears in front of the output line. At these iterations, the residual ve tor of the self{dual model has been re omputed (to avoid a umulation
of numeri al errors).
For resear h purposes, SeDuMi an produ e a plot of the iterative v{values. This feature is
a tivated by setting pars:vplot = 1. For problem truss1, this results in the plots of Figure 1.
For ea h iteration, the rst plot shows all the v{values, divided by the mean of the v-ve tor
in that iteration. It also gives a horizontal line at value 1, representing the entral path, and
a horizontal line at the entral region threshold, pars.theta = 1/4. Any v-values below this
23

threshold will be orre ted by the entering omponent in the su eeding iteration. The se ond
plot shows the rate of linear redu tion, whi h is simply
duality gap in iteration k
duality gap in iteration k 1 :
The rate of linear redu tion is also listed in the olumn `rate' in the on-s reen output of SeDuMi,
and the iterative duality gap is listed under `gap'. This is the duality gap in an arti ial self{dual
model, in whi h your original model is embedded by SeDuMi, using the te hnique of Ye, Todd
and Mizuno [31. The self{dual model gives rise to a feasibility indi ator, listed in the olumn
`feas'. Ideally, the indi ator onverges to +1 for feasible problems, and to 1 for (primal and/or
dual) infeasible problems.
Termination ontrol is provided by the elds maxiter, and eps in the pars stru ture. SeDuMi
will terminate su essfully if it nds a solution that violates feasibility and optimality requirements by no more than pars.eps. The parameter pars.maxiter allows you to set a maximum
on the number of iterations. By default, pars.eps = 1E-9 and pars.maxiter = 150. A possible experiment with these parameters is to set pars.eps = 0 in the example of minimizing
1=x1 , whi h was dis ussed in Se tion 3.1.
I thank T. Terlaky for en ouraging me to write this manual, and for pointing
out a bug in the rst publi release of SeDuMi. P. Apkarian, M. Bengtsson, T.N. Davidson, F. Glineur,
V. Prodanovi and A. Ross helped to improve the software, by providing bug reports and suggestions on
the rst release of the software. Numerous other users have ontributed in the same manner sin e.Two
anonymous referees have ontributed in improving this do ument.
A knowledgments.

Referen es
[1 F. Alizadeh, J.A. Haeberly, M.V. Nayakkankuppann, M. Overton, and S. S hmieta. SDPPa k user's guide. New York University, New York, USA, 1997.
[2 B. Bor hers. CSDP, a C library for semidenite programming. Te hni al report, New
Mexi o Te h Mathemati s Fa ulty, USA, 1997.
[3 N. Brixius and F.A. Potra. Sdpha: A MATLAB implementation of homogeneous interiorpoint algorithms for semidenite programming. Te hni al report, Department of Computer
S ien e, University of Iowa, Iowa City, Iowa, USA, 1998.
[4 T.N. Davidson, Z.-Q. Luo, and J.F. Sturm. A (primal form of the) positive real lemma for
FIR systems. Te hni al report, Communi ations Resear h Laboratory, M Master University, Hamilton, Canada, 1998. To appear.

24

[5 S. Deng and H. Hu. Computable error bounds for semidenite programming. Te hni al
report, Department of Mathemati al S ien es, Northern Illinois University, DeKalb, Illinois,
USA, 1996.
[6 S. Dussy and L. El Ghaoui. Multiobje tive robust ontrol toolbox for LMI-based ontrol.
Te hni al report, Laboratoire de Mathematiques Appliquees, ENSTA, Paris, Fran e, 1997.
[7 L. El Ghaoui and H. Lebret. Robust solutions to least-squares problems with un ertain
data. SIAM Journal on Matrix Analysis and Appli ations, 18(4):1035{1064, 1997.
[8 L. El Ghaoui, R. Nikoukhah, and F. Delebe que. LMITOOL: A front-end for LMI optimization, user's guide. Laboratoire de Mathematiques Appliquees, ENSTA, Paris, Fran e,
1995.
[9 J. Faraut and A. Koranyi. Analysis on Symmetri Cones. Oxford Mathemati al Monographs. Oxford University Press, New York, 1994.
[10 K. Fujisawa, M. Kojima, and K. Nakata. Exploiting sparsity in primal-dual interior-point
methods for semidenite programming. Mathemati al Programming, 79:235{253, 1997.
[11 K. Fujisawa, M. Kojima, and K. Nakata. SDPA (semidenite programming algorithm)
user's manual | version 4.10. Te hni al report, Dept. of Information S ien es, Tokyo
Institute of Te hnology, 2-12-1 Oh-Okayama, Meguro-ku, Tokyo 152, Japan, 1998.
[12 D. Goldfarb and K. S heinberg. On parametri semidenite programming. Te hni al report,
Columbia University, Department of IEOR, New York, USA, 1997.
[13 D. Goldfarb and K. S heinberg. Stability and e ien y of matrix fa torizations in interior
point methods. Conferen e presentation, HPOPT IV Workshop, June 16{18, Rotterdam,
The Netherlands.
[14 M.S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret. Appli ations of se ond{order one
programming. Te hni al report, Information Systems Lab, Stanford University, 1997. To
appear in Linear Algebra and Appli ations.
[15 Z.-Q. Luo, J.F. Sturm, and S. Zhang. Duality and self-duality for oni onvex programming. Te hni al Report 9620/A, E onometri Institute, Erasmus University Rotterdam,
Rotterdam, The Netherlands, 1996.
[16 Z.-Q. Luo, J.F. Sturm, and S. Zhang. Duality results for oni onvex programming. Te hni al Report 9719/A, E onometri Institute, Erasmus University Rotterdam, Rotterdam,
The Netherlands, 1997.
[17 S. Mehrotra. On the implementation of a primal{dual interior point method. SIAM Journal
on Optimization, 2:575{601, 1992.
25

[18 S. Mehrotra and Y. Ye. Finding an interior point in the optimal fa e of linear programs.
Mathemati al Programming, 62:497{515, 1993.
[19 H.D. Mittelmann.
Several SDP- odes on problems from SDPLIB.
ftp://plato.la.asu.edu/pub/sdplib.txt, 1998.
[20 R.D.C. Monteiro and S. Mehrotra. A general parametri analysis approa h and its impli ation to sensitivity analysis in interior point methods. Mathemati al Programming, 72:65{82,
1996.
[21 Y. Nesterov and M.J. Todd. Self{s aled barriers and interior{point methods for onvex
programming. Mathemati s of Operations Resear h, 22(1):1{42, 1997.
[22 D. Peau elle, D. Henrion and Y. Labit. User's guide for SeDuMi Interfa e 1.01: Solving LMI problems with SeDuMi. LAAS - CNRS, Toulouse, Fran e. Available from
http://www.laas.fr/ peau ell/SeDuMiInt.html.
[23 C. Roos, T. Terlaky, and J.-Ph. Vial. Theory and algorithms for linear optimization. An
interior point approa h. Series in dis rete mathemati s and optimization. John Wiley &
Sons, New York, 1997.
[24 A.M. Ross. Optimization over ones: SDPpa k versus SeDuMi. Te hni al Report 1998-1022, Department IEOR, University of California at Berkeley, 1998.
[25 J.F. Sturm. Primal{Dual Interior Point Approa h to Semidenite Programming, volume
156 of Tinbergen Institute Resear h Series. Thesis Publishers, Amsterdam, The Netherlands, 1997.
[26 J.F. Sturm and S. Zhang. On a wide region of enters and primal{dual interior point
algorithms for linear programming. Mathemati s of Operations Resear h, 22(2):408{431,
1997.
[27 M.J. Todd and Y. Ye. Approximate Farkas lemmas and stopping rules for iterative
infeasible{point algorithms for linear programming. Mathemati al Programming, 81:1{21,
1998.
[28 K.C. Toh, M.J. Todd, and R.H. Tutun u. SDPT3 - a MATLAB pa kage for semidenite
programming. Te hni al report, S hool of Operations Resear h and Industrial Engineering,
Cornell University, Itha a, NY, USA, 1996.
[29 L. Vandenberghe and S. Boyd. SP: Software for semidenite programming. Information
Systems Laboratory, Ele tri al Engineering Department, Stanford University, Stanford,
USA, 1994.
26

[30 S. Wu, Z.-Q. Luo, and K. Wong. Dire tion nding for oherent sour es via Toeplitz approximation. Te hni al report, Communi ations Resear h Laboratory, M Master University,
Hamilton, Ontario, Canada, 1997.
[31 Y. Ye, M.J. Todd, and S. Mizuno. An O(pnL)-iteration homogeneous and self{dual linear
programming algorithm. Mathemati s of Operations Resear h, 19:53{67, 1994.
[32 Y. Zhang. Solving large{s ale linear programs by interior{point methods under the MATLAB environment. Te hni al report, Department of Computational and Applied Mathemati s, Ri e University, Houstan, Texas, USA, 1997.
A Calling Sequen e
The primal anoni al form for solving optimization problems with SeDuMi is
minf T x j Ax = b; x 2 Kg;
and the dual anoni al form is
maxfbTy j ATy 2 Kg:
The general alling sequen e for solving the above primal{dual pair is
[x,y,info = sedumi(A,b, ,K,pars)

Here, K is a MATLAB stru ture to dene the symmetri one K; it onsists of the following
(optional) elds:
K.f The number of free primal variables, i.e. the number of dual equality onstraints
K.l The number of nonnegativity onstraints
K.q A list of dimensions of quadrati one onstraints
K.r A list of dimensions of rotated quadrati one onstraints
K.x omplex A list of primal variables in the f,l,q,r blo ks that are allowed to have a nonzero
imaginary part. The imaginary parts of the asso iated dual onstraints are then expli itly
restri ted.
K.s A list of orders of positive semideniteness onstraints
K.s omplex A list of matrix variables that are restri ted to be Hermitian positive semidenite.
27

This eld is not related to K. It lists the omponents of the y-variables that are
omplex valued. Equivalently, it lists the primal equality onstraints (Ax) = b that have
to be satised not only in their real parts, but also in their imaginary parts.
The stru ture K denes K to be
K = <K f  <K+l  (Q one 


 Q one)  (R one 


 R one)  (S one 


 S one);
In total, the number of Q one (R one, S one) omponents is length(K.q) (length(K.r),
length(K.s)). The ith Q one is
Q one = f(x1 ; x2 ) 2 <  C K q[i 1 j x1  kx2 kg;
where C denotes the spa e of omplex n{tuples. The j th R one is
R one = (x1 ; x2 ; x3) 2 <  <  C K r[j 2 x1 x2  12 kx3 k2 ; x1 + x2  0 ;
and the kth S one is
S one = x 2 C K s[k j mat(x) is Hermitian positive semidenite ;
for primal omponents. For dual omponents z = ATy, we use a slightly milder denition of
S one, viz.
S one = z 2 C K s[k j mat(z) + mat(z)0 is positive semidenite :
If k is not listed in K.s omplex, then the denition in the dual is even milder, namely,
S one = z 2 C K s[k j Re mat(z) + mat(z)0 is positive semidenite :
The length of the ve tor should be
K.y omplex

length( ) = K.f + K.l + sum(K.q) + sum(K.r) + sum(K.s.^2)

If the data (A; b; ) is real valued, then x and y will also be real valued.
The parameter pars is a MATLAB stru ture, onsisting of the following (optional) elds:
pars.fid By default, pars.fid=1, whi h tells SeDuMi to produ e iterative statisti s on the
s reen. If pars.fid=0, then SeDuMi runs quietly, i.e. no s reen output. In general, output
is written to the le or devi e that is identied by the le handle pars.fid. A le handle
is assigned to a le by the MATLAB fun tion fopen, as in
pars.fid = fopen('truss1.out','w').

28

pars.alg By

default, pars.alg=2. If pars.alg=0, then a rst-order algorithm is used, whi h

is not re ommendable. If pars.alg=1, then SeDuMi uses the entering-predi tor- orre tor
algorithm with v-linearization. If pars.alg=2 then xz-linearization is used in the orre tor,
similar to Mehrotra's algorithm. All 3 algorithms are spe ial instan es of the generi wideregion algorithm, as dis ussed in Chapter 7 of Sturm [25.
pars.theta, pars.beta By default, pars.theta=0.25 and pars.beta=0.5. These are the
wide region and neighborhood parameters. Valid hoi es are 0 <  <= 1 and 0 <  < 1.
pars.stepdif, pars.w By default, pars.stepdif=1 and pars.w = [1 1. This means that
primal/dual step length dierentiation is enabled (disabled if pars.stepdif=0). The priorities of the relative primal, dual and gap residuals are weighted as w(1):w(2):1, in order
to nd the optimal step dierentiation.
pars.vplot If this eld is 1, then SeDuMi produ es a fan y v -plot, for resear h purposes.
Default: vplot = 0.
pars.eps The desired a ura y.
pars.bigeps The required a ura y to get info.numerr < 2.
pars.maxiter Maximum number of iterations, before termination.
pars.denq Proportion of x(i)'s for whi h the sparsity in A(:,i) is onsidered normal. Default:
0.75.
pars.denf A olumn is treated as dense if it has pars.denf times more nonzeros than normal.
Default: 10.
pars.stopat Enters MATLAB debugging mode at the beginning of iteration pars.stopat.
Default: -1.
pars. g Various parameters for ontroling the Pre onditioned onjugate gradient method
(CG), whi h is only used if results from Cholesky are ina urate. Type `help sedumi'
for details.
pars. hol Various parameters for ontroling the Cholesky solve. Type `help sedumi' for details.
The output parameter info is a MATLAB stru ture, with the following elds:
info.pinf and info.dinf The feasibility status of the primal-dual problem pair, as dete ted
by SeDuMi. There are three ases:
1. pinf = dinf = 0 Then x and y are (approximate) optimal solutions, i.e. Ax = b,
x 2 K, AT y 2 K, and T x  bT y (approximately).
29

2.

pinf = 1 Primal is infeasible, i.e. fx 2 K j Ax = bg = ;. Then y is a Farkas-type

solution, i.e. bTy > 0 and ATy 2 K.
3. dinf = 1 Dual is infeasible, i.e. fy j AT y 2 Kg = ;g. Then x is a Farkas-type
solution, i.e. T x < 0, Ax = 0 and x 2 K.
info.numerr A positive value of info.numerr means that SeDuMi terminated without a hieving the desired a ura y, be ause of numeri al problems. If info.numerr = 1 then the
results are merely ina urate: the solution has still a hieved the a ura y denoted by
pars.bigeps, whi h is 1E-3 by default. If info.numerr = 2 then SeDuMi failed ompletely.

30