Second Order Freeness and Random Orthogonal Matrices Jamie Mingo (Queen’s University) (joint work with Mihai Popa and Emily Redelmeier) AMS San Diego Meeting, January 11, 2013 1 / 15 Random Matrices 1 √ (x ) d ij I Xd = Xd∗ = I questions: eigenvalues, largest, smallest, gaps, density I problem: {assumptions about distributions of xij } { {conclusions about eigenvalues of Xd } I {xii }i ∪ {xij }i<j independent with mean 0 and E(|xij |2 ) = 1 I {xii }i real random variables identically distributed I {xij }i<j real random variables identically distributed with xij random variables Wigner’s semi-circle law: eigenvalue distribution converges to distribution with density: 0.3 0.2 0.1 -1 0 1 2 2 / 15 Using traces to find the eigenvalue distribution I Xd is self-adjoint and the eigenvalues are λ1 , . . . , λd , we P make a random measure νd = d1 dk=1 δλk I for a function f , R P f dν = 1d dk=1 f (λk ) = d1 Tr(f (Xd )) = tr(f (Xd )) I thus we can study the eigenvalues by studying traces of powers: i.e. moments {tr(Xdk )}k I we only expect to get a simple answer in the large d limit so we want to know for each k, limd tr(Xdk ) I for many examples the limit is not random and can be found by finding limd E(tr(Xdk )), the moments of the limiting eigenvalue distribution I for many ensembles Tr(Xdk − E(tr(Xdk ))I) converges to a random variable and so we have fluctuation moments limd cov(Tr(Xdk ), Tr(Xdl )). 3 / 15 Unitarily Invariant Ensembles I X = √1 (xij ) is unitarily invariant if the joint distribution of d its entries is unchanged when we conjugate X by a unitary matrix; this means that if U is a d × d unitary matrix and Y = UXU∗ then for all i1 , . . . , ik , j1 , . . . , jk we have E(xi1 j1 · · · xik jk ) = E(yi1 j1 · · · yik jk ) examples 1 I if X = X∗ is the gue ensemble: X = X∗ = √ (x ) with d ij {xij }i<j ∪ {xii }i independent Gaussian random variables of mean 0 and (complex) variance 1; I X = 1 G∗ G with G = (gij ) i.i.d. complex Gaussian random d variables of mean 0 and variance 1 (complex Wishart); I X = X∗ is distributed according to the law eTr(V(X)) dX, √ 2 /2 + · · · is a polynomial, x = s + where V(x) = x −1tij ij ij Q Q and dX = i dsii i<j dsij dtij (unitary ensembles in phys.). 4 / 15 Asymptotic Freeness (unitary & orthogonal cases) I let Xd be an ensemble of random matrices and suppose that there is a non-commutative probability space (A, ϕ) and x ∈ A such that for all k, limd E(tr(Xdk )) = ϕ(xk ). Then we say that Xd has a limit distribution thm if Xd and Yd have limit distributions, are independent, and one is unitarily invariant, then Xd and Yd are asymptotically free(1) I ( ) ( ) if limd E(tr(Xd 1 · · · Xd k )) = ϕ(x(1 ) · · · x(k ) ) for all k = 1, 2, 3, . . . and all 1 , 2 , 3 , . . . then we sat that Xd has a limit t-distribution [X(−1) = Xt and x(−1) = xt ] thm if Xd is has a limit t-distribution and is independent from O, a Haar distributed random orthogonal matrix, then {Xd , Xdt } and {O, Ot } are asymptotically free(2) (1) (2) Voiculescu 1991, 1996, & M-Śniady-Speicher 2007 Collins-Śniady 2006 5 / 15 Orthogonal Case I O Haar distributed d × d orthogonal matrix, U Haar distributed d × d unitary matrix, A1 , A2 , A3 , A4 constant matrices I E(Tr(OA1 O−1 A2 ) = d−1 Tr(A1 )Tr(A2 ) I E(Tr(UA1 U−1 A2 ) = d−1 Tr(A1 )Tr(A2 ) I E(Tr(UA1 UA2 )) = 0 I E(Tr(OAOB)) = d−1 Tr(ABt ) = tr(ABt ) I E(Tr(OA1 OA2 OA3 OA4 )) = tr(A1 At4 )tr(A2 At3 ) + d−1 tr(A1 At2 A3 At4 ) − tr(A1 At2 )tr(A3 At4 ) + tr(A1 At4 A3 At2 ) − tr(A1 At4 )tr(A2 At3 ) − d−2 tr(A1 At2 A4 At3 ) + tr(A1 At4 A3 At2 ) + tr(A1 At3 A2 At4 ) + tr(A1 At3 A4 At2 ) 6 / 15 Second Order Probability Spaces (c̄ Nica & Speicher) I I I I I Xd random matrix ensemble with limit distribution x ∈ (A, ϕ) suppose for each m, n limd cov(Tr(Xdm ), Tr(Xdn )) exists then we define ϕ2 (xm , xn ) to be this limit — these are the fluctuation moments of Xd ϕ2 : A ⊗ A → C is a bi-trace with ϕ2 (1, a) = ϕ2 (1, a) = 0 for all a (A, ϕ, ϕ) is a second order probability space fluctuation moments exist for many random matrix models and are described by planar objects 1 8 2 5 7 3 4 6 8 1 2 9 10 7 3 4 6 5 7 / 15 Second Order Freeness (c̄ R. Speicher) I A1 , A2 ⊂ (A, ϕ, ϕ2 ) are second order free if they are free in Voiculescu’s sense and whenever we have centred a1 , . . . , am , b1 , . . . , bn ∈ A with ai ∈ Aki and bj ∈ Alj with k1 , k2 , · · · , km , k1 and l1 , l2 , · · · , ln , l1 then I I for m , n, ϕ2 (a1 · · · am , b1 · · · bn ) = 0 for m = n > 1 (indices of b are mod m) ϕ2 (a1 · · · am , b1 · · · bn ) = m Y m X ϕ(ai bk−i ) k=1 i=1 b2 a3 a1 a1 b1 b1 b2 b3 a2 a3 a1 b1 b3 b2 a2 a3 b3 a2 8 / 15 Real Second Order Freeness (Emily Redelmeier) I (A, ϕ, ϕ2 , t) real second order non-commutative probability space (as before but with addition of the transpose t) I real second order freeness (same as before but also use transposes) ϕ2 (a1 a2 a3 , b1 b2 b3 ) = a1 b1 b2 b1 + b3 a3 b2 a2 a3 b2 a3 a2 a3 b3 a2 a1 bt1 bt3 b1 + b3 a1 + a1 a1 bt1 + bt2 a2 a1 bt3 a3 bt1 + bt2 bt3 a2 a3 bt2 a2 9 / 15 Example: Covariance of O’s and A’s Suppose O is a Haar distributed d × d orthogonal matrix and A1 , . . . , A6 are constant matrices (1 + d−1− 2d−2 )cov(Tr(OA1 O−1 A2 ), Tr(OA3 O−1 A4 )) = d−4 {Tr(A1 )Tr(A2 )Tr(A3 )Tr(A4 ) + Tr(A1 )Tr(A2 )Tr(At3 )Tr(At4 )} − d−3 {Tr(A1 A3 )Tr(A2 )Tr(A4 ) + Tr(A1 At3 )Tr(A2 )Tr(At4 ) + Tr(A1 )Tr(A2 A4 )Tr(A3 ) + Tr(A1 )Tr(A2 At4 )Tr(At3 )} + (d−2 + d−3 ){Tr(A1 A3 )Tr(A2 A4 ) + Tr(A1 At3 )Tr(A2 At4 )} − d−3 {Tr(A1 At3 )Tr(A2 A4 ) + Tr(A1 A3 )Tr(A2 At4 )}. subleading terms can produce non-orientable maps 10 / 15 Main Theorems (c̄ Popa & Redelmeier, arXiv:1210.6079) I {Ad,1 , . . . , Ad,s }d ensemble of random matrices with real second order limiting distribution I Od Haar distributed random orthogonal matrix independent from A’s I thm: {Ad,1 , . . . , Ad,s } and Od are asymptotically real free of second order I thm: independent Haar distributed random orthogonal matrices are asymptotically real free of second order I thm: if {Ai }i and {Bj }j have a real second order limit distribution and are independent and the joint distribution of the entries of A’s is invariant under conjugation by a orthogonal matrix then {Ai } and {Bj } are asymptotically real second order free. 11 / 15 Orthogonal versus unitary if we put together the main theorems of M-Śniady-Speicher with M-Popa-Redelmeier we get; as a unitarily invariant ensemble is orthogonally invariant I if A = {A1 , . . . , Ar } and B = {B1 , . . . , Bs } are independent and A is unitarily invariant then A and B are both asymptotically real second order free and asymptotically (complex) second order free I thus lim E(tr(Ai Btj )) = 0 d 12 / 15 Unitary Invariance and t-distributions: (c̄ M. Popa) let U be a Haar distributed random unitary matrix and U = (U∗ )t be the matrix with ij entry uij ; U and U are Haar distributed random unitary matrices (by the centrality of the Weingarten function) I U = {U, U∗ , Ut , U } has a second order limit t-distribution I U is orthogonally invariant (works because Ot = O−1 ) but not unitarily invariant; e.g. E((U)1,2 (U)1,2 ) = d−1 but E((VUV −1 )1,2 (VUV −1 )1,2 ) = −d−1 where V = diag(i, 1, . . . , 1) is a unitary (but not orthogonal) matrix I thm: if {A1 , . . . , As } has a second order limit distribution and is unitarily invariant then it has a real second order limit distribution 13 / 15 Real and Complex Together Suppose I A1 is unitarily invariant and has a second order limit distribution I A2 is independent form A2 and has a second order limit t-distribution then A1 and A2 are asymptotically real second order free. thm: if U is a Haar distributed random unitary matrix then {U, U∗ } and {Ut , U } are asymptotically real second order free, in particular they are first order free (in the sense of Voiculescu) thm if {A1 , . . . , An } are unitarily invariant and have a second order limit distribution then {A1 , . . . , An } and {At1 , . . . , Atn } are asymptotically second order free. 14 / 15 Weingarten function (Collins & Śniady 2006) I I I I I O = (oij ), d × d Haar distributed orthogonal matrix E(oi1 i−1 oi2 i−2 · · · oin i−n ) = 0 for n odd p, q ∈ P2 (n) (a pair of pairings) hϕ(p), qi = d#(p∨q) , ϕ : C[P2 (n)] → C[P2 (n)] is invertible, Wg = ϕ−1 δipδqδ = 1 only when ir = ip(r) & i−r = i−q(r) , ∀r X E(oi1 i−1 oi2 i−2 · · · oin i−n ) = hWg(p), qiδipδqδ p,q∈P2 (n) I I 1 , 2 , . . . , n ∈ {−1, 1} E(Trγ (O1 A1 , O2 A2 , . . . , On An )) X ηn 1 = hWg(p), qiE(Trπp · q (Aη 1 , . . . , An )) p,q∈P2 (n) I (πp · q , ηp · q ) is the Kreweras complement of the pair of pairings (p, q), πp · q ∈ Sn , η1 , . . . , ηn ∈ {−1, 1} 1 2 3 4 ∨ 1 2 3 4 = 1 2 3 4 15 / 15