AN EXACT DUALITY THEORY FOR SEMIDEFINITE

PROGRAMMING BASED ON SUMS OF SQUARES

IGOR KLEP AND MARKUS SCHWEIGHOFER

Abstract. Farkas’ lemma is a fundamental result from linear programming

arXiv:1207.1691v2 [math.OC] 11 Nov 2012

providing linear certificates for infeasibility of systems of linear inequalities.

In semidefinite programming, such linear certificates only exist for strongly in-
feasible linear matrix inequalities. We provide nonlinear algebraic certificates
for all infeasible linear matrix inequalities in the spirit of real algebraic geom-
etry: A linear matrix inequality A(x)  0 is infeasible if and only if −1 lies
in the quadratic module associated to A. We also present a new exact duality
theory for semidefinite programming, motivated by the real radical and sums
of squares certificates from real algebraic geometry.

1. Introduction
A linear matrix inequality (LMI) is a condition of the form
n
X
A(x) = A0 + xi Ai  0 (x ∈ Rn )
i=1

where the Ai are symmetric matrices of the same size and one is interested in the
solutions x ∈ Rn making A(x) positive semidefinite (A(x)  0). The solution
set to such an inequality is a closed convex semialgebraic subset of Rn called a
spectrahedron or an LMI domain. Optimization of linear objective functions over
spectrahedra is called semidefinite programming (SDP) [BV96, To01, WSV00], and
is a subfield of convex optimization. In this article, we are concerned with the
duality theory of SDP from a viewpoint of a real algebraic geometer, and with the
important SDP feasibility problem: When is an LMI feasible; i.e., when is there an
x ∈ Rn satisfying A(x)  0?
A diagonal LMI, where all Ai are diagonal matrices, is just a system of linear
inequalities, and its solution set is a polyhedron. Optimization of linear objective
functions over polyhedra is called linear programming (LP). The ellipsoid method
developed by Shor, Yudin, Nemirovskii and Khachiyan showed at the end of the
1970s for the first time that the LP feasibility problem (and actually the problem
of solving LPs) can be solved in (deterministically) polynomial time (in the bit
model of computation assuming rational coefficients) [Sr86, Chapter 13]. Another
breakthrough came in the 1980s with the introduction of the more practical interior

Date: November 10, 2012.

2010 Mathematics Subject Classification. Primary 13J30, 90C22, 15A22, Secondary 14P10,
15A48, 15A39.
Key words and phrases. linear matrix inequality, LMI, spectrahedron, semidefinite programming,
SDP, quadratic module, infeasibility, duality theory, real radical, Farkas’ lemma.
This research was supported through the programme “Research in Pairs” (RiP) by the Mathema-
tisches Forschungsinstitut Oberwolfach in 2010. The first author was supported by the Faculty
Research Development Fund (FRDF) of The University of Auckland (project no. 3701119), and
partially supported by the Slovenian Research Agency under Project no. J1-3608 and Program
no. P1-0222. Part of the research was done while the first author was on leave from the University
of Maribor, and held a visiting professorship at the Universität Konstanz in 2011.
1
2 IGOR KLEP AND MARKUS SCHWEIGHOFER

point methods by Karmarkar and their theoretical underpinning by Nesterov and

Nemirovskii [NN94, Ne07].
The motivation to replace the prefix “poly” by “spectra” is to replace “many”
values of linear polynomials (the diagonal values of A(x)) by the “spectrum” of
A(x) (i.e., the set of its eigenvalues). The advantage of LMIs over systems of
linear inequalities (or of spectrahedra over polyhedra, and SDP over LP, respec-
tively) is a considerable gain of expressiveness which makes LMIs an important
tool in several areas of applied and pure mathematics. Many problems in control
theory, system identification and signal processing can be formulated using LMIs
[BEFB94, Par00, HG05, Du07, Ce10]. Combinatorial optimization problems can
often be modeled or approximated by SDPs [Go97]. LMIs also find application in
real algebraic geometry for finding sums of squares decompositions of polynomials
[Las10, Ma08]. There is even a hierarchy of SDP approximations to polynomial
optimization problems [Lau09] consisting of the so-called Lasserre moment relax-
ations. In this article, rather than trying to solve polynomial optimization problems
by using SDP, we borrow ideas and techniques from real algebraic geometry and
polynomial optimization in order to get new results in the theory of semidefinite
programming.
The price to pay for the increased expressivity of SDP is that they enjoy some
less good properties. First of all, the complexity of solving general SDPs is a very
subtle issue. For applications in combinatorial optimization, it typically follows
from the theory of the ellipsoid method [Sr86] or interior point methods [NN94] that
the translation into SDP yields a polynomial time algorithm (see [dK02, Section
1.9] for exact statements). However, the complexity status of the LMI feasibility
problem (the problem of deciding whether a given LMI with rational coefficients has
a solution) is largely unknown. What is known is essentially only that (in the bit
model) LMI feasibility lies either in NP∩co-NP or outside of NP∪co-NP. Therefore
it cannot be NP-complete unless NP = co-NP. This follows from our work below,
but has been already proven by Ramana [Ra97] in 1997; Porkolab and Khachiyan
[PK97] have proved that either for fixed number variables or for fixed matrix size,
the LMI feasibility problem lies in P. Second, the standard (Lagrange-Slater) dual
of a semidefinite program works well when the feasible set is full-dimensional (e.g.
if there is x ∈ Rn with A(x) ≻ 0). However, in general, strong duality can fail
badly, and there is no easy way of reducing to the full-dimensional case. Even the
corresponding version of Farkas’ lemma fails for SDP.
We prove in this paper a nonlinear Farkas’ lemma for SDP by giving algebraic
certificates for infeasibility of an LMI. Furthermore, we present a new exact duality
theory for SDP. The inspiration for our sums of squares dual comes from real
algebraic geometry, more precisely from sums of squares representations and the
Real Nullstellensatz [Ma08, PD01, Sc09]. We believe that this new connection will
lead to further insights in the future.
Reader’s guide. The paper is organized as follows: We fix terminology and nota-
tion in Section 2. Our main results, including the sums of squares dual of an SDP,
are presented in Section 3. The two crucial ingredients needed in the proof are a
low-dimensionality certificate for spectrahedra (see Subsection 3.3), and a new Pos-
itivstellensatz for linear polynomials nonnegative on a spectrahedron (see Theorem
3.4.1 in Subsection 3.4). Finally, in Subsection 3.7, we present the sums of squares
dual (Dsos ) of an SDP, and Theorem 3.7.3, an algorithmic variant of the linear
Positivstellensatz. Section 4 contains applications of these results. For example, in
Subsection 4.2 we interpret Theorems 3.4.1 and 3.7.3 in the language of real alge-
braic geometry, and in Subsection 4.3 we prove a nonlinear Farkas’ lemma for SDP
by giving nonlinear algebraic certificates for infeasibility of an LMI. These results
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 3

use quadratic modules from real algebraic geometry. As a side product we intro-
duce a hierarchy for infeasibility of LMIs, whose first stage coincides with strong
infeasibility. Subsection 4.4 contains certificates for boundedness of spectrahedra
which are used to give a Putinar-Schmüdgen-like Positivstellensatz for polynomi-
als positive on bounded spectrahedra. Finally, the article concludes with two brief
sections containing examples illustrating our results, and an application to positive
linear functionals.

2. Notation and terminology

We write N := {1,2, . . . } and R for the sets of natural and real numbers, respec-
tively. Let R be a unital commutative ring. For any matrix A over R, we denote
by A∗ its transpose. Then SRm×m := {A ∈ Rm×m | A = A∗ } denotes the set of
all symmetric m × m matrices. Examples of these include hermitian squares, i.e.,
elements of the form A∗ A for some A ∈ Rn×m .
Recall that a matrix A ∈ Rm×m is called positive semidefinite (positive definite)
if it is symmetric and v ∗ Av ≥ 0 for all vectors v ∈ Rm (v ∗ Av > 0 for all v ∈
Rm r {0}). For real matrices A,B ∈ Rm×m , we write A  B (respectively A ≺ B)
to express that B − A is positive semidefinite (respectively positive definite). We
denote by SRm×m 0 and SRm×m ≻0 the convex cone of all positive semidefinite and
positive definite matrices of size m, respectively.

2.1. Matrix polynomials. Let x = (x1 , . . . ,xn ) be an n-tuple of commuting vari-

ables and R[x] the polynomial ring. With R[x]k we denote the vector space of all
polynomials of degree ≤ k. A (real) matrix polynomial is a matrix whose entries
are polynomials from R[x]. It is linear or quadratic if its entries are from R[x]1
or R[x]2 , respectively. A matrix polynomial is an element of the ring R[x]m×n for
some m,n ∈ N, and can be viewed either as a polynomial with matrix coefficients,
or as a matrix whose entries are polynomials. For a comprehensive treatment of
the theory of matrix polynomials we refer the reader to the book [GLR82] and the
references therein.

2.2. Linear pencils and spectrahedra. We use the term linear pencil as a syn-
onym and abbreviation for symmetric linear matrix polynomial. A linear pencil
A ∈ R[x]α×α can thus be written uniquely as
A = A0 + x1 A1 + · · · + xn An
with Ai ∈ SRα×α . If A ∈ R[x]α×α is a linear pencil, then the condition A(x)  0
(x ∈ Rn ) is called a linear matrix inequality (LMI) and its solution set
SA := {x ∈ Rn | A(x)  0}
is called a spectrahedron (or also an LMI set ). We say that A is infeasible if SA = ∅,
and A is feasible if SA 6= ∅.
Obviously, each spectrahedron is a closed convex semialgebraic subset of Rn .
Optimization of linear objective functions over spectrahedra is called semidefinite
programming (SDP) [BV96, To01, WSV00]. If A ∈ R[x]α×α is a diagonal linear pen-
cil, then A(x)  0 (x ∈ Rn ) is just a (finite) system of (non-strict) linear inequalities
and SA is a (closed convex) polyhedron. Optimization of linear objective functions
over polyhedra is called linear programming (LP). The advantage of LMIs over sys-
tems of linear inequalities (or of spectrahedra over polyhedra, and SDP over LP, re-
spectively) is a considerable gain of expressiveness which makes LMIs an important
tool in many areas of applied and pure mathematics [BEFB94, Go97, Par00, Las10].
SDPs can be solved efficiently using interior point methods [NN94, St00, dK02].
4 IGOR KLEP AND MARKUS SCHWEIGHOFER

2.3. Sums of squares. Another example of symmetric matrix polynomials that

are of special interest to us are sums of hermitian squares in R[x]m×m . They are
called sos-matrices. More explicitly, S ∈ R[x]m×m is an sos-matrix if the following
equivalent conditions hold:
P ∗ P for some s ∈ N and some P ∈ R[x]s×m ;
(i) S = P
(ii) S = Pri=1 Q∗i Qi for some r ∈ N and Qi ∈ R[x]m×m ;
(iii) S = si=1 vi vi∗ for some s ∈ N and vi ∈ R[x]m .
A special case are sums of squares in the polynomial ring R[x]. They are called
sos-polynomials and they are nothing else but sos-matrices of size 1. We denote the
set of all sos-matrices (of any size) over R[x] by
Σ2 := P ∗ P | s,m ∈ N, P ∈ R[x]s×m .


In particular, Σ2 ∩ R[x] is the set of sos-polynomials.

Note that an sos-matrix S ∈ R[x]m×m is positive semidefinite on Rn but not vice-
versa, since e.g. a polynomial nonnegative on Rn is not necessarily sos [Ma08, PD01].
√
2.4. Radical
√ ideals. Recall that for any ideal I ⊆ R[x], its radical I and its real
radical R I are the ideals defined by
√
I := {f ∈ R[x] | ∃k ∈ N : f k ∈ I} and
√
I := {f ∈ R[x] | ∃k ∈ N : ∃s ∈ Σ2 ∩ R[x] : f 2k + s ∈ I}.
R

√ √
An ideal I ⊆ R[x] is called radical if I = I and real radical if I = R I. We refer
the reader to [BCR98] for further details.

3. Duality theory of semidefinite programming

In this section we present a sums of squares inspired dual for SDP, see Subsection
3.7. It is derived from two core ingredients, which are of independent interest. First,
Proposition 3.3.1 below allows us to detect low-dimensionality of spectrahedra,
thus leading to a codimension-reduction technique. Second, Theorem 3.4.1 gives
a nonlinear algebraic certificate (i.e., a Positivstellensatz) for linear polynomials
nonnegative on a spectrahedron.

3.1. Weakly feasible and weakly infeasible linear pencils. Recall that the
linear pencil A(x) ∈ R[x]1α×α is infeasible if SA = ∅. We call A strongly infeasible
if
dist {A(x) | x ∈ Rn }, SRα×α


0 > 0,
and weakly infeasible if it is infeasible but is not strongly infeasible. A feasible linear
pencil A is strongly feasible if there is an x ∈ Rn such that A(x) ≻ 0, and weakly
feasible otherwise. To A we associate the convex cone
n X o
CA := c + u∗i Aui | c ∈ R≥0 , ui ∈ Rα
i
n o
= c + tr(AS) | c ∈ R≥0 , S ∈ SRα×α
0 ⊆ R[x]1 .

Note that CA consists of linear polynomials nonnegative on SA .

The following is an extension of Farkas’ lemma from LP to SDP due to Sturm
[St00, Lemma 2.18]. We include its simple proof based on a Hahn-Banach separation
argument.
Lemma 3.1.1 (Sturm). A linear pencil A is strongly infeasible if and only if −1 ∈
CA .
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 5

Proof. Suppose
n
X
A = A0 + xi Ai ∈ R[x]α×α
1
i=1
is strongly infeasible. Then the non-empty convex sets {A(x) | x ∈ Rn } and SRα×α0
can be strictly separated by an affine hyperplane (since their Minkowski sums with a
small ball are still disjoint and can therefore be separated [Ba02, Theorem III.1.2]).
This means that there is a non-zero linear functional
ℓ : SRα×α → R
α×α
and γ ∈ R, with ℓ(SR0 ) ⊆ R>γ and ℓ({A(x) | x ∈ Rn }) ⊆ R<γ . Choose
α×α
B ∈ SR such that
ℓ(A) = tr(AB)
α×α α×α
for all A ∈ SR . Since ℓ(SR0 ) is bounded from below, by the self-duality of
the convex cone of positive semidefinite matrices, 0 6= B  0. Similarly, we obtain
ℓ(Ai ) = 0 for i ∈ {1, . . . ,n}. Note that γ <P0 since 0 = ℓ(0) ∈ R>γ so we can
assume ℓ(A0 ) = −1 by scaling. Writing B = i ui u∗i with ui ∈ Rα , we obtain
X X
−1 = ℓ(A0 ) = ℓ(A(x)) = tr(A(x) ui u∗i ) = u∗i A(x)ui .
i i
n ∗
P
for all x ∈ R . Hence −1 = i ui Aui
∈ CAP
.
Conversely, if −1 ∈ CA , i.e., −1 = c + i u∗i Aui for some c ≥ 0 and ui ∈ Rα ,
α×α
then with B := i ui u∗i ∈ SR0
P
we obtain a linear form
ℓ : SRα×α → R, A 7→ tr(AB)
satisfying ℓ(SRα×α
0 ) ⊆ R≥0 and ℓ({A(x) | x ∈ Rn }) = {−1 − c} ⊆ R≤−1 . So A is
strongly infeasible.
Lemma 3.1.2. Let A ∈ SR[x]α×α 1 be an infeasible linear pencil. Then the following
are equivalent:
(i) A is weakly infeasible;
(ii) SA+εIα 6= ∅ for all ε > 0.
Proof. Since all norms on a finite-dimensional vector space are equivalent, we can
without loss of generality use the operator norm on Rα×α .
Suppose that (i) holds and ε > 0 is given. Choose B ∈ SRα×α n
0 and x ∈ R with
kB − A(x)k < ε. Then A(x) + εIα  0, i.e., x ∈ SA+εIm .
Conversely, suppose that (ii) holds. To show that
dist {A(x) | x ∈ Rn }, SRα×α


0 } = 0,

we let ε > 0 be given and have to find B ∈ SRα×α n

0 and x ∈ R with

kA(x) − Bk ≤ ε.
n
But this is easy: choose x ∈ R with A(x) + εIα  0, and set B := A(x) + εIα .
The following lemma is due to Bohnenblust [Bo48] (see also [Ba01, Theorem 4.2]
for an easier accessible reference). While Bohnenblust gave a non-trivial bound
on the number of terms that are really needed to test condition (i) below, we will
not need this improvement and therefore take the trivial bound α. Then the proof
becomes easy and we include it for the convenience of the reader.
Lemma 3.1.3 (Bohnenblust). For A1 , . . . ,An ∈ SRα×α the following are equiva-
lent:
Pα
(i) Whenever u1 , . . . ,uα ∈ Rα with i=1 u∗i Aj ui = 0 for all j ∈ {1, . . . ,n}, then
u1 = · · · = uα = 0;
6 IGOR KLEP AND MARKUS SCHWEIGHOFER

(ii) span(A1 , . . . ,An ) contains a positive definite matrix.

Proof. It is P trivial that (ii) implies (i). To prove Pα that (i) implies P
(ii), note that
α α
SRα×α
0 = { i=1 u i u ∗
i | u 1 , . . . ,u α ∈ R α
} and u
i=1 i
∗
Bu i = tr(B ∗
i=1 ui ui ) for
α α×α
all u1 , . . . ,uα ∈ R . The hypotheses thus says that, given any B ∈ SR0 , we have
tr(A1 B) = · · · = tr(An B) = 0 =⇒ B = 0. (1)
Now suppose that span(A1 , . . . ,An ) ∩ SRα×α
= ∅. By the standard separation
≻0
theorem for two non-empty disjoint convex sets (see for example [Ba02, Theorem
III.1.2]), span(A1 , . . . ,An ) and SRα×α
≻0 can be separated by a hyperplane (the sep-
arating affine hyperplane must obviously contain the origin). Therefore there is a
non-zero linear functional L : SRα×α → R with
α×α
L(SR≻0 ) ⊆ R≥0 and L(span(A1 , . . . ,An )) ⊆ R≤0 .
α×α
Then of course L(SR0 ) ⊆ R≥0 and L(span(A1 , . . . ,An )) = {0}. Now choose
α×α
B ∈ SR such that
L(A) = tr(AB) for all A ∈ SRα×α .
Then B 6= 0, B ∈ SRα×α
0 and tr(A1 B) = · · · = tr(An B) = 0, contradicting (1).

Lemma 3.1.4. Let A ∈ R[x]α×α

1 be a linear pencil which is either weakly infeasible
or weakly feasible. Then there are k ≥ 1 and u1 , . . . ,uk ∈ Rα r {0} such that
Pk ∗
i=1 ui Aui = 0.

Proof. Assume that the conclusion is false. By Lemma 3.1.3, we find x0 ,x1 , . . . ,xn ∈
R such that
x0 A0 + x1 A1 + · · · + xn An ≻ 0.
Of course it is impossible that x0 > 0 since otherwise A xx10 , . . . , xxn0 ≻ 0. Also

x0 = 0 is excluded (since otherwise A(cx1 , . . . ,cxn ) ≻ 0 for c > 0 large enough).

Hence without loss of generality x0 = −1, i.e., x1 A1 + · · · + xn An ≻ A0 . Choose
ε > 0 such that
x1 A1 + · · · + xn An ≻ A0 + 2εIα .
By Lemma 3.1.2, we can choose some y ∈ SA+εIα . But then
A0 + (x1 + 2y1 )A1 + · · · + (xn + 2yn )An ≻ 2(A0 + εIα + y1 A1 + · · · + yn An )  0,
contradicting the hypotheses.
3.2. An algebraic glimpse at standard SDP duality. We now recall briefly
the standard duality theory of SDP. We present it from the viewpoint of a real
algebraic geometer, i.e., we use the language of polynomials in the formulation of
the primal-dual pair of SDPs and in the proof of strong duality. This is necessary
for a better understanding of the sums of squares dual (Dsos ) given in Subsection
3.7 below.
A semidefinite program (P) and its standard dual (D) is given by a linear pencil
A ∈ R[x]α×α
1 and a linear polynomial ℓ ∈ R[x]1 as follows:

(P) minimize ℓ(x) (D) maximize a

subject to x ∈ Rn subject to S ∈ SRα×α , a ∈ R
A(x)  0 S0
ℓ − a = tr(AS)

To see that this corresponds (up to some minor technicalities) to the formulation
in the literature, just write the polynomial constraint ℓ − a = tr(AS) of the dual as
n + 1 linear equations by comparing coefficients.
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 7

The optimal values of (P) and (D) are defined to be

P ∗ := inf{ℓ(x) | x ∈ Rn , A(x)  0} ∈ R ∪ {±∞} and
∗
D := sup{a | S ∈ SRα×α
0 , a ∈ R, ℓ − a = tr(AS)} ∈ R ∪ {±∞},
respectively, where the infimum and the supremum is taken in the ordered set
{−∞} ∪ R ∪ {∞} (where inf ∅ = ∞ and sup ∅ = −∞). By weak duality, we mean
that P ∗ ≥ D∗ , or equivalently, that the objective value of (P) at any of its feasible
points is greater or equal to the objective value of (D) at any of its feasible points.
Fix a linear pencil A. It is easy to see that weak duality holds for all primal
objectives ℓ if and only if
f ∈ CA =⇒ f ≥ 0 on SA
holds for all f ∈ R[x]1 , which is of course true. By strong duality, we mean that
P ∗ = D∗ (zero duality gap) and that (the objective of) (D) attains this common
optimal value in case it is finite. It is a little exercise to see that strong duality for
all primal objectives ℓ is equivalent to
f ≥ 0 on SA ⇐⇒ f ∈ CA
for all f ∈ R[x]1 .
Unlike weak duality, strong duality fails in general (cf. Subsection 4.6 below;
Pataki recently characterized all linear pencils A such that there exists a linear
objective function ℓ for which strong duality fails [Pat]). However, it is well-known
that it does hold when the feasible set SA of the primal (P) has non-empty interior
(e.g. if A is strongly feasible). Here is a real algebraic geometry flavored proof of
this:
Proposition 3.2.1 (Standard SDP duality). Let A ∈ SR[x]α×α
1 be a linear pencil
such that SA has non-empty interior. Then
f ≥ 0 on SA ⇐⇒ f ∈ CA
for all f ∈ R[x]1 .
Proof. In a preliminary step, we show that the convex cone CA is closed in R[x]1 .
To this end, consider the linear subspace U := {u ∈ Rα | Au = 0} ⊆ Rα . The map
α
X
ϕ : R × (Rα /U )α → CA , (a,ū1 , . . . ,ūα ) 7→ a2 + u∗i Aui
i=1

is well-defined and surjective.

Suppose ϕ maps (a,ū1 , . . . ,ūα ) ∈ (Rα /U )α to 0. Fix i ∈ {1, . . . ,α}. Then
∗
ui A(x)ui = 0 for all x ∈ SA . Since A(x)  0, this implies A(x)ui = 0 for all
x ∈ SA . Using the hypothesis that SA has non-empty interior, we conclude that
Aui = 0, i.e., ui ∈ U . Since i was arbitrary and a = 0, this yields (a,ū1 , . . . ,ūα ) = 0.
This shows ϕ−1 (0) = {0}. Together with the fact that ϕ is a (quadratically)
homogeneous map, this implies that ϕ is proper (see for example [PS01, Lemma
2.7]). In particular, CA = im ϕ is closed.
Suppose now that f ∈ / R[x]1 r CA . The task is to find x ∈ SA such that
f (x) < 0. Being a closed convex cone, CA is the intersection of all closed half-
spaces containing it. Therefore we find a linear map ψ : R[x]1 → R such that
ψ(CA ) ⊆ R≥0 and ψ(f ) < 0. We can assume ψ(1) > 0 since otherwise ψ(1) = 0
and we can replace ψ by ψ + ε evy for some small ε > 0, where y ∈ SA is chosen
arbitrarily. Hereby evx : R[x]1 → R denotes the evaluation in x ∈ Rn . Finally, after
a suitable scaling we can even assume ψ(1) = 1.
8 IGOR KLEP AND MARKUS SCHWEIGHOFER

Now setting x := (ψ(x1 ), . . . ,ψ(xn )) ∈ Rn , we have ψ = evx . So ψ(CA ) ⊆ R≥0

means exactly that A(x)  0, i.e., x ∈ SA . At the same time f (x) = ψ(f ) < 0 as
desired.

3.3. Certificates for low dimensionality of spectrahedra. The problems with

the standard duality theory for SDP thus arise when one deals with spectrahedra
having empty interior. Every convex set with empty interior is contained in an affine
hyperplane. The basic idea is now to code the search for such an affine hyperplane
into the dual SDP and to replace equality in the constraint ℓ − a = tr(AS) of (D) by
congruence modulo the linear polynomial f ∈ R[x]1 defining the affine hyperplane.
However, this raises several issues:
First, SA might have codimension bigger than one in Rn . This will be resolved
by iterating the search up to n times.
Second, we do not see any possibility to encode the search for the linear polyno-
mial f directly into an SDP. What we can implement is the search for a non-zero
quadratic sos-polynomial q together with a certificate of SA ⊆ {q = 0}. Note that
{q = 0} is a proper affine subspace of Rn . It would be best to find a q such that
{q = 0} is the affine hull of SA since then we could actually avoid the n-fold itera-
tion just mentioned. However, as demonstrated in Example 4.6.3 below, this is in
general not possible.
Third, we need to carefully implement congruence modulo linear polynomials f
vanishing on {q = 0}. This will be dealt with by using the radical ideal from real
algebraic geometry together with Schur complements.
We begin with a result which ensures that a suitable quadratic sos-polynomial q
can always be found. In fact, the following proposition says that there exists such a
q which is actually a square. The statement is of interest in itself since it provides
certificates for low-dimensionality of spectrahedra. We need quadratic (i.e., degree
≤ 2) sos-matrices for this.

Proposition 3.3.1. For any linear pencil A ∈ SR[x]α×α

1 , the following are equiv-
alent:
(i) SA has empty interior;
(ii) There exists a non-zero linear polynomial f ∈ R[x]1 and a quadratic sos-matrix
S ∈ SR[x]α×α
2 such that
− f 2 = tr(AS). (2)
Proof. From (ii) it follows that −f 2 ≥ 0 on SA , which implies f = 0 on SA . So it
is trivial that (ii) implies (i).
For the converse, suppose that SA has empty interior. If there is u ∈ Rα r {0}
such that Au = 0 then, by an orthogonal change of coordinates on Rα , we could
assume that u is the first standard basis vector e1 . But then we delete the first
column and the first row from A. We can iterate this and therefore assume from
now on that there is no u ∈ Rα r {0} with Au = 0.
We first treat the easy case where A is stronglyPinfeasible. By Lemma 3.1.1,
∗
there are c ∈ R≥0 and ui ∈ Rα with P −1 − c = i ui Aui . By scaling the ui
∗ α×α
we can assume P c = 0. Setting S := i ui ui ∈ SR and f := 1, we have
−f 2 = −1 = i u∗i Aui = tr(AS) for the constant sos-matrix S and the constant
non-zero linear polynomial f .
Now we assume that A is weakly infeasible or feasible. In case that A is feasible,
it is clearly weakly feasible since otherwise SA would have non-empty interior. Now
Lemma 3.1.4 justifies the following case distinction:
Case 1. There is u ∈ Rα r {0} with u∗ Au = 0.
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 9

Write A = (ℓij )1≤i,j≤α . Again by an orthogonal change of coordinates on Rα , we

can assume that u = e1 , i.e., ℓ11 = 0. Moreover, we may assume f := ℓ12 6= 0 (since
Ae1 = Au 6= 0). Setting f ′ := 12 (−1 − ℓ22 ), v := [f ′ f 0 . . . 0]∗ and S := vv ∗ , we
have
tr(AS) = v ∗ Av = 2f ′ f ℓ12 + f 2 ℓ22 = f 2 (ℓ22 + 2f ′ ) = −f 2 .
Case 2. Case 1 does not apply but there are k ≥ 2 and u1 , . . . ,uk ∈ Rα r {0}
such that ki=1 u∗i Aui = 0.
P
Here we set f := u∗1 Au1 6= 0 and write −f = f12 − f22 where f1 := 21 (−f + 1) ∈
R[x]1 and f2 := 12 (−f − 1) ∈ R[x]1 . Then we can use the quadratic sos-matrix
Pk
S := f12 u1 u∗1 + f22 i=2 ui u∗i to get
 k k
X
 X
tr(AS) = tr A f12 u1 u∗1 + f22 ui u∗i = f12 u∗1 Au1 + f22 u∗i Aui
i=2 i=2
= f12 u∗1 Au1 − f22 u∗1 Au1 = (f12 − f22 )u∗1 Au1 = −f 2 .
The certificate (2) of low-dimensionality exists for some but in general not for
every affine hyperplane containing the spectrahedron. We illustrate this in Ex-
ample 4.6.3 below, where the spectrahedron has codimension two and is therefore
contained in infinitely many affine hyperplanes only one of which allows for a cer-
tificate of the form (2).
3.4. Linear polynomials positive on spectrahedra. We now carry out the
slightly technical but straightforward iteration of Proposition 3.3.1 announced in
Subsection 3.3, and combine it with Proposition 3.2.1. We get a new type of Pos-
itivstellensatz for linear polynomials on spectrahedra with bounded degree com-
plexity. In what follows, we shall use (p1 , . . . ,pr ) to denote the ideal generated by
p1 , . . . , pr .
Theorem 3.4.1 (Positivstellensatz for linear polynomials on spectrahedra).
Let A ∈ SR[x]α×α
1 be a linear pencil and f ∈ R[x]1 . Then
f ≥ 0 on SA
if and only if there exist ℓ1 , . . . ,ℓn ∈ R[x]1 , quadratic sos-matrices S1 , . . . ,Sn ∈
SR[x]α×α
2 , a matrix S ∈ SRα×α
0 and c ≥ 0 such that

ℓ2i + tr(ASi ) ∈ (ℓ1 , . . . ,ℓi−1 ) for i ∈ {1, . . . ,n}, and (3)

f − c − tr(AS) ∈ (ℓ1 , . . . ,ℓn ). (4)
Proof. We first prove that f ≥ 0 on SA in the presence of (3) and (4).
The traces in (3) and (4) are obviously nonnegative on SA . Hence it is clear
that constraint (4) gives f ≥ 0 on SA if we show that ℓi vanishes on SA for all
i ∈ {1, . . . ,n}. Fix i ∈ {1, . . . ,n} and assume by induction that ℓ1 , . . . ,ℓi−1 vanish
on SA . Then (3) implies ℓ2i + tr(ASi ) vanishes on SA and therefore also ℓi .
Conversely, suppose now that f ≥ 0 on SA . We will obtain the data with
properties (3) and (4) by induction on the number of variables n ∈ N0 .
To do the induction basis, suppose first that n = 0. Then f ≥ 0 on SA just means
that the real number f is nonnegative if A ∈ SRα×α is positive semidefinite. But if
f ≥ 0, then it suffices to choose c := f ≥ 0 and S := 0 to obtain (4) with n = 0, and
the condition (3) is empty since n = 0. We now assume that f < 0 and therefore
A 6 0. Then we choose u ∈ Rα with u∗ Au = f . Setting S := uu∗ ∈ SRα×α 0 and
c := 0, we have
f − c − tr(AS) = f − u∗ Au = f − f = 0,
as required.
10 IGOR KLEP AND MARKUS SCHWEIGHOFER

For the induction step, we now suppose that n ∈ N and that we know already
how to find the required data for linear pencils in n − 1 variables. We distinguish
two cases and will use the induction hypothesis only in the second one.
Case 1. SA contains an interior point.
In this case, we set all ℓi and Si to zero so that (3) is trivially satisfied. Property
(4) can be fulfilled by Proposition 3.2.1.
Case 2. The interior of SA is empty.
In this case, we apply Proposition 3.3.1 to obtain 0 6= ℓ1 ∈ R[x]1 and a quadratic
sos-matrix S1 ∈ SR[x]α×α with
ℓ21 + tr(AS1 ) = 0. (5)
The case where ℓ1 is constant is trivial. In fact, in this case we can choose all
remaining data being zero since (ℓ1 , . . . ,ℓi ) = (ℓ1 ) = R[x] for all i ∈ {1, . . . ,n}.
From now on we therefore assume ℓ1 to be non-constant. But then the reader
easily checks that there is no harm carrying out an affine linear variable trans-
formation which allows us to assume ℓ1 = xn . We then apply the induction hy-
pothesis to the linear pencil A′ := A(x1 , . . . ,xn−1 ,0) and the linear polynomial
f ′ := f (x1 , . . . ,xn−1 ,0) in n − 1 variables to obtain ℓ2 , . . . ,ℓn ∈ R[x]1 , quadratic
sos-matrices S2 , . . . ,Sn ∈ SR[x]α×α2 , a matrix S ∈ SRα×α 0 and a constant c ≥ 0
such that
ℓ2i + tr(A′ Si ) ∈ (ℓ2 , . . . ,ℓi−1 ) for i ∈ {2, . . . ,n} and (6)
′ ′
f − c − tr(A S) ∈ (ℓ2 , . . . ,ℓn ). (7)
′ ′ ′
Noting that both f − f and tr(ASi ) − tr(A Si ) = tr((A − A )Si ) are contained in
the ideal (xn ) = (ℓ1 ), we see that (6) together with (5) implies (3). In the same
manner, (7) yields (4).
3.5. Constructing SDPs for sums of squares problems. The (coefficient tu-
ples of) sos-polynomials in R[x] of bounded degree form a projection of a spec-
trahedron. In other words, the condition of being (the coefficient tuple of) an
sos-polynomial in R[x] of bounded degree can be expressed with an LMI by means
of additional variables. This is the well-known Gram matrix method [Lau09, Ma08].
As noted by Kojima [Ko03] and nicely described by Hol and Scherer [SH06], the
Gram matrix method extends easily to sos-matrices (see Example (??) below).
3.6. Real radical computations. Let A ∈ SR[x]α×α 1 be a linear pencil and q ∈
R[x]2 a (quadratic) sos-polynomial such that −q = tr(AS) for some (quadratic)
sos-matrix S like in (2) above. In order to resolve the third issue mentioned in
Subsection 3.3, we would like to get our hands on (cubic) polynomials vanishing on
{q = 0}. That is, we want to implement the ideals appearing in (3) and (4) in an
SDP.
By the Real Nullstellensatz [BCR98, Ma08,
p PD01], each polynomial vanishing
on the real zero set {q = 0} of q lies in R (q). This gives us a strategy of how to
find the cubic polynomials vanishing on {q = 0}, cf. Proposition 3.6.1 and Lemma
3.6.2 below. The Real Nullstellensatz playsp only a motivating role for us; we only
use its trivial converse: Each element of R (q) vanishes on {q = 0}.
The central question is how to model the search for elements in the real radical
ideal using SDP. The key to this will be to represent polynomials by matrices as is
done in the Gram matrix method mentioned in Section 3.5. For this we introduce
some notation.
For each d ∈ N0 , let s(d) := dim R[x]d = d+n


n denote the number of monomials
#„ s(d)
of degree at most d in n variables and [x]d ∈ R[x] the column vector
#„ 2
∗
[x]d := 1 x1 x2 . . . xn x1 x1 x2 . . . . . . xdn

 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 11

consisting of these monomials ordered first with respect to the degree and then
lexicographic.
The following proposition shows how to find elements
p of degree at most d + e
(represented by a matrix W ) in the real radical I := R (q) of the ideal generated by
#„ ∗ #„
a polynomial q ∈ R[x]2d (represented by a symmetric matrix U , i.e., q = [x]d U [x]d ).
We will later use it with d = 1 and e = 2 since q will be quadratic and we will be
interested in cubic polynomials in I. Note that
 
∗ I W
U  W W ⇐⇒ 0
W∗ U

by the method of Schur complements.

Proposition 3.6.1. Let d,e ∈ N0 , let I be a real radical ideal of R[x] and U ∈
#„ ∗ #„
SRs(d)×s(d) be such that [x]d U [x]d ∈ I. Suppose W ∈ Rs(e)×s(d) with U  W ∗ W .
#„ ∗ #„
Then [x]e W [x]d ∈ I.

Proof. Since U − W ∗ W is positive semidefinite, we find B ∈ Rs(d)×s(d) with U −

#„
W ∗ W = B ∗ B. Now let pi ∈ R[x] denote the i-th entry of W [x]d and qj the j-th
#„
entry of B [x]d . From
#„ #„ #„ #„
p21 + · · · + p2s(e) + q12 + · · · + qs(d)
2
= (W [x]d )∗ W [x]d + (B [x]d )∗ B [x]d
#„ ∗ #„ #„ ∗ #„
= [x]d (W ∗ W + B ∗ B)[x]d = [x]d U [x]d ∈ I

it follows that p1 , . . . ,ps(e) ∈ I since I is real radical. Now

#„ ∗ #„ #„ ∗ #„
[x]e W [x]d = [x]e [p1 . . . ps(e) ]∗ = [p1 . . . ps(e) ][x]e ∈ I

since I is an ideal.

The following lemma is a weak converse to Proposition 3.6.1. Its proof relies
heavily on the fact that only linear and quadratic polynomials are involved.

Lemma 3.6.2. Let ℓ1 , . . . ,ℓt ∈ R[x]1 , and q1 , . . . ,qt ∈ R[x]2 . Suppose that U ∈
SRs(1)×s(1) satisfies
#„ ∗ #„
[x]1 U [x]1 = ℓ21 + · · · + ℓ2t . (8)
Then there exists λ > 0 and W ∈ Rs(2)×s(1) satisfying λU  W ∗ W and
#„ ∗ #„
[x]2 W [x]1 = ℓ1 q1 + · · · + ℓt qt .

Proof. Note that the U satisfying (8) is unique and hence positive semidefinite.
Suppose that at least one qi 6= 0 (otherwise take W = 0). Choose column vectors
#„
ci ∈ Rs(2) such that c∗i [x]2 = qi . Now let W ∈ Rs(2)×s(1) be the matrix defined by
#„ Pt
W [x]1 = i=1 ℓi ci , so that
t t t
#„ ∗ #„ X #„ ∗ X #„ X
[x]2 W [x]1 = ℓi [x]2 ci = ℓi c∗i [x]2 = ℓi qi .
i=1 i=1 i=1

Moreover, we get
t
#„ ∗ #„ #„ #„ X
[x]1 W ∗ W [x]1 = (W [x]1 )∗ W [x]1 = (ℓi ci )∗ (ℓj cj ),
i,j=1
12 IGOR KLEP AND MARKUS SCHWEIGHOFER

and therefore for all x ∈ Rn ,

 
1
x t
 ∗  1 X
(ℓi (x)ci )∗ (ℓj (x)cj )

1 x1 . . . xn W W  .  =

 ..  i,j=1
xn
t
1 X
(ℓi (x)ci )∗ (ℓi (x)ci ) + (ℓj (x)cj )∗ (ℓj (x)cj )


≤
2 i,j=1
t
X t
X
=t (ℓi (x)ci )∗ (ℓi (x)ci ) ≤ λ ℓi (x)2 ,
i=1 i=1
Pt ∗
where we set λ := t i=1 ci ci > 0. Therefore
∗
xn (λU − W ∗ W ) 1 x1
  
1 x1 ... ... xn ≥0
for all x ∈ R . By homogeneity and continuity this implies y (λU − W ∗ W )y ≥ 0
n ∗

for all y ∈ Rs(1) , i.e., λU  W ∗ W .

3.7. A new exact duality theory for SDP. Given an SDP of the form (P)
described in Subsection 3.2, the following is what we call its sums of squares dual :
(Dsos ) maximize a
subject to S ∈ SRα×α 0 , a ∈ R
S1 , . . . ,Sn ∈ SR[x]α×α
2 quadratic sos-matrices
U1 , . . . ,Un ∈ SRs(1)×s(1)
W1 , . . . ,Wn ∈ Rs(2)×s(1)
#„ ∗ #„ #„ ∗ #„
[x]1 Ui [x]1 + [x]2 Wi−1 [x]1 + tr(ASi ) = 0 (i ∈ {1, . . . ,n})
Ui  Wi∗ Wi (i ∈ {1, . . . ,n})
#„ ∗ #„
ℓ − a + [x]2 Wn [x]1 − tr(AS) = 0,
where W0 := 0 ∈ Rs(2)×s(1) .
Remark 3.7.1. Just like Ramana’s extended Lagrange-Slater dual [Ra97], (Dsos )
can be written down in polynomial time (and hence has polynomial size) in the
bit size of the primal (assuming the latter has rational coefficients) and it guaran-
tees that strong duality (i.e., weak duality, zero gap and dual attainment) always
holds. Similarly, the facial reduction [BW81, TW] gives rise to a good duality the-
ory of SDP. We refer the reader to [Pat00] for a unified treatment of these two
constructions.
As mentioned in Section 3.5, the quadratic sos-matrices can easily be modeled by
SDP constraints using the Gram matrix method, and the polynomial identities can
be written as linear equations by comparing coefficients. The Si serve to produce
negated quadratic sos-polynomials vanishing on SA (cf. Proposition 3.3.1) which
are captured by the matrices Ui . From this, cubics vanishing on SA are produced
(cf. Subsection 3.6) and represented by the matrices Wi . These cubics serve to
implement the congruence modulo the ideals from (3) and (4). Then the entire
procedure is iterated n times. We present an explicit example in Section 4.6.
Just as Proposition 3.2.1 corresponds to the standard SDP duality, Theorem
3.7.3 below translates into the strong duality for the sums of squares dual (Dsos ).
Before we come to it, we need a folk lemma, well-known from the theory of Gröbner
bases.
Lemma 3.7.2. Suppose d ∈ N, f ∈ R[x]d and ℓ1 , . . . ,ℓt ∈ R[x]1 are linear polyno-
mials such that f ∈ (ℓ1 , . . . ,ℓt ). Then at least one of the following is true:
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 13

(a) there exist p1 , . . . ,pt ∈ R[x]d−1 such that f = p1 ℓ1 + · · · + pt ℓt ;

(b) there are λ1 , . . . ,λt ∈ R such that λ1 ℓ1 + · · · + λt ℓt = 1.
Proof. Suppose that (b) is not fulfilled. Then we may assume by Gaussian elimina-
tion and after renumbering the variables that ℓi = xi −ℓ′i where ℓ′i ∈ R[xi+1 , . . . ,xn ]1 .
We now proceed by induction on t ∈ N0 to prove (a). For t = 0, there is nothing to
show. Now let t ∈ PN and suppose the lemma is already proved with t replaced by
t − 1. Write f = |α|≤d aα xα with aα ∈ R. Setting g := f (ℓ′1 ,x2 , . . . ,xn ), we have
X
′α1 α2 ′
f −g = aα (xα αn
1 − ℓ1 )x2 · · · xn = p1 (x1 − ℓ1 ) = p1 ℓ1 ,
1

|α|≤d
1≤α1

where
1 −1
αX
!
X
p1 := aα xi1 ℓ′α
1
1 −1−i
xα αn
2 · · · xn ∈ R[x]d−1 .
2

|α|≤d i=0
1≤α1
Moreover, g ∈ (ℓ2 , . . . ,ℓt ) and therefore g = p2 ℓ2 + · · · + pt ℓt for some p2 , . . . ,pt ∈
R[x]d−1 by the induction hypothesis. Now
f = (f − g) + g = p1 ℓ1 + · · · + pt ℓt .
Theorem 3.7.3 (Sums of squares SDP duality). Let A ∈ SR[x]α×α 1 be a linear
pencil and f ∈ R[x]1 . Then
f ≥ 0 on SA
if and only if there exist quadratic sos-matrices S1 , . . . ,Sn ∈ SR[x]α×α 2 , matrices
U1 , . . . ,Un ∈ SRs(1)×s(1) , W1 , . . . ,Wn ∈ Rs(2)×s(1) , S ∈ SRα×α
0 and c ∈ R≥0 such
that
#„ ∗ #„ #„ ∗ #„
[x]1 Ui [x]1 + [x]2 Wi−1 [x]1 + tr(ASi ) = 0 (i ∈ {1, . . . ,n}), (9)
Ui  Wi∗ Wi (i ∈ {1, . . . ,n}), (10)
#„ ∗ #„
f − c + [x]2 Wn [x]1 − tr(AS) = 0, (11)
where W0 := 0 ∈ Rs(2)×s(1) .
Proof. We first prove that existence of the above data implies f ≥ 0 on SA . All
we will use about the traces appearing in (9) and (11) is that they are polynomials
nonnegative on SA . Let I denote the real radical ideal of all polynomials vanishing
#„ ∗ #„
on SA . It is clear that (11) gives f ≥ 0 on SA if we show that [x]2 Wn [x]1 ∈ I. In
#„ ∗ #„
fact, we prove by induction that [x]2 Wi [x]1 ∈ I for all i ∈ {0, . . . ,n}.
The case i = 0 is trivial since W0 = 0 by definition. Let i ∈ {1, . . . ,n} be given
#„ ∗ #„ #„ ∗ #„
and suppose that [x]2 Wi−1 [x]1 ∈ I. Then (9) shows [x]1 Ui [x]1 ≤ 0 on SA . On the
#„ ∗ #„
other hand, (10) implies in particular Ui  0 and therefore [x]1 Ui [x]1 ≥ 0 on SA .
#„ ∗ #„ #„ ∗ #„
Combining both, [x]1 Ui [x]1 ∈ I. Now Proposition 3.6.1 implies [x]2 Wi [x]1 ∈ I by
(10). This ends the induction and shows f ≥ 0 on SA as claimed.
Conversely, suppose now that f ≥ 0 on SA . By Theorem 3.4.1 and Lemma 3.7.2,
we can choose ℓ1 , . . . ,ℓn ∈ R[x]1 , quadratic sos-matrices S1′ , . . . ,Sn′ ∈ SR[x]α×α ,
S ∈ SRα×α
0 and qij ∈ R[x]2 (1 ≤ j ≤ i ≤ n) such that
i−1
X
ℓ21 + · · · + ℓ2i + tr(ASi′ ) = q(i−1)j ℓj (i ∈ {1, . . . ,n}) and (12)
j=1
Xn
f − c − tr(AS) = qnj ℓj . (13)
j=1
14 IGOR KLEP AND MARKUS SCHWEIGHOFER

There are two little arguments involved in this: First, (3) can trivially be rewritten
as ℓ21 + · · · + ℓ2i + tr(ASi′ ) ∈ (ℓ1 , . . . ,ℓi−1 ) for i ∈ {1, . . . ,n}. Second, in Lemma 3.7.2
applied to ℓ1 , . . . ,ℓi−1 (i ∈ {1, . . . ,n + 1}) we might fall into case (b). But then we
may set ℓi = · · · = ℓn = 0 and Si′ = · · · = Sn′ = S = 0.
#„ ∗ #„
Now define Ui′ ∈ SRs(1)×s(1) by [x]1 Ui′ [x]1 = ℓ21 + · · · + ℓ2i for i ∈ {1, . . . ,n}.
Using Lemma 3.6.2, we can then choose λ > 0 and W1′ , . . . ,Wn′ ∈ Rs(2)×s(1) such
that
λUi′  Wi′∗ Wi′ (14)
#„ ∗ ′ #„ Pi
and [x]2 Wi [x]1 = − j=1 qij ℓj for i ∈ {1, . . . ,n}. Setting Wn := Wn′ , equation
(13) becomes (11). Moreover, (12) can be rewritten as
#„ ∗ #„ #„ ∗ ′ #„
[x]1 Ui′ [x]1 + [x]2 Wi−1 [x]1 + tr(ASi′ ) = 0 (i ∈ {1, . . . ,n}) (15)
To cope with the problem that λ might be larger than 1 in (14), we look for
λ1 , . . . ,λn ∈ R>0 such that Ui  Wi∗ Wi for all i ∈ {1, . . . ,n} if we define Ui := λi Ui′
′
and Wi−1 := λi Wi−1 for all i ∈ {1, . . . ,n} (in particular W0 = W0′ = 0). With
this choice, the desired linear matrix inequality (10) is now equivalent to λi Ui′ 
λ2i+1 Wi′∗ Wi′ for i ∈ {1, . . . ,n − 1} and λn Un′  Wn′2 . Looking at (14), we therefore
see that any choice of the λi satisfying λi ≥ λλ2i+1 for i ∈ {1, . . . ,n − 1} and λn ≥ λ
ensures (10). Such a choice is clearly possible. Finally, equation (15) multiplied by
λi yields (9) by setting Si := λi Si′ for i ∈ {1, . . . ,n}.

4. Positivity of polynomials on spectrahedra

In this section we present applications of the results presented in Section 3. We
interpret Theorems 3.4.1 and 3.7.3 in the language of real algebraic geometry in
Subsection 4.2, and prove a nonlinear Farkas’ lemma for SDP, i.e., nonlinear alge-
braic certificates for infeasibility of an LMI, in Subsection 4.3. These results use
quadratic modules from real algebraic geometry, which we recall in Subsection 4.1.
As a side product we obtain a hierarchy for infeasibility of LMIs, whose first stage
coincides with strong infeasibility. Subsection 4.4 contains certificates for bound-
edness of spectrahedra and a Putinar-Schmüdgen-like Positivstellensatz for poly-
nomials positive on bounded spectrahedra. Finally, the section concludes with two
brief subsections containing examples illustrating our results, and an application to
positive linear functionals.
4.1. Quadratic module associated to a linear pencil. Let R be a (commuta-
tive unital) ring. A subset M ⊆ R is called a quadratic module in R if it contains
1 and is closed under addition and multiplication with squares, i.e.,
1 ∈ M, M + M ⊆ M, and a2 M ⊆ M for all a ∈ R,
see for example [Ma08]. The support of M is defined to be supp M := M ∩ (−M ).
A quadratic module M ⊆ R is called proper if −1 6∈ M . If 21 ∈ R, then the identity
4a = (a + 1)2 − (a − 1)2 for all a ∈ R, (16)
shows that supp M is an ideal and therefore any improper quadratic module M
equals R (since 1 is contained in its support).
An LMI A(x)  0 can be seen as the infinite family of simultaneous linear
inequalities u∗ A(x)u ≥ 0 (u ∈ Rα ). In optimization, when dealing with families of
linear inequalities, one often considers the convex cone generated by them (cf. CA in
Subsection 3.1). Real algebraic geometry handles arbitrary polynomial inequalities
and uses the multiplicative structure of the polynomial ring. Thence one considers
more special types of convex cones, like quadratic modules. One of the aims of
this section is to show that it is advantageous to consider quadratic modules for the
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 15

study of LMIs. Since quadratic modules in polynomial rings are infinite-dimensional

convex cones, we will later on also consider certain finite-dimensional truncations
of them, see Subsection 4.3.
Definition 4.1.1. Let A be a linear pencil of size α in the variables x. We introduce
n X X o
MA := s + vi∗ Avi | s ∈ R[x]2 , vi ∈ R[x]α
i
n o
= s + tr(AS) | s ∈ R[x] sos-polynomial, S ∈ R[x]α×α sos-matrix ⊆ R[x],
and call it the quadratic module associated to the linear pencil A.
Note that for any linear pencil A, each element of MA is a polynomial nonneg-
ative on the spectrahedron SA . In general, MA does not contain all polynomials
nonnegative on SA (e.g. when A is diagonal and dim SA ≥ 3 [Sc09]; another sim-
ple example is presented in Example 4.2.5 below). For diagonal A, the quadratic
module MA (and actually the convex cone CA ) contains however all linear poly-
nomials nonnegative on the polyhedron SA by Farkas’ lemma. For non-diagonal
linear pencils A even this can fail, see Example 4.2.5 below. To certify nonnegativ-
ity of a linear polynomial on a spectrahedron, we therefore employ more involved
algebraic certificates (motivated by the real radical), cf. Theorem 3.4.1 and its
SDP-implementable version Theorem 3.7.3. These two theorems yield algebraic
certificates for linear polynomials nonnegative on SA . While they have the advan-
tage of being very well-behaved with respect to complexity issues, their statements
are somewhat technical. Leaving complexity issues aside, one can use them to de-
duce a cleaner algebraic characterization of linear polynomials nonnegative on SA .
Indeed, given f ∈ R[x]1 with f ≥ 0 on SA , the certificates
√ in Theorems 3.4.1 and
3.7.3 can be seen to be equivalent to f ∈ CA + supp MA√by means of Prestel’s
theory of semiorderings. Note that each √ element of CA + supp MA is obviously
nonnegative on SA since the elements of supp MA vanish on SA .
Finally, this will allow us to come back to the quadratic module MA . We will
show that it contains each linear polynomial nonnegative on SA after adding an
arbitrarily small positive constant, see Corollary 4.2.4.
In this subsection, basic familiarity with real algebraic geometry as presented
e.g. in [BCR98, Ma08, PD01] is needed. The following proposition follows easily
from Prestel’s theory of semiorderings on a commutative ring, see for example [Sc09,
1.4.6.1].
Proposition 4.1.2. Let M be a quadratic module in R[x]. Then
p p \
supp M = R supp M = {supp S | S semiordering of R[x], M ⊆ S}.
We explicitly extract the following consequence since this is exactly what is
needed in the sequel.
Lemma 4.1.3. Let M be a quadratic module in R[x]. Then
p p
( supp M − M ) ∩ M ⊆ supp M . (17)
√
Proof. To prove (17), suppose p ∈ M can be written p = g − q with g ∈ supp M
and q ∈ M . By Proposition 4.1.2, we have to show that p ∈ supp S for each
semiordering S of R[x] with M ⊆ S. But if such S is given, then g ∈ supp S and
therefore p = g − q ∈ −S as well as p ∈ M ⊆ S. Hence p ∈ supp S.
4.2. Linear polynomials positive on spectrahedra – revisited. With Lemma
4.1.3 at hand, we can now give a conceptual interpretation of the certificates ap-
pearing in Theorem 3.4.1, disregarding the complexity of the certificate.
16 IGOR KLEP AND MARKUS SCHWEIGHOFER

Proposition 4.2.1. If A ∈ SR[x]1α×α is a linear pencil, f,ℓ1 , . . . ,ℓn ∈ R[x]1 are

linear polynomials, S1 , . . . ,Sn ∈ SR[x]α×α
2 are quadratic sos-matrices, S ∈ SRα×α
0
√
and c ∈ R≥0 are such that (3) and (4) hold, then f ∈ CA + supp MA .
√
Proof. Set I := supp MA . It is clear that (4) gives f ∈ CA + I if we prove that
ℓi ∈ I for all i ∈ {1, . . . ,n}. Fix i ∈ {1, . . . ,n} and assume by induction that
ℓ1 , . . . ,ℓi−1 ∈ I. Then (3) implies ℓ2i + tr(ASi ) ∈ I and therefore ℓ2i ∈ (I − MA ) ∩
R[x]2 ⊆ (I − MA ) ∩ MA ⊆ I by (17).
P

We get the same interpretation for the certificates from Theorem 3.7.3.
Proposition 4.2.2. If A ∈ SR[x]α×α 1 is a linear pencil, f ∈ R[x]1 is a linear polyno-
mial, S1 , . . . ,Sn ∈ SR[x]2α×α are quadratic sos-matrices, U1 , . . . ,Un ∈ SRs(1)×s(1) ,
α×α
W1 , . . . ,Wn ∈ Rs(2)×s(1) , S ∈ SR0 and c ∈ R≥0 are such that (9), (10) and (11)
√
hold, then f ∈ CA + supp MA .
√
Proof. Set I := supp MA . It is clear that constraint (11) gives f ∈ CA + I if we
#„ #„ #„ #„
show that [x]2 Wn [x]1 ∈ I. In fact, we show by induction that [x]2 Wi [x]1 ∈ I for all
i ∈ {0, . . . ,n}.
The case i = 0 is trivial since W0 = 0 by definition. Let i ∈ {1, . . . ,n} be
#„ ∗ #„
given and suppose that we know already [x]2 Wi−1 [x]1 ∈ I. Then (9) shows
#„ ∗ #„
[x]1 Ui [x]1 ∈ I − MA . On the other hand (10) implies in particular Ui  0 and
#„ ∗ #„ #„ ∗ #„
R[x]2 ⊆ MA . But then [x]1 Ui [x]1 ∈ (I − MA ) ∩ MA ⊆ I
P
therefore [x]1 Ui [x]1 ∈
#„ ∗ #„
by (17). Now (10) yields [x]2 Wi [x]1 ∈ I by Proposition 3.6.1 since I is real radical
by Proposition 4.1.2. This ends the induction.
The following corollary is now a generalization of Proposition 3.2.1 working also
for low-dimensional SA (note that supp MA = (0) if SA has non-empty interior).
Corollary 4.2.3. Let A ∈ SR[x]α×α
1 be a linear pencil. Then
p
f ≥ 0 on SA ⇐⇒ f ∈ CA + supp MA
for all f ∈ R[x]1 .
Proof. Combine either Theorem 3.4.1 with Proposition 4.2.1, or Theorem 3.7.3 with
Proposition 4.2.2.
Corollary 4.2.4. Let A ∈ SR[x]α×α
1 be a linear pencil. Then
f ≥ 0 on SA ⇐⇒ ∀ε > 0 : f + ε ∈ MA
for all f ∈ R[x]1 .
Proof. To prove the non-trivial implication, let f ∈ R[x]1 with f ≥ 0 on SA be given.
It suffices to show f + ε ∈ MA for the special case ε = 1 (otherwise replace f by εf
and divide by ε). By Corollary 4.2.3, there exists g ∈ CA , p ∈ R[x] and k ∈ N such
that f = g+p and pk ∈ I := supp MA . Now f +ε = f +1 = g+(f −g)+1 = g+(p+1)
and it is enough to show that p + 1 ∈ MA . This will follow from the fact that the
image of p + 1 is a square in the quotient ring R[x]/I. Indeed, since the image of p
in R[x]/I is nilpotent (in fact the image of pk is zero), we can simply write down
a square root of this element using the finite Taylor expansion at 1 of the square
root function in 1 given by the binomial series:
k−1
!2
X 1
2 i
p+1≡ p mod I.
i=0
i

We point out that adding an ε > 0 in Corollary 4.2.4 cannot be avoided:

 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 17

Example 4.2.5. Consider  

1 x
A= .
x 0
Then SA = {0}. Hence obviously x ≥ 0 on SA . But it is easy to see that x 6∈ MA
[Za, Example 2].
4.3. Infeasibility of linear pencils. Let A be a linear pencil. In Subsection
3.1 we have introduced the convex cone CA ⊆ R[x]1 and above we have defined
the quadratic module MA ⊆ R[x] associated to A, both consisting of polynomials
which are obviously nonnegative on the spectrahedron SA . The convex cone CA
is in general too small to detect infeasibility of A (in particular, strong duality for
the standard primal-dual pair of an SDP fails; see also Subsection 4.6 below). On
the other hand, as an easy consequence of Corollary 4.2.4, we have the following
version of Farkas’ lemma for SDP which unlike Lemma 3.1.1 does not only work for
strongly but also for weakly infeasible linear pencils. The price we pay is that we
have to replace the finite-dimensional convex cone CA by the infinite-dimensional
quadratic module MA .
Corollary 4.3.1 (Nonlinear Farkas’ lemma for semidefinite programming). Let A
be a linear pencil. Then
SA = ∅ ⇐⇒ −1 ∈ MA .
In other words, A is infeasible if and only if MA is improper.
To pass from the infinite-dimensional MA to finite-dimensions, we introduce
truncations of MA :
Definition 4.3.2. Given a linear pencil A of size α and k ∈ N0 , let
nX o
(k)
X
MA := p2i + vj∗ Avj | pi ∈ R[x]k , vj ∈ R[x]α
k
i j
n o
= s + tr(AS) | s ∈ R[x]2k sos-polynomial, S ∈ R[x]α×α
2k sos-matrix
⊆ R[x]2k+1 ,
be the truncated quadratic module with degree restriction k associated to A. Note
(0)
that MA = CA .
Our main result in this subsection is the following quantitative strengthening of
Corollary 4.3.1:
Theorem 4.3.3. Let A be an infeasible linear pencil of size α in n variables. Then
(2min{α−1,n} −1)
−1 ∈ MA .
Observe that if A is strictly infeasible, then this statement follows by Lemma
3.1.1. We present a self-contained proof of Theorem 4.3.3 based on the following
two lemmas.
Lemma 4.3.4. Let A be an infeasible linear pencil in n variables. Then
(2n −1)
−1 ∈ MA .
Proof. We shall prove this by induction on n. The statement is clear for n = 0.
Given n ∈ N and
Xn
A = A0 + xi Ai ∈ SR[x]α×α
1 ,
i=1
we assume the statement has been established for all infeasible linear pencils with
n − 1 variables.
18 IGOR KLEP AND MARKUS SCHWEIGHOFER

2
By Lemma 3.1.4, there is an u ∈ Rα r {0} with u∗ (Iα ⊗ A)u = 0. Replacing A
(k)
by Iα ⊗ A changes neither MA nor SA = ∅. Without loss of generality, we assume
therefore that there is u ∈ Rα r {0} with u∗ Au = 0. Writing A = (ℓij )1≤i,j≤α and
performing a linear coordinate change on Rα , we can moreover assume ℓ11 = 0.
Furthermore, without loss of generality, ℓ12 6= 0. Setting ℓ′ := 12 (−1 − ℓ22 ), v :=
[ℓ′ ℓ12 0 . . . 0]∗ and S := vv ∗ , we have
(1)
tr(AS) = v ∗ Av = 2ℓ′ ℓ212 + ℓ212 ℓ22 = ℓ212 (ℓ22 + 2ℓ′ ) = −ℓ212 ∈ MA . (18)
If ℓ12 ∈ R, we are done. Otherwise after possibly performing an affine linear change
of variables on Rn , we may assume ℓ12 = xn .
Now A′ := A(x1 , . . . ,xn−1 ,0) is an infeasible linear pencil in n − 1 variables. By
(2n−1 −1)
our induction hypothesis, −1 ∈ MA′ . In particular, there are
pi ∈ R[x]2n−1 −1 and vj ∈ R[x]α
2n−1 −1

satisfying X X
−1 = p2i + vj∗ A′ vj .
i j
vj∗ An vj ∈ R[x]2n −2 . Then
P
Let q := j
X X
−1 = 2 p2i + 2 vj∗ A′ vj + 1
i j
X X
=2 p2i +2 vj∗ Avj − 2qxn + 1 (19)
i j
X X
=2 p2i + 2 vj∗ Avj + (1 − qxn )2 + q 2 (−x2n ).
i j
(2n −1)
Since deg q ≤ 2n − 2, we have q 2 (−xn ) ∈ MA by (18). Taken together with
2 (2n −1) (2n −1)
(1 − qxn ) ∈ MA , (19) implies −1 ∈ MA .
Lemma 4.3.5. Let A be an infeasible linear pencil of size α. Then
(2α−1 −1)
−1 ∈ MA .
Proof. We prove this by induction on α. The statement is clear for α = 1. Given
n
X
A = A0 + xi Ai ∈ SR[x]α×α
1
i=1

of size α ≥ 2, we assume the statement has been established for all infeasible linear
(0)
pencils of size α − 1. If A is strongly infeasible, then −1 ∈ CA = MA by Lemma
3.1.1. So we may assume A is weakly infeasible.
Claim. There is an affine linear change of variables after which A assumes the
form
b∗
 
b
A= 0 ,
b A′
∗
where b0 ∈ R[x]1 , b = b1 · · · bα−1 ∈ R[x]α−1 , A′ is a linear pencil of size α−1,

1
and bj ∈ R[x]1 satisfy
(1)
− b2j ∈ MA for j = 0, . . . ,α − 1. (20)
Furthermore, b0 can be chosen to be either 0 or x1 .
Explanation. By Lemma 3.1.4, there is k ∈ N and u1 , . . . ,uk ∈ Rα r {0} with
Pk ∗
i=1 ui Aui = 0. We distinguish two cases.
Case 1. There is u ∈ Rα r {0} with u∗ Au = 0.
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 19

Write A = (ℓij )1≤i,j≤α . By changing coordinates on Rα , we can assume that u is

the first standard basis vector, i.e., ℓ11 = 0. Hence
0 b∗
 
A= ,
b A′
∗
where b = b1 · · · bα−1 ∈ R[x]α−1 and A′ is a linear pencil of size α − 1. As in

1
(1)
the proof of Lemma 4.3.4, we deduce that −b2j ∈ MA for all j = 1, . . . ,α − 1.
Case 2. Case 1 does not apply but there are k ≥ 2 and u1 , . . . ,uk ∈ Rα r {0}
Pk
such that i=1 u∗i Aui = 0.
In this case,
k
(0) (0)
X
ℓ11 := u∗1 Au1 = − u∗i Aui ∈ CA ∩ −CA = MA ∩ −MA .
i=2

Since Case 1 does not apply, ℓ11 6= 0. Furthermore, since A is assumed to be weakly
infeasible, ℓ11 6∈ R. Hence after an affine linear change of variables on Rn , we can
assume ℓ11 = x1 . Thus
b∗
 
x
A= 1 ,
b A′
∗
where b = b1 · · · bα−1 ∈ R[x]α−1 and A′ is a linear pencil of size α − 1. Note

1
that
−4x21 = (1 − x1 )2 x1 + (1 + x1 )2 (−x1 )
(1)
shows that −x21 ∈ ML . Using this, one gets similarly as above that also each of
(1)
the entries bj of b satisfies −b2j ∈ MA . This proves our claim. 
If one of the bj ∈ R r {0}, we are done by (20). Otherwise we consider two cases.
Case a. The linear system b0 (x) = 0, b(x) = 0 is infeasible.
Then we proceed as follows. There are γ0 , . . . ,γα−1 ∈ R satisfying
α−1
X
γj bj = 1. (21)
j=0

For each j = 0, . . . ,α − 1 and δ ∈ R we have

 δ 2 δ 2 (1)
1 + δbj = 1 + bj + (−b2j ) ∈ MA
2 4
by (20). Hence (21) implies
α−1 α−1
X X 1 
(1)
−1 = 1 − 2 = 1 − 2 γj b j = − 2γj bj ∈ ML .
j=0 j=0
α

Case b. The linear system b0 (x) = b(x) = 0 is feasible.

Then we perform an affine linear change of variables on Rn to ensure
{x ∈ Rn | b0 (x) = 0, b(x) = 0} = {0}r × Rn−r
for some r ∈ N. Moreover, we may assume x1 , . . . ,xr are among the entries bj ,
j = 0, . . . ,α − 1.
Now A′′ := A′ (0, . . . ,0,xr+1 , . . . , xn ) is an infeasible linear pencil of size α−1. By
(2α−2 −1)
our induction hypothesis, −1 ∈ MA′′ . In particular, there are s ∈ Σ2 ∩ R[x]
α−1 α−1
with deg s ≤ 2 − 2, and vi ∈ R[x]2α−2 −1 satisfying
X
−1 = s + vi∗ A′′ vi .
i
20 IGOR KLEP AND MARKUS SCHWEIGHOFER

Introducing
 
X 0
qt := vi∗ At vi ∈ R[x]2α−1 −2 and wi := ∈ R[x]α
2α−2 −1
vi
i

we have
X
vi∗ A′′ vi + 1


−1 = 2s + 2
i
r r 
X X
X 1 √
2 
= 2s + 2 vi∗ A′ vi − 2qt xt + √ − rqt xt + 2qt xt + rqt2 (−x2t )
i t=1 t=1
r
r r
X X 1 √
2 X
= 2s + 2 wi∗ Awi + √ − rqt xt + rqt2 (−x2t ).
i t=1
r t=1
(22)
(2α−1 −1) √ (2α−1 −1)
Combining qt2 (−x2t ) ∈ MA with ( √1r − rqt xt )2 ∈ MA , (22) implies
(2α−1 −1)
−1 ∈ MA .

Proof of Theorem 4.3.3. Immediate from Lemma 4.3.4 and Lemma 4.3.5.

Remark 4.3.6. With the aid of truncated quadratic modules associated to a linear
pencil A, we can introduce a hierarchy of infeasibility: A is called k-infeasible for
(k) (k−1)
k ∈ N0 , if −1 ∈ MA and −1 6∈ MA . By Lemma 3.1.1, A is strongly infeasible
if and only if it is 0-infeasible, and A is weakly infeasible if and only if it is k-
infeasible for some k ∈ N. Detecting k-infeasibility can be implemented as an SDP,
cf. Subsection 3.5.
In [HL06] Henrion and Lasserre extend Lasserre’s hierarchy [Las01] for optimiz-
ing over scalar polynomial inequalities to polynomial matrix inequalities (PMIs).
Motivated by problems of systems control theory, the authors of [HL06] develop
the primal-moment/dual-sos approach for (non-convex) PMIs, a particular case
of which are the (convex) LMIs treated here. Our Theorem 4.3.3 shows that for
infeasible LMIs, the SDP hierarchy described by Henrion and Lasserre in [HL06]
generates a certificate of infeasibility at a finite relaxation order.

4.4. Bounded spectrahedra. In this section we establish algebraic certificates for

boundedness of a spectrahedron. As a corollary, we obtain a Putinar-Schmüdgen-
type Positivstellensatz for polynomials positive on bounded spectrahedra, see Corol-
lary 4.4.4.
A quadratic module M ⊆ R[x] is said to be archimedean if one of the following
equivalent conditions holds:
(i) ∀f ∈ R[x] ∃N ∈ PNn : N + f ∈ M ;
(ii) ∃N ∈ N : N ± i=1 x2i ∈ M ;
(iii) ∀i ∈ {1, . . . ,n} ∃N ∈ N : N ± xi ∈ M ;
see [Ma08, Corollary 5.2.4].
Obviously, if MA is archimedean for a linear pencil A, then SA is bounded. In
[HKM] complete positivity (see e.g. [ER00, BL04, Pau02, Pis03]) was used to deduce
that for strictly feasible linear pencils the converse holds. Subsequently, a certain
generalization of this result for projections of spectrahedra has been proved by
other techniques in [GN11]. In this section we will establish the result for arbitrary
bounded SA . We deal separately with the relatively easy case of non-empty SA
(possibly with empty interior).
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 21

Lemma 4.4.1. Let A be a linear pencil with SA 6= ∅. Then

SA is bounded ⇐⇒ MA is archimedean.
Proof. The direction (⇐) is obvious as remarked above. Let us consider the con-
verse. We first establish the existence of finitely many linear polynomials in CA
certifying the boundedness of SA .
There is a ball B ⊆ Rn with SA ⊆ B and SA ∩ ∂B = ∅. For every x ∈ ∂B there
is a vector u ∈ Rn with
u∗ A(x)u < 0. (23)
By continuity, (23) holds for all x in a neighborhood Ux of x. From {Ux | x ∈ ∂B}
we extract by compactness a finite subcovering {Ux1 , . . . , Uxr } of ∂B. Let ℓi :=
u∗i Aui ∈ R[x]1 and
S := {x ∈ Rn | ℓ1 (x) ≥ 0, . . . ,ℓr (x) ≥ 0}.
Clearly, SA ⊆ S and S ∩ ∂B = ∅. Since SA is non-empty by hypothesis and
contained in B, it follows that S contains a point of B. But then it follows from the
convexity of S together with S ∩ ∂B = ∅ that S ⊆ B. In particular, S is bounded.
Now every ℓi ∈ CA ⊆ MA . Hence the quadratic module M generated by the ℓi
is contained in MA . Choose N ∈ N with N ± xi > 0 on S for all i. Fix a k and
δ ∈ {−1,1}. The system of linear inequalities
−N + δxk ≥ 0, ℓ1 (x) ≥ 0, . . . ,ℓr (x) ≥ 0
is infeasible. Hence by Farkas’ lemma [Fa02], there are αj ∈ R≥0 satisfying
− 1 = α0 (−N + δxk ) + α1 ℓ1 + · · · + αr ℓr . (24)
Note α0 6= 0 since S 6= ∅. Rearranging terms in (24) yields N − δxk ∈ CA . Since
k and δ were arbitrary and CA ⊆ MA , we conclude that MA is archimedean.
Corollary 4.4.2. Let A be a linear pencil. Then
SA is bounded ⇐⇒ MA is archimedean.
Proof. If SA 6= ∅, then this is Lemma 4.4.1. If SA = ∅, then −1 ∈ MA by Corollary
4.3.1, so MA is archimedean.
Remark 4.4.3. Note that the above corollary is a strong variant of Schmüdgen’s
characterization [Sm91] of bounded basic closed semialgebraic sets as being exactly
those whose describing finitely generated preorderings are archimedean. Preorder-
ings have the tendency of being much larger than quadratic modules. In general, a
finitely generated quadratic module might describe a bounded or even an empty set
without being archimedean, see [PD01, Example 6.3.1] and [Ma08, Example 7.3.2].
Corollary 4.4.2 says that quadratic modules associated to linear pencils behave very
well in this respect.
We conclude this section with a version of Putinar’s Positivstellensatz [Pu93] for
bounded spectrahedra:
Corollary 4.4.4. Let A be a linear pencil and assume that SA is bounded. If
f ∈ R[x] satisfies f |SA > 0, then f ∈ MA .
Proof. By Corollary 4.4.2, MA is archimedean. Now apply (a slight generalization
of) Putinar’s Positivstellensatz [Ma08, Theorem 5.6.1].
Remark 4.4.5. Let A be a linear pencil with bounded SA .
(1) In the case A is strongly feasible, Corollary 4.4.4 has already been proved in
[HKM, §7] by completely different techniques, namely complete positivity from
operator algebras. Note however that the more involved case in our approach
occurs when A is infeasible.
22 IGOR KLEP AND MARKUS SCHWEIGHOFER

(2) From Corollary 4.4.2 it is easy to see that the quadratic module in the sense
of rings with involution (see [KS10]) associated to a linear pencil A of size α
in the ring of β × β matrix polynomials is archimedean (in the sense of [KS10,
Subsection 3.1] or [HKM, Sections 6, 7]) if the spectrahedron SA defined by A
is bounded (cf. [HKM, Section 7]). Among other consequences, this implies a
suitable generalization of Corollary 4.4.4 for matrix polynomials positive definite
on the bounded spectrahedron SA (cf. [SH06, Corollary 1], [KS10, Theorem 3.7]
and [HKM, Theorem 7.1]).
4.5. An application to positive linear functionals. In this brief subsection we
explain how our results pertain to positive linear functionals.
Definition 4.5.1. Suppose R ⊆ SRα×α is a vector subspace, and let R0 :=
R ∩ SRα×α
0 . A linear functional f : R → R is called positive if f (R0 ) ⊆ R≥0 .

Detecting positivity of a linear functional can be implemented with an SDP

using our Theorem 3.7.3. Let {A1 , . . . ,An } be a basis for R, and introduce the
linear pencil
A = A1 x1 + · · · + An xn ∈ SR[x]α×α
1 .
To f we associate the linear polynomial
fA = f (A1 )x1 + · · · + f (An )xn ∈ R[x]1 .
Then f is positive if and only if fA |SA ≥ 0, and this is a situation completely
characterized by our Theorem 3.4.1. In turn, it can be implemented using (Dsos ).
Remark 4.5.2. In subsequent work [HKMNS] we shall exploit how the ideas pre-
sented here apply to operator algebras and complete positivity [ER00, BL04, Pau02,
Pis03], by using the results from [HKM, HKM12].
4.6. Examples. We conclude this paper with a series of examples pertaining to
the theory developed.
Example 4.6.1. The standard textbook example [St00, WSV00] of a weakly in-
feasible linear pencil seems to be
 
x 1
A := .
1 0
(1) ∗
Then −1 6∈ CA , but −1 ∈ MA . Indeed, for u := 1 −1 − 2x , we have


−2 = u∗ Au.
Hence A is 1-infeasible.
Example 4.6.2. Let  
0 x1 0
A := x1 x2 1 .
0 1 x1
(1)
Then A is weakly infeasible and −1 6∈ MA .
Assume otherwise, and let
X
−1=s+ vj∗ Avj , (25)
j

where vj ∈ R[x]31 and s ∈ Σ2 ∩ R[x]2 . We shall carefully analyze the terms vj∗ Avj .
Write  ∗
vj = q1j q2j q3j , and qij = aij + bij x1 + cij x2
with aij ,bij ,cij ∈ R. Then the x32 coefficient of vj∗ Avj equals c22j , so c2j = 0 for all
j. Next, by considering the x1 x22 terms, we deduce c3j = 0. Now the only terms
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 23

possibly contributing to x22 come from s, so s ∈ R[x1 ]2 . The coefficient of x2 in

vj∗ Avj is a square, so a2j = 0. But now vj∗ Avj does not have a constant term
anymore, leading to a contradiction with (25).
(3) (2)
From Theorem 4.3.3 it follows that −1 ∈ MA . In fact, −1 ∈ MA since
−2 = u∗ Au
h i∗
x22
for u = 1
2 + x2
2 + 8 −1 1 + x22 ∈ R[x]32 . Thus A is 2-infeasible.
Example 4.6.3. Let
 
0 x1 0
A = x1 x2 x3  .
0 x3 x1
Then SA = {(0,x2 ,0) ∈ R3 | x2 ≥ 0} and the (affine) hyperplanes containing SA
are {x1 = 0} and {ax1 + x3 = 0} (a ∈ R). As is shown in Case 1 of the proof of
Proposition 3.3.1, the certificate of low-dimensionality (2) exists for the hyperplane
{x1 = 0}, i.e., there is a quadratic sos-matrix S such that −x21 = tr(AS). However,
none of the other hyperplanes containing SA allow for a certificate of the form (2).
Indeed, assume that there is a ∈ R such that {ax1 + x3 = 0} has a corresponding
certificate. Combining it with the one for {x1 = 0}, we get a quadratic sos-matrix
S such that
−(2a2 )x21 − 2(ax1 + x3 )2 = tr(AS)
which implies
(1)
−x23 = (2ax1 + x3 )2 + (−(2a2 )x21 − 2(ax1 + x3 )2 ) ∈ MA .
(1)
Specializing x3 to 1, one gets the contradiction −1 ∈ MA′ where A′ is the linear
pencil from Example 4.6.2.
The next example gives a more explicit presentation of our dual (Dsos ) applied
to a classical example of an SDP with nonzero duality gap.
Example 4.6.4. Consider
 
α + x2 0 0
A= 0 x1 x2  =: A0 + x1 A1 + x2 A2 ,
0 x2 0
where α ∈ R>0 and Aj ∈ SR3×3 , and let ℓ = x2 . If x ∈ R2 satisfies A(x)  0, then
x2 = 0. Thus the SDP (P) associated to A and ℓ
min x2
(26)
s.t. A(x)  0
has optimal value 0. Its standard dual (D) is
max a
s.t. − a = tr(A0 S)
0 = tr(A1 S) (27)
1 = tr(A2 S)
S  0.
From the second equation we see S2,2 = 0 and thus (since S  0), S2,j = Sj,2 = 0
for all j. Now the last equation implies S1,1 = 1, and thus −a = α by the first
equation. Hence the optimal value of (27) is −α, yielding a duality gap of α > 0.
We next present our sums of squares dual (Dsos ) for (26).
24 IGOR KLEP AND MARKUS SCHWEIGHOFER

max a
s.t. S ∈ SR3×3
0 , a ∈ R

S1 ,S2 ∈ SR[x]3×3
2 quadratic sos-matrices (28)
3×3 6×3
U1 ,U2 ∈ SR , W1 ,W2 ∈ R
#„ ∗ #„
[x]1 U1 [x]1 + tr(AS1 ) = 0 (29)
U1  W1∗ W1 (30)
#„ ∗ #„ #„ ∗ #„
[x]1 U2 [x]1 + [x]2 W1 [x]1 + tr(AS2 ) = 0 (31)
U2  W2∗ W2 (32)
#„ ∗ #„
ℓ − a + [x]2 W2 [x]1 − tr(AS) = 0. (33)

To express (28) using LMI constraints, note that

2
X X
D+ xk Ek + xk xℓ Fk,ℓ ∈ SR[x]3×3
2 (34)
k=1 1≤k≤ℓ≤2

is an sos-matrix iff there are 3 × 3 matrices Ξ1 ,Ξ2 and Φ1,2 satisfying

 
D Ξ1 Ξ2
Ξ∗1 F1,1 Φ1,2   0
Ξ∗2 Φ∗1,2 F2,2 (35)
Ξk + Ξ∗k = Ek , k = 1,2
Φ1,2 + Φ∗1,2 = F1,2 .

Let us use the notation of (34) for the quadratic sos-matrix S1 , and consider the
left hand side of (29). Its constant coefficient is tr(A0 D) + (U1 )1,1 = 0. Since
A0 ,D,U1  0, this implies (U1 )1,j = (U1 )j,1 = 0 and D1,j = Dj,1 = 0. Next,
the x31 term of (29) is tr(A1 F1,1 ) = 0,h whence
i (F1,1 )2,j = (F1,1 )j,2 = 0. From
0 0 0
(35) it follows that Ξ1 is of the form ∗0∗ . Hence (E1 )2,2 = 0. Finally, by
∗0∗
considering the x21 term in (29), we obtain tr(A0 F1,1 ) + tr(A1 E1 h) + (Ui1 )2,2 = 0.
00 0
Since (E1 )2,2 = 0, A1 E1 = 0. As A0 ,F1,1 ,U1  0, we deduce U1 = 0 0 0 for some
00u
u ∈ R≥0 . In particular, from (30) we see the first two columns of W1 are 0. Hence
#„ ∗ #„
[x]2 W1 [x]1 ∈ x2 R[x].
Using this information on W1 , we can analyze (31) as in the previous paragraph,
and deduce that the first two columns of W2 are 0. Next, we turn to (33). All its
#„ ∗ #„
terms of degree ≥ 2 come from [x]2 W2 [x]1 , so (W2 )j,k = 0 for all (j,k) 6= (1,3).
This reduces (33) to the system of linear equations

−a = tr(A0 S)
0 = tr(A1 S)
1 + (W2 )1,3 = tr(A2 S).

It is instructive to compare this to (27) above. Again S2,j = Sj,2 = 0 for all j.
Since tr(A0 S) ≥ 0 and we are maximizing a, we set S1,1 = 0, yielding (W2 )1,3 = −1
and a = 0. It is now easy to see that this W2 and a can be extended to a feasible
(and thus optimal) point for the above sums of squares dual.
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 25

Finally, we give Ramana’s dual for (26); we refer to [Ra97, p. 142] for details.
max a
s.t. 0 = tr(Aj U1 ), j = 0,1,2
U1  W1 W1∗


0 = tr Aj (U2 + W1 ) , j = 0,1,2
(36)
U2  W2 W2∗


0 = tr Aj (U + W2 ) , j = 1,2


−a = tr A0 (U + W2 )
U  0.
The reader will have no problems verifying that the optimal value of (36) is 0.
Example 4.6.4 demonstrates that Ramana’s dual is generally smaller in size than
the sums of squares dual. However, the advantage of our sums of squares dual is
that it admits a nice real algebraic geometric interpretation, and naturally leads
itself to the Positivstellensätze we presented in Section 4.

Acknowledgments. The authors thank three anonymous referees for their de-
tailed reading and many helpful comments.

References
[Ba01] A. Barvinok: A remark on the rank of positive semidefinite matrices subject to affine
constraints, Discrete Comput. Geom. 25 (2001), no. 1, 23–31 5
[Ba02] A. Barvinok: A course in convexity, Graduate Studies in Mathematics 54, Amer.
Math. Soc., 2002 5, 6
[BCR98] J. Bochnak, M. Coste, M.F. Roy: Real algebraic geometry, Springer-Verlag, 1998 4, 10,
15
[BEFB94] S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan: Linear matrix inequalities in system
and control theory, Studies in Applied Mathematics 15, SIAM, 1994 2, 3
[BL04] D.P. Blecher and C. Le Merdy: Operator algebras and their modules—an operator
space approach, London Mathematical Society Monographs. New Series 30, Oxford
University Press, 2004 20, 22
[Bo48] F. Bohnenblust: Joint positiveness of matrices, Technical report, California Institute
of Technology, 1948; available at
http://orion.uwaterloo.ca/~hwolkowi/henry/book/fronthandbk.d/Bohnenblust.pdf 5
[BV96] S. Boyd, L. Vandenberghe: Semidefinite programming, SIAM Rev. 38 (1996), no. 1,
49–95 1, 3
[BW81] J.M. Borwein, H. Wolkowicz: Facial reduction for a cone-convex programming problem,
J. Austral. Math. Soc. Ser. A 30 (1980/81), no. 3, 369–380 12
[Ce10] G. Chesi: LMI techniques for optimization over polynomials in control: a survey, IEEE
Trans. Automat. Control 55 (2010), no. 11, 2500–2510 2
[dK02] E. de Klerk: Aspects of semidefinite programming, Interior point algorithms and se-
lected applications. Applied Optimization 65, Kluwer Acad. Publ., 2002 2, 3
[Du07] B. Dumitrescu: Positive trigonometric polynomials and signal processing applications,
Springer-Verlag, 2007 2
[ER00] E.G. Effros, Z.-J. Ruan: Operator spaces, London Mathematical Society Monographs.
New Series 23, Oxford University Press, 2000 20, 22
[Fa02] J. Farkas: Theorie der einfachen Ungleichungen, J. reine angew. Math. 124 (1902),
1–27 21
[GLR82] I. Gohberg, P. Lancaster, L. Rodman: Matrix polynomials, Computer Science and
Applied Mathematics, Academic Press, 1982 3
[GN11] J. Gouveia, T. Netzer: Positive polynomials and projections of spectrahedra, SIAM J.
Optim. 21 (2011), no. 3, 960–976 20
26 IGOR KLEP AND MARKUS SCHWEIGHOFER

[Go97] M. Goemans: Semidefinite programming in combinatorial optimization, Lectures on

mathematical programming, Math. Program. 79 (1997), no. 1-3, Ser. B, 143–161 2, 3
[HKM] J.W. Helton, I. Klep, S. McCullough: The matricial relaxation of a linear matrix
inequality, to appear in Math. Program. http://arxiv.org/abs/1003.0908 20, 21, 22
[HKM12] J.W. Helton, I. Klep, S. McCullough: The convex Positivstellensatz in a free algebra,
Adv. Math. 231 (2012) 516–534 22
[HKMNS] J.W. Helton, I. Klep, S. McCullough, C.Nelson, M. Schweighofer: Complete positivity
for nonunital operator systems and the hollow Positivstellensatz, in preparation 22
[HG05] D. Henrion, A. Garulli (editors): Positive polynomials in control, Lecture Notes in
Control and Information Sciences 312, Springer-Verlag, 2005 2
[HL06] D. Henrion, J.B. Lasserre: Convergent relaxations of polynomial matrix inequalities
and static output feedback, IEEE Trans. Autom. Control 51 (2006), no. 2, 192–202 20
[Ko03] M. Kojima: Sums of squares relaxations of polynomial semidefinite programs, Technical
report, Tokyo Institute of Technology, 2003 10
[KS10] I. Klep, M. Schweighofer: Pure states, positive matrix polynomials and sums of Her-
mitian squares, Indiana Univ. Math. J. 59 (2010), no. 3, 857–874 22
[Las01] J.B. Lasserre: Global optimization with polynomials and the problem of moments,
SIAM J. Optim. 11 (2000/01), no. 3, 796–817 20
[Las10] J.B. Lasserre: Moments, positive polynomials and their applications, Optimization
Series 1, Imperial College Press, 2010 2, 3
[Lau09] M. Laurent: Sums of squares, moment matrices and optimization over polynomials,
Emerging applications of algebraic geometry, 157–270, IMA Vol. Math. Appl. 149,
Springer, New York, 2009; updated version available at
http://homepages.cwi.nl/~monique/files/moment-ima-update-new.pdf 2, 10
[Ma08] M. Marshall: Positive polynomials and sums of squares, Mathematical Surveys and
Monographs 146, Amer. Math. Soc., 2008 2, 4, 10, 14, 15, 20, 21
[Ne07] A. Nemirovski: Advances in convex optimization: conic programming, International
Congress of Mathematicians, Vol. I, 413–444, Eur. Math. Soc., 2007 2
[NN94] Y. Nesterov, A. Nemirovskii: Interior-point polynomial algorithms in convex program-
ming, Studies in Applied Mathematics 13, SIAM, 1994 2, 3
[Par00] P.A. Parrilo: Structured Semidefinite Programs and Semialgebraic Geometry Methods
in Robustness and Optimization, Ph.D. thesis, California Institute of Technology, 2000
2, 3
[Pat00] G. Pataki: A Simple Derivation of a Facial Reduction Algorithm, and Extended Dual
Systems, Technical report, Columbia University, 2000
http://www.unc.edu/~pataki/papers/fr.pdf 12
[Pat] G. Pataki: Bad semidefinite programs: they all look the same, preprint
http://arxiv.org/abs/1112.1436 7
[Pau02] V. Paulsen: Completely bounded maps and operator algebras, Studies in Advanced
Mathematics 78, Cambridge University Press, 2002 20, 22
[PD01] A. Prestel, C.N. Delzell: Positive polynomials. From Hilbert’s 17th problem to real
algebra, Springer-Verlag, 2001 2, 4, 10, 15, 21
[Pis03] G. Pisier: Introduction to operator space theory, London Mathematical Society Lecture
Note Series 294, Cambridge University Press, 2003 20, 22
[PK97] L. Porkolab, L. Khachiyan: On the complexity of semidefinite programs, J. Global
Optim. 10 (1997), no. 4, 351–365 2
[PS01] V. Powers, C. Scheiderer: The moment problem for non-compact semialgebraic sets,
Adv. Geom. 1 (2001), no. 1, 71–88 7
[Pu93] M. Putinar: Positive polynomials on compact semi-algebraic sets, Indiana Univ. Math.
J. 42 (1993), no. 3, 969–984 21
[Ra97] M. Ramana: An exact duality theory for semidefinite programming and its complexity
implications, Math. Program. 77 (1997), no. 2, Ser. B, 129–162 2, 12, 25
[Sc09] C. Scheiderer: Positivity and sums of squares: a guide to recent results, Emerging
applications of algebraic geometry, 271–324, IMA Vol. Math. Appl. 149, Springer-
Verlag, 2009 2, 15
[SH06] C.W. Scherer, C.W.J. Hol: Matrix sum-of-squares relaxations for robust semi-definite
programs, Math. Program. 107 (2006), no. 1-2, Ser. B, 189–211 10, 22
 A SUMS OF SQUARES DUAL FOR SDP AND INFEASIBLE LMI 27

[Sm91] K. Schmüdgen: The K-moment problem for compact semi-algebraic sets, Math. Ann.
289 (1991), no. 2, 203–206 21
[Sr86] A. Schrijver: Theory of linear and integer programming, Wiley-Interscience Series in
Discrete Mathematics, John Wiley & Sons, 1986 1, 2
[St00] J. Sturm: Theory and algorithms of semidefinite programming, High performance op-
timization, 1–194, Appl. Optim. 33, Kluwer Acad. Publ., 2000 3, 4, 22
[To01] M.J. Todd: Semidefinite optimization, Acta Numer. 10 (2001), 515—560 1, 3
[TW] L. Tuncel, H. Wolkowicz, Strong Duality and Minimal Representations for Cone Opti-
mization, to appear in Comput. Optim. Appl.
http://orion.math.uwaterloo.ca/~hwolkowi/henry/reports/regclosegap.pdf 12
[WSV00] H. Wolkowicz, R. Saigal, L. Vandenberghe (editors): Handbook of semidefinite pro-
gramming. Theory, algorithms, and applications, Kluwer Acad. Publ., 2000 1, 3, 22
[Za] A. Zalar: A note on a matrix version of the Farkas lemma, Comm. Algebra 40 (2012)
3420–3429 17

Igor Klep, Department of Mathematics, The University of Auckland, Private Bag

92019, Auckland 1142, New Zealand
E-mail address: [email protected]

Markus Schweighofer, Universität Konstanz, Fachbereich Mathematik und Statistik,

78457 Konstanz, Allemagne
E-mail address: [email protected]
28 IGOR KLEP AND MARKUS SCHWEIGHOFER

NOT FOR PUBLICATION

Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Reader’s guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Notation and terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1. Matrix polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2. Linear pencils and spectrahedra. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3. Sums of squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4. Radical ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Duality theory of semidefinite programming. . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1. Weakly feasible and weakly infeasible linear pencils . . . . . . . . . . . . . . . . 4
3.2. An algebraic glimpse at standard SDP duality . . . . . . . . . . . . . . . . . . . . 6
3.3. Certificates for low dimensionality of spectrahedra . . . . . . . . . . . . . . . . . 8
3.4. Linear polynomials positive on spectrahedra . . . . . . . . . . . . . . . . . . . . . . 9
3.5. Constructing SDPs for sums of squares problems . . . . . . . . . . . . . . . . . . 10
3.6. Real radical computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.7. A new exact duality theory for SDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4. Positivity of polynomials on spectrahedra. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.1. Quadratic module associated to a linear pencil . . . . . . . . . . . . . . . . . . . . 14
4.2. Linear polynomials positive on spectrahedra – revisited . . . . . . . . . . . . . 15
4.3. Infeasibility of linear pencils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.4. Bounded spectrahedra. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.5. An application to positive linear functionals . . . . . . . . . . . . . . . . . . . . . . 22
4.6. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25