Stability of closedness of convex cones

under linear mappings

1
Jonathan M. Borwein and Warren B. Moors

Abstract. In this paper we reconsider the question of when the continuous linear
image of a closed convex cone is closed in Euclidean space. In particular, we show
that although it is not true that the closedness of the image is preserved under small
perturbations of the linear mappings it is “almost” true that the closedness of the
image is preserved under small perturbations, in the sense that, for “almost all” linear
mappings from Rn into Rm if the image of the cone is closed then there is a small
neighbourhood around it whose members also preserve the closedness of the cone.

AMS (2002) subject classification: Primary 47N10 ; Secondary 90C25, 90C22.

Keywords: closed convex cone; linear mapping; stability; linear programming.

1 Introduction
We say that a nonempty subset K of a vector space V is a cone if for each λ ∈ [0, ∞) and
each x ∈ K, λx ∈ K. If {a1 , a2 , . . . , an } is a finite subset of a vector space V then we shall
denote by ha1 , a2 , . . . , an i the cone generated by {a1 , a2 , . . . , an } i.e.,
( n )
X
ha1 , a2 , . . . , an i := λk ak : 0 ≤ λk < ∞ for all 1 ≤ k ≤ n .
k=1

Further, we shall say that a convex cone K in a vector space V is finitely generated if there
exists a finite set {a1 , a2 , . . . , an } ⊆ V such that K = ha1 , a2 , . . . , an i. By [2, page 25] we
know that each finitely generated cone in a normed linear space X is closed. In fact, each
finitely generated cone is a polyhedral set i.e., a finite intersection of closed half-spaces, [2,
page 99].
If X and Y are finite dimensional normed linear spaces then we shall denote by, L(X, Y ) the
set of all linear transformations from X into Y . Throughout this paper we shall assume that
L(X, Y ) is endowed with a Hausdorff linear topology. Since all Hausdorff linear topologies on
finite dimensional spaces are homeomorphic, [3, page 51] we shall, with out loss of generality,
assume that the topology on L(X, Y ) is generated by the operator norm on L(X, Y ) and
that the topologies on X and Y are generated by the Euclidean norms.
In this paper we will examine the question of whether the continuous linear image of a
closed convex cone is closed. The motivation for this will be well known to many readers:
1
This paper is dedicated to our friend and colleague Stephen Simons on the occasion of his 70th birthday.

1
the abstract version of the Farkas lemma [2, p. 24] or the Krein-Rutman theorem [2, Cor.
3.3.13] asserts that for a closed convex cone K in Y and A in L(X, Y ) one has

+
A−1 K = A∗ (K + ) (1)

if, and only, if A∗ (K + ) is closed. Here K + := {x ∈ Y : hx, yi ≥ 0, for all y ∈ K} is the

positive dual cone and A∗ denotes the transpose operator. Formula (1) in turn is the basis
of strong duality in abstract linear programming [2, §5.3] and of the Karush-Kuhn-Tucker
theorem [2, §6.1] and its generalizations.
A naive guess—based on two-dimensional reasoning—might be that the continuous linear
image of a closed convex cone is always closed and in fact this is the case for finitely generated
cones (or, equivalently, polyhedral cones). There are, however, simple examples (see for
instance Example 1) that show that this naive guess is false. One might then speculate
that the closedness of the image of a closed convex cone under a continuous linear mapping
might, at least, be preserved under small perturbations of the linear map. This is made more
plausible by the recent literature on the distance to inconsistency for abstract inequality
systems [2, p.122] viewed as a generalization of the condition number. As pioneered by
Renegar and others, there is a strictly positive distance to inconsistency for a system {x ∈
X : Ax − b ∈ K} and one might hope this is true for closure.

2 Two Limiting Examples

Such speculation is also refuted by Example 1 which brings us to the other purpose of this
paper. This is to (i) analyze when the continuous linear image of a closed convex cone
is closed and to (ii) provide some sufficiency conditions under which the continuous linear
image of a closed convex cone is closed. There is surprisingly little literature on the issue of
when precisely a conical linear image is closed, see for example [4] and [1].

Example 1 Let

K := {(w, x, y, z) ∈ R4 : 0 ≤ w, 0 ≤ x and y 2 + (z − x)2 ≤ x2 }.

Then K is an inverse linear image of the right-circular cone and so closed and convex. For
each λ ≥ 0 define the linear mapping Tλ : R4 → R3 by, Tλ (w, x, y, z) := (x − λw, y, z), a
rank-one linear perturbation of T0 . Note that if λ = 0 then Tλ (K) is a closed cone in R3 ,
but for every λ > 0 the image

Tλ (K) ≡ {(x, y, z) : z > 0} ∪ {(x, 0, 0) : x ∈ R}

which is convex but not closed. k



Example 1 shows that the closedness of the image of a closed convex cone under a linear
mapping is not stable even under arbitrarily small rank-one perturbations. A more concrete,
but closely related example, is given by the following abstract linear program.

2
Example 2 Let us consider the following closed convex cone in R7 . We simplify things by
letting z := (x1 , x2 , x3 , x4 , y1 , y2 , y3 ) denote a point in R7 and let

K := {z ∈ R7 : 0 ≤ x1 , 0 ≤ x2 , x23 + (x4 − x2 )2 ≤ x22 , 0 ≤ y1 , y22 + (y3 − y1 )2 ≤ y12 }

and let the linear mapping z ∗ : R7 → R be defined by z ∗ (z) := x4 + y3 . For each 0 ≤ λ < ∞
and µ ∈ R let us define    
−λ 1 0 0 0 0 0 1
Aλ :=  0 0 1 0 0 0 0  and β µ :=  1  .
0 0 0 0 0 1 0 µ
Then for each 0 ≤ λ < ∞ and µ ∈ R we can consider the optimization problem.

E(λ, µ) := inf{z ∗ (x) : x ∈ K and Aλ x = β µ }.

It is easy to check that this infimum is obtained if, and only if, λ = µ = 0. k


Example 2 shows that the existence of minima in abstract linear programming problems is
not stable under arbitrarily small rank-one perturbations. This again highlights the difficulty
of exactly characterizing closure of a conical linear image.

3 Preliminary Positive Results

We first collect and improve various known results. Versions of each are to be found in [2].
Our first result shows that the closedness of the image of a closed convex cone is related to
the closedness of the sum of a closed convex cone with a finite dimensional subspace.

Proposition 1 Let T ∈ L(Rn , Rm ) and let ∅ 6= K ⊆ Rn be an arbitrary set. Then T (K) is

closed in Rm if, and only if, K + ker(T ) is closed in Rn .

Proof: Suppose that T (K) is closed. Then T −1 (K) = K + ker(T ) is closed in Rn , since T is
continuous. Conversely, suppose that K + ker(T ) is closed in Rn . Then C := [K + ker(T )] ∩
[ker(T )]⊥ is also closed in Rn , and moreover, T (K) = T (C). Now, T |ker(T )⊥ is a 1-to-1 linear
mapping (and hence a homeomorphism) onto T ([ker(T )]⊥ ); which is a closed subspace of
Rm . Therefore, T (K) = T (C) is closed in T ([ker(T )]⊥ ) and hence closed in Rm . k


Proposition 2 Suppose that K is a finitely generated convex cone in a vector space X. If

T : X → Y is a linear mapping into a normed linear space Y then T (K) is a closed convex
cone in Y .

Proof: Since K is finitely generated there exists a finite set {a1 , a2 , . . . , an } in X such that
K = ha1 , a2 , . . . , an i. A simple calculation then reveals that T (K) = hT (a1 ), T (a2 ), . . . , T (an )i;
which is finitely generated and hence closed, [2, page 25] k


Next we give some sufficiency conditions for the image of a closed convex cone to be closed.

3
Proposition 3 Let T ∈ L(Rn , Rm ) and let K be a closed cone (not necessarily convex) in
Rn . If
K ∩ ker(T ) = {0}
then there exists a neighbourhood N of T in L(Rn , Rm ) such that S(K) is closed in Rm for
each S ∈ N .

Proof: Let C := {k ∈ K : kkk = 1}. Then both C and T (C) are compact and 0 6∈ T (C)
therefore dist(0, T (C)) > 0 and so there exists a neighbourhood N of T in L(Rn , Rm ) such
that dist(0, S(C)) > 0 for each S ∈ N . Since
S(K) = {λc : c ∈ S(C) and 0 ≤ λ < ∞}
and S(C) is compact for each S ∈ N , it follows that S(K) is closed in Rm for each S ∈ N ,
as claimed. k


Proposition 4 Let T ∈ L(Rn , Rm ) and let K be a closed convex cone in Rn . If K ∩ ker(T )

is a linear subspace then T (K) is closed convex cone in Rm .

Proof: Let M := K ∩ ker(T ) and let N := M ⊥ ∩ [K + M ] = M ⊥ ∩ K. Then T (N ) = T (K)

and N ∩ ker(T ) = {0}. Therefore, T (K) = T (N ) is a closed convex cone on application of
Proposition 3. k


For a subset D of a vector space V , the core of D, denoted, cor(D), is the set of all points
d ∈ D where for each x ∈ V \ {d} there exists an 0 < r < 1 such that λx + (1 − λ)d ∈ D
for all 0 ≤ λ < r. Clearly if the affine span aff(D) 6= V then cor(D) = ∅. In this case the
following concept is useful.
Given a subset C of a vector space V , the intrinsic core of C, denoted icor(A), is the set of all
points c ∈ C where for each x ∈ aff(C) there exists an 0 < r < 1 such that λx + (1 − λ)c ∈ C
for all 0 ≤ λ < r.
One of the most important properties of the intrinsic core is that if C is a convex subset of
a finite dimensional vector space V then icor(C) 6= ∅, [3, page 7]. In fact, if V is a finite
dimensional topological vector space then icor(C) is dense in C for each convex subset C
of the space V . Another important property of the core is that for a convex subset C of a
finite dimensional topological vector space, cor(C) = int(C), [2, Theorem 4.1.4].
The reason for our interest in the intrinsic core is based in the following result.

Proposition 5 Let Y be a normed linear space, T : Rn → Y be a linear transformation and

let K be a closed cone in Rn . If
ker(T ) ∩ icor(K) 6= ∅
then T (K) is a finite dimensional linear subspace of Y and hence a closed convex cone.

Proof: By [2, Problem 13 part (e)], {0} = T (ker(T ) ∩ icor(K)) ⊆ T (icor(K)) ⊆ icor(T (K)).
Since T (K) is a cone we see that T (K) = aff(T (K)) = span(T (K)). The result now follows,
since every finite dimensional subspace of a normed linear space is closed. k


From Proposition 4 and Proposition 5 we see that:

4
Corollary 1 The only way T (K) can fail to be closed is if
ker(T ) ∩ K ⊆ K \ icor(K)
and that at the same time ker(T ) ∩ K is not a linear subspace.

4 The Main Results

We are now ready for our principle positive results.
Lemma 1 Suppose that K is a closed convex cone in Rn , Y := K − K, T ∈ L(Rn , Rm ) and
T |Y ∈ L(Y, Rm ) has rank m. If
ker(T ) ∩ icor(K) 6= ∅
then there exists a neighbourhood W of T in L(Rn , Rm ) such that
ker(S) ∩ icor(K) 6= ∅
for all S ∈ W. In particular, S(K) is a closed convex cone in Rm for each S ∈ W.
Proof: Let M ⊆ L(Y, Rm ) be the family of all mappings with rank m. It is routine to show
that M is a dense open subset of L(Y, Rm ) since T |Y ∈ M and so m ≤ Dim(Y ).
(a) We shall consider first the case when Y = Rn . For each S ∈ L(Rn , Rm ), let AS denote
the matrix representation of S with respect to the standard bases on Rn and Rm . Since
rank(S) = m for all S ∈ M, the rows of each AS , S ∈ M are linearly independent. Therefore,
for each S ∈ M, the matrix AtS (AS AtS )−1 AS is well-defined and represents the orthogonal
projection of Rn onto the row space of AS . Since the null space of AS is perpendicular to
the row space of AS , In − AtS (AS AtS )−1 AS is the matrix representation of the projection of
Rn onto the null space of AS . Here In denotes the n × n identity matrix.
Next, we shall consider M(n,n) , the space of all n × n matrices, endowed with the linear
topology of component-wise convergence. With respect to this topology the mapping ϕ :
M → M(n,n) defined by, ϕ(S) := In − AtS (AS AtS )−1 AS is continuous. Moreover, for any fixed
x ∈ Rn , the mapping S 7→ ϕ(S)(x) is continuous on M and ϕ(S)(x) ∈ ker(S) for all S ∈ M.
Therefore, if we choose x ∈ ker(T ) ∩ icor(K) = ker(T ) ∩ cor(K) = ker(T ) ∩ int(K) then
there exists a neighbourhood W of T in L(Rn , Rm ) such that ϕ(S)(x) ∈ ker(S) ∩ int(K) =
ker(S) ∩ cor(K) = ker(S) ∩ icor(K) for all S ∈ W. This completes the proof for the special
case when Y = Rn .
(b) In the general case, consider the mapping R : L(Rn , Rm ) → L(Y, Rm ) defined by,
R(S)(x) := S(x) for all x ∈ Y . Then R is a continuous linear mapping from L(Rn , Rm ) into
L(Y, Rm ). We now apply the first part of the proof to R(T ) ∈ M to obtain a neighbourhood
W 0 of R(T ) in L(Y, Rm ) such that ker(S) ∩ icor(K) 6= ∅ for all S ∈ W 0 . Therefore, if we let
W := R−1 (W 0 ) then W is an open neighbourhood of T in L(Rn , Rm ) and ker(S) ∩ icor(K) ⊇
ker(R(S)) ∩ icor(K) 6= ∅ for all S ∈ W. k


The next result shows—as promised—that although it is not true that, if T (K) is closed for
some closed convex cone K then S(K) is closed for all S in some neighbourhood of T , it is
“almost” true, in the sense that for “almost all” T ∈ L(Rn , Rm ) if T (K) is closed then there
exists a neighbourhood W of T such that S(K) is closed for all S ∈ W.

5
Theorem 1 Suppose that K is a closed convex cone in Rn then

int{T ∈ L(Rn , Rm ) : T (K) is closed}

is a dense open subset of L(Rn , Rm ).

Proof: Let Y := K − K, and let M ⊆ L(Rn , Rm ) be the family of all linear mappings T
such that T |Y has maximal rank (i.e., rank(T |Y ) = min{m, Dim(Y )}). It is standard that
M is a dense open subset of L(Rn , Rm ). Hence it will be sufficient to show that

int{T ∈ L(Rn , Rm ) : T (K) is closed}

is dense in M. If Dim(Y ) ≤ m then T |Y is one-to-one for each member of M and so

ker(T ) ∩ K = {0} for each T ∈ M and thus we are done by Proposition 3. Hence we shall
assume that m < Dim(Y ). Let T be any element of M and let N be any neighbourhood of T
in M. If ker(T ) ∩ K = {0} then we are again done by Proposition 3. So let us suppose that
{0} =
6 ker(T ) ∩ K. If ker(T ) ∩ icor(K) 6= ∅ then by Lemma 1 there exists a neighbourhood
N 0 of T in N such that S(K) is closed for each S ∈ N 0 . Thus, we will suppose that

{0} =
6 ker(T ) ∩ K ⊆ K \ icor(K).

Choose k0 ∈ [ker(T ) ∩ K] \ [{0} ∪ icor(K)]. Then since inf{kT (k)k : k ∈ icor(K) ∩ V} = 0 for
each neighbourhood V of k0 there exists a k 0 ∈ icor(K) and S ∈ N such that S(k 0 ) = 0. We
now re-apply Lemma 1 to obtain a neighbourhood U of S in N such that ker(S 0 ) ∩ icor(K) 6=
∅ for all S 0 ∈ U and so S 0 (K) is a closed subspace for each S 0 ∈ U. k


This result should be compared to a corresponding result on negligibility in Hausdorff mea-

sure in [5]. These results are largely motivated by semi-definite programming, [6].

Corollary 2 For any given closed convex cone K, the abstract Farkas lemma of equation
(1) holds for a dense open set of operators.

Proof: The adjoint mapping between L(Rn , Rm ) and L(Rm , Rn ) preserves both density and
openness. Indeed, for a dense open set of operators both A(K) and A∗ (K + ) are simultane-
ously closed. k


References
[1] H. H. Bauschke, J. M. Borwein and P. Tseng, “Metric regularity, strong CHIP, and CHIP are
distinct properties,” J. Convex Analysis, 7 (2000), 395–412.

[2] J. M. Borwein and A. S. Lewis, Convex analysis and nonlinear optimization. Theory and ex-
amples. Second edition. CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC,
3, Springer, New York, 2006.

[3] John R. Giles, Convex analysis with applications in the differentiation of convex functions.
Research Notes in Mathematics 58 Pitman (Advanced Publishing Program) Boston, Mass-
London, 1982.

6
[4] Gabor Pataki, “On the closedness of the linear image of a closed convex cone,” Math of Oper.
Res., 32 (200x), 395–412.

[5] Gabor Pataki and L. Tuncel, “On the Generic Properties of Convex Optimization Problems in
Conic Form,” Mathematical Programming A, 89 (2001), 449–457.

[6] H. Wolkowicz, R. Saigal, and L. Vandenberghe, “Handbook of Semidefinite Programming: The-

ory, Algorithms, and Applications”, International Series in Operations Research & Management
Science, 27, Springer-Verlag, 2001.

Jonathan M. Borwein,
Faculty of Computer Science,
Dalhousie University,
Halifax, Nova Scotia B3H 1W5,
Canada.
Email: [email protected]
URL: http://users.cs.dal.ca/∼jborwein/
Warren B. Moors,
Department of Mathematics,
The University of Auckland,
Private Bag 92019, Auckland,
New Zealand.
Email: [email protected]
URL: http://www.math.auckland.ac.nz/∼moors/