8 The Set of Low Rank Matrices
8.3 Lifted Sets: A Dictionary
Proposition 8.2.10(tangent space, (Vandereycken, 2013, Proposition 2.1)). Let S≤rbe the set of matrices inRm×nwith rank at most r, let X ∈S≤rwithrank(X) =r, and let UΣV>be
Example 8.2.12. Now we consider a simple example with the assumption that n > m > 2r:
Take x¯ = at most 2r while the only matrix in C with rank less than 2r is x itself. This is an example¯ of an affine subspace that intersects S≤r at only one point and which is tangential to S≤r. By (Hesse, 2014, Theorem 5.19 c)), since max
hu,viu∈TS≤r(x¯)∩B,v ∈TC(x¯)∩B = 1, we cannot have a locally linearly regular intersection C∩S≤r. See Figure 8.3 for an example of the resulting slow convergence of alternating projections between C and S≤r.
8.3 Lifted Sets: A Dictionary
In Chapter 10, we will study the behavior of the method of alternating projections where one of the sets is the set of matrices of rank one. As a preparation of this, we establish a link between constraint sets inCn and constraint sets of matrices in Cn×n. The motivation in our case is the fact that we can reformulate the physical problem of phase retrieval inCnin terms of minimizing the rank of a matrix over an affine subspace of matrices inCn×n.
8 The Set of Low Rank Matrices
Figure 8.3: Convergence of alternating projections between the tangent space ofS≤rand S≤ritself.
1. for all Hermitian X∈ C ∩S≤1there exists x∈ C such that xx∗ =X;
2. xx∗ ∈ Cfor all x ∈C.
Remark 8.3.2. For a given set C ⊂ Cn, Definition 8.3.1 does not determine the lift of C uniquely. The idea is more totranslateconstraints, formulated inCn, into well-described con-straints in the matrix space. Then an analysis of the intersection of the lift with the set of rank one matrices also gives further insight on C. In particular, for a subset C⊂Cn, it would be fully sufficient to takeC B{xx∗|x∈C}. However, in order to benefit from additional structure of lifts of C, we can also take for example an affine hull of{xx∗|x∈C}.
Proposition 8.3.3(uniqueness up to a global phase). For all Hermitian X∈ C ∩S≤1, the vector x∈C such that xx∗ = X is determined up to global phase. This means that from some rank one matrixxx∗ the vector xis unique at least up to a global phaseθ. Assume now that there existsy∈Cnsuch thatyy∗ = xx∗andy ,eiθx for allθ ∈ [0, 2π]. Becauseyy∗ = xx∗, we know thatxjxj = yjyj for allj. Hence, there existθ1, . . . ,θn such thatyj = eiθjxj for all j. Ify , eiθx for allθ ∈ [0, 2π], then there exists at least oneksuch thatθk , θjfor allj, k. We look at one of those entries ofyy∗
88
8.3 Lifted Sets: A Dictionary
indexed bykandjwithθj ,θk:
yy∗jk =yjyk =eiθj xje−iθkxk =ei(θj−θk)xx∗jk ,xx∗jk. (8.14) This is a contradiction. Hence, there exists someθ ∈[0, 2π]such thaty=eiθx.
8.3.1 Lifts of Linear Spaces and Cones
We focus now on affine subspaces and cones ofCn. The first and easiest example would be the lift of an affine subspace ofCn. Additionally, we consider two examples moti-vated by the physical application of phase retrieval that will be revisited in Definition 9.2.2.
Proposition 8.3.4(lifts of affine spaces). Let A : Cm → Cn be a linear mapping and let b∈Cm. Define the affine subspace
BB{x∈Cn|Ax=b}. (8.15)
Then the set
B BX∈Cn×nAXA∗ = bb∗ (8.16)
is an affine subspace, and it is a lift of B.
Proof. If there is some x ∈ B, then the rank one matrix xx∗ satisfies the equation Axx∗A∗ =bb∗.
The setBis an affine subspace ofCn×nif for everyλ∈Cwe haveλx+ (1−λ)y∈ B for allx,y∈ B. Choose some arbitraryλand someX,Y∈ B. Then
A(λX+ (1−λ)Y)A∗ = AλXA∗+A(1−λ)YA∗
=λAXA∗+ (1−λ)AYA∗ =bb∗.
Lemma 8.3.5(support constraints). Let I be an index set in{1, . . . ,n}and define
CsB
x∈Cnxj =0if j<I . (8.17) Then the set
CsBX∈Cn×n Xjk =0if j< I or if k< I is a lift of Cs.
Proof. LetX∈ Cand rank(X) =1. ThenXis Hermitian, and it can be written as some xx∗forx∈Cn. Suppose there existsj< I. ThenXjk=0 for all 1≤k ≤n. If there exists ksuch thatXkk,0, then this means thatxj =0, and thusx∈Cs.
On the other hand, ifx ∈Cs, then, for allj<I, we havexj =0. This gives usxx∗jk=0 for all 1≤k≤ n, which is all that is required to show thatxx∗ ∈ Cs.
8 The Set of Low Rank Matrices
Proposition 8.3.6(lift of a convex cone). Let A:Cm →Cnbe a linear mapping. Define the convex cone
K={x ∈Cn|Ax≥0}. (8.18)
The set
KBX∈Cn×n AXA∗ ≥0 (8.19)
is a lift of K, and it is a convex cone.
Proof. Note that for every convex coneKwe have 0∈ K. This means that for everyx∈ Kwe haveAxx∗A∗ ≥ 0. On the other hand, if a positive semidefinite rank one matrix xx∗ satisfies Axx∗A∗ ≥ 0, then either x ∈ Kor −x ∈ K. LetX ∈ K, then˜ AXA∗ ≥ 0, and naturally, for allλ ≥ 0, we have AλXA∗ = λAXA∗ ≥ 0. On the other hand, let λ∈[0, 1]andX,Y. ThenA(λX+ (1−λ)Y)A∗ = λAXA∗+ (1−λ)AYA∗≥0.
Lemma 8.3.7(positive real cone). For
CBRn≥0=x∈Cnx ∈Rn,xj ≥0for all j , (8.20) the setC BnX∈Cn×nX∈ Rn≥×0n,X∗ = Xo
is a lift of C.
Proof. LetX∈ Csand rank(X) =1. ThenXis Hermitian, and it can be written as some xx∗ for x ∈ Cn. Let j be such that Xjj , 0. If such a j does not exist, then X is the zero matrix. Choosexj = pXjj > 0. Because xjxk ∈ R≥0 for allk, we conclude that
xk ∈R≥0.
8.3.2 Quadratic Constraints
The examples so far gave subspaces as lifts of subspaces. An interesting example of the benefits of lifted sets is the following: suppose we are given an unitary mapping A:Cn →Cnand, for someb∈Rn≥0, the set
MB n
x
|hak,xi|2=bkfor all 1≤k≤ no
, (8.21)
where ak are the complex conjugates of the rows of A. This is due to the fact that hak,xi = a∗kx. We reformulate this problem. We define for a vectorx ∈ Cnthe matrix X B xx∗, which is of rank one and positive semidefinite. Similarly, define for every k=1, . . . ,mthe matrixAkBaka∗k.
Proposition 8.3.8. Define the mappingAkviaAk(X) =Tr A∗kX
. The set
BBX∈Cn×nAk(X) =b for all1≤k≤n (8.22) is an affine subspace ofCn×n, andBis a lift of M.
Proof. The restrictions can be written as
|hak,xi|2 =Tr(x∗aka∗kx) =Tr(aka∗kxx∗) =Tr(AkX) =Ak(X). (8.23)
90
8.4 Second-Order Subdifferentials at the Set of Rank-Constrained Matrices In other words, for every x ∈ M there existsX = xx∗ ∈ B, and for everyX ∈ B ∩S≤1
there existsx∈Cnsuch thatxx∗ = X.
If A is the discrete Fourier transform, then the above model can be applied to the phase retrieval problem (Chapter 9). The idea of reformulating quadratic constraints as linear equations in the matrix space comes from semidefinite programming. Its appli-cation to phase retrieval goes back to the article by Cand`es et. al in (Cand`es et al., 2011, Section 2.2). We discuss this matter in Section 10.1.