MINIMAX THEOREMS ON HERMITIAN MATRICES MINH KHA There are some interesting inequalities involving eigenvalues of Hermitian matrices. Almost of these I just summarize (I may add some little comments) from the new book of Terence Tao “Topics in random matrix theory” [1]. Minimax theorems A beautiful theorem of Courant and Fischer gives us an alternate definition of eigenvalues of Hermitian matrices: Theorem 0.1. (Courant-Fischer theorem) Let A be a Hermitian n × n matrix. Let λ1 (A) ≥ . . . ≥ λn (A) are all eigenvalues of A. Then λi (A) = supdimV =i inf v∈V,kvk=1 v ∗ Av = inf dimV =n+1−i supv∈V,kvk=1 v ∗ Av. The proof of this is a short and nice application of the spectral theorem of hermitian matrices. There is a generalisation which is called Wielandt minimax formula: Theorem 0.2. For each collection 1 ≤ i1 < i2 < ... < ik ≤ n, we define a partial flag to be a nested collection V1 ⊂ V2 ⊂ ... ⊂ Vk ⊂ Cn such that dimVj = ij . Define the associated Schubert variety X(V1 , ..., Vk ) to be the collection of all k-dimensional subspace W such that dim(W ∩ Vj ) ≥ j. Then λi1 (A) + λi2 (A) + ... + λik (A) = supV1 ,..,Vk inf W ∈X(V1 ,..,Vk ) tr(A|W ) for all n × n hermitian matrix A. By duality, we have λi1 (A) + λi2 (A) + ... + λik (A) = inf V1 ,..,Vk supW ∈X(V1 ,..,Vk ) tr(A|W ) if dimVj = n − ij + 1 and Vk ⊂ ... ⊂ V1 . Pm ∗ n Here the partial trace T r(A|W ) = i=1 vi Avi , where W is a subspace of C spanned by an orthonormal basis v1 , ..., vn . It is not difficult to see this definition is independent of the choice of orthonormal basis of W . Note that this trace operator is defined similarly for trace class operators on a separable Hilbert space: the Schatten S1 class (or noncommutative L1 space). The proof of this one is involved with some combinatorial and dimension counting arguments. We sketch a little bit some ideas used in the proof. Sketch: The key in this proof is to construct a suitable vector space W from Claim 2. Note that since the sequence λi (A) is decreasing, the sum λi1 (A) + ... + λik (A) Pk Pn could be approximated (≥) by a kind of “Riemann sum” j=1 ( i=ij λi (A)γi,j ) ∼ Pk Pn ∗ j=1 ( i=ij γj Aγj ) for some suitable weighted vectors γj = (γij ,j , ..., γn,j ). This suggests us to find a subspace W generated by k orthonormal vectors from subspaces 2000 Mathematics Subject Classification. 20C07, (20E99). Key words and phrases. . 1 2 MINH KHA spanned by {en , ..., eij }, and since we would like to control the sizes of W ∩ Vj are big enough (≥ j) each time j increases, so W should be chosen in a way such that it contains “enough” linearly independent vectors vi ∈ Vi . This is the main idea of the following two claims. Proof. Here are some useful claims which are interesting independently: Claim 1: Given a nest of subspaces Vk ⊂ ... ⊂ V1 such that dimVi ≥ k − i + 1 for every i ∈ N. Suppose that there is an orthonormal family of vectors wi ∈ Vi for any i = 1, ..., k − 1. Let we denote U as the vector subspace spanned by these wi . Then we can always extend and redefine this family in such a way that there exists a vector u ∈ V1 U so that U ⊕ Cu is spanned by an orthonormal basis h1 , ..., hk where each hi ∈ Vi . Proof of claim 1: To prove the first claim, we use induction on k. The initial case k = 2 is clear. Suppose we are at induction step k − 1. Let S = span{w2 , ..., wk−1 }. To “find u”, we should use our induction argument on the nest of k − 1 elements V2 , ..., Vk . Clearly, the restriction on dimensions of these elements Vi is still satisfied in the case k − 1. Thus, we could find a normalize vector v ∈ V2 S such that S ⊕ Cv is spanned by an orthonormal basis h2 , ..., hk where each hi ∈ Vi for any i = 2, ..., k. Now our vector v could be in U or not. In the first case, if v ∈ U then since dimU = dimS + 1, we must have U = S ⊕ Cv or equivalently, U = span{h2 , ..., hk }. Thus, it suffices for us to choose any normalize vector u ∈ V1 U then our claim is proved in this case. This is done since dimV1 ≤ k = dimU + 1. In the second case, we redefine v by its projection on the orthogonal complement ⊥ U of U in V1 . We could see that all hi and w1 must be orthogonal to each other. Since then, we could let u = v and U ⊕ Cv = Cw1 ⊕ span{h2 , ..., hk } is our desire. Claim 2: Given a partial flag V1 ⊂ ... ⊂ Vk with dimVj = ij , and a nest of vector subspaces Wk ⊂ ... ⊂ W1 with dimWj ≥ n − ij + 1. Then we could select two orthonormal families of vectors vi ∈ Vi and wi ∈ Wi such that they span the same vector subspace W = span{v1 , ..., vk } = span{w1 , ..., wk }. Proof of Claim 2 : We use induction on k too. k = 1 is obvious by the estimate dim(A ∩ B) ≥ dim(A) + dim(B) − n for any two subspaces A, B. Assume we are at induction step k − 1. We want to prove for the case k. So by the induction hypothesis, we choose two orthonormal families of vectors vi ∈ Vi and wi ∈ Wi for any i = 1, ..., k − 1 and we denote their span is U . Define Sj = Wj ∩ Vk , j = 1, ..., k − 1. Note that each wj ∈ Sj . Then dimSj ≥ dimWj + dimVk − n ≥ ik − ij + 1 ≥ k − j + 1 and Sk−1 ⊂ ... ⊂ S1 . Applying the first claim, we could find a normalize vector u ∈ S1 U so that U ⊕ Cu is spanned by an orthonormal basis h1 , ..., hk , where hi ∈ Si ⊂ Wi and if we let vk = u then W = U ⊕ Cvk satisfies our requirement. Proof of theorem: Given a hermitian matrix A, we recall e1 , ..., en to be the eigenvectors basis of A such that Aei = λi (A)ei . MINIMAX THEOREMS ON HERMITIAN MATRICES 3 Let Wj = span{en , ..., eij }, then it is clear that the nest Wj satisfies the assumption of Claim 2. We choose two orthonormal families of vectors vi ∈ Vi and wi ∈ Wi such that they span the same vector subspace W = span{v1 , ..., vk } = span{w1 , ..., wk }. Then W ∈ X(V1 , ..., Vk ) and: T r(A|W ) = k X vi∗ Avi = i=1 ≤ k X n X k X wi∗ Awi = i=1 k X n X |e∗l wj |2 e∗l Ael j=1 l=ij |e∗l wj |2 λij (A) = λi1 (A) + ... + λik (A) j=1 l=ij , which proves λi1 (A) + λi2 (A) + ... + λik (A) ≥ supV1 ,..,Vk inf W ∈X(V1 ,..,Vk ) tr(A|W ). The other direction is much easier since if we take Vj = span{e1 , ..., eij }, then for any W ∈ X(V1 , .., Vk ), we can easily choose recursively an orthonormal basis wj for W such that wj ∈ W ∩ Vj for any j = 1, ..., k. Thus, T r(A|W ) = Pk Pij ∗ 2 l=1 |el wj | λl (A) ≥ λi1 (A) + ... + λik (A), which completes the proof of our j=1 theorem. References [1] T. Tao, Topics in random matrix theory, Graduate Studies in Mathematics, vol. 132, American Mathematical Society, Providence, RI, 2012. M.K., Department of Mathematics, Texas A&M University, College Station, TX 77843-3368, USA E-mail address: kha@math.tamu.edu