Failure of Strong Convergence of Matrices with Fermionic Entries
Dimitri Shlyakhtenko · 2026
Read Failure of Strong Convergence of Matrices with Fermionic Entries by Dimitri Shlyakhtenko (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-28648.txt.
Introduction
Dimitri Shlyakhtenko, writing alone, exhibits a failure of a convergence that random matrix theory had led many to expect. Strong convergence says random matrices should mimic their limiting free-probability objects not just in average spectrum but in the norms of every polynomial, the strongest and most useful sense.
His matrices have fermionic entries, operators obeying the canonical anticommutation relations, scaled to variance one over N. Individually, each matrix's real part does converge in operator norm to the semicircular limit, matching the classical picture. Jointly, the family refuses: even the operator space layout of the spans of the real and imaginary parts fails to approach the semicircular limit. Fermionic randomness, whatever its individual good behavior, does not assemble into the free limit that bosonic and classical constructions achieve.
Catalog record
| CREATOR | Dimitri Shlyakhtenko |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.28648
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Failure of Strong Convergence of Matrices with Fermionic Entries Dimitri Shlyakhtenko Abstract Let Q N ( k ) Q^{(k)}_{N} be an N × N N\times N matrices with entries satisfying CAR, normalized to have variance 1 / N 1/\sqrt{N} with respect to the trace of the CAR algebra. We show that, although the operator norm of the real part of an individual matrix Q N ( k ) Q^{(k)}_{N} converges as N → ∞ N\to\infty to the semicircular limit, the family of matrices does not converge to the free probability limit strongly. In fact, even the operator space structure of the linear spans of the real and imaginary parts of Q N ( k ) Q^{(k)}_{N} ’s, k = 1 , … , M k=1,\dots,M , does not converge to the semicircular limit. 1 Introduction. In this paper, we consider the asymptotics of N\times N block matrices A^{(k)}_{N} , 1\leq k\leq M with entries given by generators satisfing Canonical Anti-Commutation Relations (CAR) (see § 1.2 for the precise definitions and notations). These matrices are fermionic analogs of GUE matrices, in the sense that the Gaussian random entries of GUE matrices are replaced by “iid” fermionic variables. It has been known, since the work of Voiculescu, that, under a natural trace, such “CAR-UE” matrices converge in moments to a free semicircular family S_{1},\dots,S_{M} (see [ 15 ] ). However, two questions were left open after Voiculescu’s work. The first is the behavior of the operator norm of a single matrix A_{N}=A_{N}^{(1)} . The second is the question of strong convergence: whether it is true that operator norms of arbitrary polynomials in the matrices A_{N}^{(1)},A_{N}^{(2)},\dots converge to the operator norm of the same polynomial but evaluated in the free semicircular family S_{1},S_{2},\dots . This question was communicated to the author by Gilles Pisier, and it came up, for example, in the work [ 9 ] where the matrices A_{N}^{(k)} were used as one of the ways to prove the QWEP property for C^{*} -algebras generated by generalized circular elements. Our main result is that for all N , \|A_{N}\|=\sqrt{4-3/N}<2 (in fact, we give a precise description of the spectrum of A_{N} , see Theorem 2 ); moreover, \|A_{N}\|\to 2 , which is the operator norm of the semicircular limit. However, we show that strong convergence fails if we consider simultaneously 12 or more such matrices. More precisely, if A_{N}^{(1)},B_{N}^{(2)},\dots,A_{N}^{(M)},B_{N}^{(M)} are “CAR-iid” copies of A_{N}^{(1)} , and we set Q_{N}^{(j)}=A_{N}^{(j)}+iB_{N}^{(j)} , then Q_{N}^{(j)} converge in * -moments to a family of freely independent circular elements Q^{(1)},\dots,Q^{(M)} , but \limsup_{N\to\infty}\Big\|\sum_{j=1}^{M}Q_{N}^{(j)}(Q_{N}^{(j)})^{*}\Big\|\gneq\Big\|\sum_{j=1}^{M}Q^{(j)}(Q^{(j)})^{*}\Big\|. (see Theorem 3 below). The same example shows that the operator space structure of the span of A_{N}^{(1)},B_{N}^{(2)},\dots,A_{N}^{(M)},B_{N}^{(M)} does not converge, as N\to\infty to the free probability limit. Thus the asymptotics of the “CAR-UE” matrices gives another example of a real Hilbertian operator space [ 10 ] , which is not isomorphic to the complex span of semicircular variables, and whose precise operator structure we cannot elucidate. Convergence to the semicircular limit does hold for real-linear combinations, but we are unable to prove the complex case (although we suspect that the statement is true). We summarize these results in Table 1 . Space Comparison to Free Isomorphism Class Banach space W_{\infty}^{\mathbb{R}} same \ell_{\mathbb{R}}^{2}(2M) Banach space W_{\infty}^{\mathbb{C}} unknown, same? conjecturally \ell^{2}_{\mathbb{C}}(2M) Operator space W_{\infty}^{\mathbb{C}} different unknown C^{*} -algebra C^{*}(W_{\infty}^{\mathbb{C}}) different unknown Table 1: Here W^{\mathbb{C}}_{\infty} and W^{\mathbb{R}}_{\infty} are the complex or real span of the family (A^{(j)}_{N},B^{(j)}_{N},1\leq j\leq M) in the ultraproduct of W^{*}(A^{(j)}_{N},B^{(j)}_{N}:1\leq j\leq M). There is another obstruction to strong convergence, where a different degree-two polynomial involving commutators of the “CAR-UE” matrices is used. This computation is given in Theorem 4 . Based on this norm computation, we give evidence in § 3.4 that there is not a “a single-vector obstruction to norm convergence” as is the case for permutation matrices, and showing that it does not seem possible to “restore” strong convergence by restricting to a “large enough” part of the underlying Hilbert space. To conclude the paper, we discuss a variant of the “CAR-UE” matrices involving q -creation operators [ 1 ] for -1<q<1 . We show that in those cases, strong convergence holds true, in fact not just for the “ q -UE” matrices, but for matrices with creation operators as entries (and free creation operators as limits). 1.1 AI use: How the results were obtained This paper also represents an experiment in the use of AI for research in operator algebras. The counterexample to strong convergence was found by utilizing software written by ChatGPT 5.4 Pro, and ChatGPT was able to find an expression for the special vector \eta_{0} used to give a lower bound on the norm of sums of commutators in Lemma 11 . However, it did not realize that this gives a counterexample to strong convergence. ChatGPT, however, showed remarkable performance on the question of convergence of A_{N} (though when asked directly it hallucinated incorrect operator-space inequalities). It hallucinated the combinatorial identity ( 1 ) discussed in § 2.5 for the moments of the matrices A_{N} . When explained that this identity does exist in the paper [ 5 ] and asked to give a direct proof, it retraced the steps back to the matrix A_{N} , found the representation theory proof presented in this paper, and derived the identity as a consequence. ChatGPT 5.5 Pro was also helpful in the proof of the q -deformed version. Overall, it is clear that it is a useful and powerful tool for mathematical research, even in a highly theoretical field such as the field of this paper. A partial formalization of the results of this paper in Lean is available at https://github.com/shlyakhtenko/car-matrices . The formalization covers §2, including the underlying representation-theory tools, as well as Lemma 11 , down to theorems and definitions in Mathlib v4.32.0-rc1 with no additional axioms or hypotheses. We intend to pursue formalization of the free probability norm estimates necessary to see the violation of norm convergence later on. 1.2 Setup and Notation For fixed integers M and N , consider the real Hilbert space H=\bigoplus_{k=1}^{M}\bigl(H_{k}^{(1)}\oplus H_{k}^{(2)}\bigr), where for each k=1,2,\dots,M and r=1,2 , H_{k}^{(r)} is a real Hilbert space of dimension N^{2} with basis h_{pq}^{(r,k)} , 1\leq p,q\leq N . Let \mathscr{F} denote the associated antisymmetric Fock space (see e.g. [ 2 ] ), and let \Omega be the vacuum vector. Denote by \mathbb{B} the space of (bounded) operators on the finite-dimensional space \mathscr{F} . For h\in H\otimes_{\mathbb{R}}\mathbb{C} , let a(h)^{*}\in\mathbb{B} be the creation operator given on elementary antisymmetric tensors by a(h)^{*}(h_{1}\wedge\dots\wedge h_{k})=h\wedge h_{1}\wedge\dots\wedge h_{k} , and let a(h) be its adjoint. These operators satisfy the Canonical Anti-Commutation Relations: for h,g\in H\otimes_{\mathbb{R}}\mathbb{C} , a(h)a(g)+a(g)a(h)=0,\ a(h)a(g)^{*}+a(g)^{*}a(h)=\langle h,g\rangle. For h\in H , let b(h)=a(h)+a(h)^{*} be the associated Majorana operator. Let e_{pq}\in M_{N\times N} , 1\leq p,q\leq N , be elementary N\times N matrices, and consider the operators Q^{(k)}_{N}=\frac{1}{\sqrt{N}}\sum_{p,q=1}^{N}e_{pq}\otimes\bigl(b(h^{(1,k)}_{pq})+i\,b(h^{(2,k)}_{pq})\bigr)\in M_{N\times N}\otimes\mathbb{B}. Let A^{(k)}_{N}=\frac{Q^{(k)}_{N}+(Q^{(k)}_{N})^{*}}{2},\qquad B^{(k)}_{N}=\frac{Q^{(k)}_{N}-(Q^{(k)}_{N})^{*}}{2i} be the real and imaginary parts of Q_{N}^{(k)} . Denote by \phi:\mathbb{B}\to\mathbb{C} the vector state x\mapsto\langle\Omega,x\Omega\rangle and by \phi_{N}:M_{N\times N}\otimes\mathbb{B}\to\mathbb{C} the state \phi_{N}=N^{-1}Tr\otimes\phi . By [ 15 ] , the family of matrices (A_{N}^{(1)},B_{N}^{(1)},\dots,A_{N}^{(M)},B_{N}^{(M)}) converges in distribution to a free semicircular family. 2 Norm of real linear combinations. In this section we