Dice Question Streamline Icon: https://streamlinehq.com

Strong convergence for tensor-embedded unitarily invariant random matrices

Determine whether strong convergence in operator norm holds for tensor-embedded families Y_1, ..., Y_L constructed from unitarily invariant d^2×d^2 random matrices X_1, ..., X_L via embeddings Y_j := X_j^{(0j)} ⊗ (I_d)^{⊗([L]\{j})} ∈ M_d^{⊗(L+1)}. Specifically, establish whether Y_1, ..., Y_L converge strongly to a free semicircular system s_1, ..., s_L as d → ∞ in the case X_1 = ··· = X_L, and more generally under mild unitarily invariant assumptions on the family (X_1, ..., X_L).

Information Square Streamline Icon: https://streamlinehq.com

Background

The paper proves asymptotic tensor freeness for families obtained by embedding bipartite matrices into larger tensor product spaces according to bipartite graphs, under unitary invariance and first-order limit assumptions. In the independent case, recent works (CY24, CGVH24) establish strong convergence (operator norm convergence) to a free semicircular system for such embedded families.

In contrast, when all base matrices are identical (X_1 = ··* = X_L), previous work (Lan16) shows the expectation of the operator norm of the sum converges to the corresponding free limit value, but a full strong convergence result is not established. The authors ask whether the strong convergence Y_1, ..., Y_L → s_1, ..., s_L can be proved in this identical case and, more broadly, for unitarily invariant families beyond independence.

References

Can one say anything about the strong convergence (i.e., convergence of operator norms) of these families? For example, suppose each X_j is a (normalized) GUE matrix of size d2 and Y_j:=X_j{(0j)}\otimes (I_d){\otimes [L]\setminus {j}\in M{d}{\otimes(L+1)}. It has recently been shown that Y_1,\ldots, Y_L are asymptotically strongly free as d\to \infty if X_1,\ldots, X_L are independent. On the other hand, considers the case X_1=X_2=\cdots=X_L and shows that as d\to \infty, E[|Y_1+\cdots +Y_L|_{\infty}]\to 2\sqrt{L}= |s_1+\cdots +s_L| where {s_1,\ldots, s_L} is a free semicircular system in a C*-probability space. A natural question is whether the strong convergence Y_1,\ldots, Y_L\to s_1,\ldots, s_L still holds in this case and, more generally, whether it extends to the case when the family {X_1,\ldots, X_L} is unitarily invariant with reasonably nice conditions.

Tensor free probability theory: asymptotic tensor freeness and central limit theorem (2504.01782 - Nechita et al., 2 Apr 2025) in Section 6.2 (Tensor embedding of random matrices), Question