Introduction
We study the complete continuous first-order theories of the matrix algebras M n ( C ) M_n(\mathbb{C}) M n ( C ) , equipped with the normalized trace τ n = n − 1 Tr \tau_n = n^{-1}\operatorname{Tr} τ n = n − 1 Tr and the metric induced by ∥ x ∥ 2 = τ n ( x ∗ x ) 1 / 2 \lVert x\rVert_2 = \tau_n(x^*x)^{1/2} ∥ x ∥ 2 = τ n ( x ∗ x ) 1/2 on operator-norm bounded balls. A sentence has a real value in each algebra, and two tracial algebras are elementarily equivalent if all these values agree. Farah, Hart and Sherman asked whether every sentence value converges as n → ∞ n \to\infty n → ∞ [12 Question 6.6] . To express this question in terms of ultraproducts, we use Łoś’ theorem:
φ ∏ U M n ( C ) = lim U φ M n ( C ) , \varphi\prod_{\mathcal{U}} M_n(\mathbb{C}) = \lim_{\mathcal{U}} \varphi^{M_n(\mathbb{C})}, φ U ∏ M n ( C ) = U lim φ M n ( C ) ,
so convergence is equivalent to elementary equivalence of all nonprincipal tracial ultraproducts of this sequence. Our first result gives a negative answer.
Theorem A. There are 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 pairwise non-elementarily equivalent tracial matrix ultraproducts of the sequence ( M n ( C ) , τ n ) (M_n(\mathbb{C}),\tau_n) ( M n ( C ) , τ n ) , and the same holds for metric ultraproducts of ( U ( n ) , d 2 ) (U(n),d_2) ( U ( n ) , d 2 ) in the pure metric group language, where d 2 ( u , v ) = ∥ u − v ∥ 2 d_2(u,v)=\lVert u-v\rVert_2 d 2 ( u , v ) = ∥ u − v ∥ 2 . Consequently, the sequences of complete continuous theories
Th ( M n ( C ) , τ n ) , Th ( U ( n ) , d 2 ) \operatorname{Th}(M_n(\mathbb{C}),\tau_n), \qquad\operatorname{Th}(U(n),d_2) Th ( M n ( C ) , τ n ) , Th ( U ( n ) , d 2 ) do not converge as n → ∞ n \to\infty n → ∞ .
Purely universal and purely existential sentences do converge, since every tracial matrix ultraproduct has the universal and existential theories of the hyperfinite I I 1 \mathrm{II}_1 II 1 factor [13 Remark 6.5(1), equation (6.2)] .
To distinguish the theories, we recover the diagonal and use its unitary normalizer. The quotient by the diagonal unitaries is a universal sofic group, so uniform recovery of this quotient allows us to apply the nonconvergence result of [3 ] to matrix algebras. We first fix the class of ultraproducts and the notion of internality used in this recovery.
Definition 1.1. Let U \mathcal{U} U be a nonprincipal ultrafilter on N \mathbb{N} N , and let d = ( d n ) \mathbf{d}=(d_n) d = ( d n ) be a sequence of positive integers with d n → ∞ d_n \to\infty d n → ∞ . We call
M U , d = ∏ U ( M d n ( C ) , τ d n ) (1) \mathcal{M}_{\mathcal{U},\mathbf{d}}=\prod_{\mathcal{U}}(M_{d_n}(\mathbb{C}),\tau_{d_n}) \tag*{(1)} M U , d = U ∏ ( M d n ( C ) , τ d n ) (1) a tracial matrix ultraproduct . Its elements are classes [ x n ] U [x_n]_{\mathcal{U}} [ x n ] U of operator-norm bounded sequences, modulo sequences whose 2 2 2 -norms tend to zero along U \mathcal{U} U . The operations are induced coordinatewise, and τ ( [ x n ] U ) = lim U τ d n ( x n ) \tau([x_n]_{\mathcal{U}})=\lim_{\mathcal{U}}\tau_{d_n}(x_n) τ ([ x n ] U ) = lim U τ d n ( x n ) . A unital von Neumann subalgebra A ⊆ M U , d A\subseteq\mathcal{M}_{\mathcal{U},\mathbf{d}} A ⊆ M U , d is internal if A = [ A n ] U A=[A_n]_{\mathcal{U}} A = [ A n ] U for unital ∗ * ∗ -subalgebras A n ⊆ M d n ( C ) A_n\subseteq M_{d_n}(\mathbb{C}) A n ⊆ M d n ( C ) . An internal MASA is an internal abelian subalgebra that is maximal among internal abelian subalgebras.
Throughout, M U , d \mathcal{M}_{\mathcal{U},\mathbf{d}} M U , d and M V , e \mathcal{M}_{\mathcal{V},\mathbf{e}} M V , e denote tracial matrix ultraproducts in this sense. To recognize the diagonal, we will use Lemma 2.1 : internal MASAs are maximal abelian in the ambient algebra and are unitarily conjugate to the canonical diagonal [ D d n ] U [D_{d_n}]_{\mathcal{U}} [ D d n ] U , where D d ⊆ M d ( C ) D_d\subseteq M_d(\mathbb{C}) D d ⊆ M d ( C ) denotes the diagonal algebra. Isomorphisms are complex ∗ * ∗ -isomorphisms; they preserve the trace because these ultraproducts are I I 1 \mathrm{II}_1 II 1 factors.
To use this recovery for elementary equivalence, we work in the continuous language of tracial ∗ * ∗ -algebras on operator-norm bounded sorts. A predicate is definable if it is a uniform limit of formulas, and a nonempty closed set is definable if its distance predicate is definable. We allow parameters when specified, use bounded real ranges, and write s − ˙ t = max { s − t , 0 } s\mathbin{\dot{-}}t=\max\{s-t,0\} s − ˙ t = max { s − t , 0 } . See [10 Section 3.2] for quantification over definable sets. The following theorem gives a uniform way to recover an internal diagonal from finitely many parameters.
Theorem B. There are fixed finite formulas Δ ( z ) \Delta(z) Δ ( z ) and δ ( x ; z ) \delta(x;z) δ ( x ; z ) , with z z z a tuple of six contractions, such that in every tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},\mathbf{d}} M U , d : (1) the zero-set of Δ \Delta Δ is nonempty; (2) every z z z in this zero-set defines an internal MASA D z D_z D z with
δ ( x ; z ) = dist 2 ( x , ( D z ) ≤ 1 ) ( x ∈ ( M U , d ) ≤ 1 ) . \delta(x;z)=\operatorname{dist}_2(x,(D_z)_{\leq1}) \qquad(x\in(\mathcal{M}_{\mathcal{U},\mathbf{d}})_{\leq1}). δ ( x ; z ) = dist 2 ( x , ( D z ) ≤ 1 ) ( x ∈ ( M U , d ) ≤ 1 ) . Every such MASA is unitarily conjugate to the canonical diagonal.
The diagonal cannot be the commutant of a finite tuple [30 Theorem 2.1(1)] , so we recover it as the relative join of two abelian commutants with spectral gap. The spectral gap internality theorem of Liu and the authors [ALT26, Theorem B] makes these commutants internal. Uniform abelian Hilbert–Schmidt stability then ensures that their relative join is internal as well.
The finite code and conjugacy of internal MASAs will allow us to recover the theory of the pair ( M U , d , D ) (\mathcal{M}_{\mathcal{U},\mathbf{d}},D) ( M U , d , D ) from that of the ambient algebra. Here the pair is the expansion by the distance to the operator-norm unit ball of D D D . We obtain the following reduction.
Theorem C. Let D = [ D d n ] U ⊆ M U , d D=[D_{d_n}]_{\mathcal{U}}\subseteq\mathcal{M}_{\mathcal{U},\mathbf{d}} D = [ D d n ] U ⊆ M U , d and F = [ D e n ] V ⊆ M V , e F=[D_{e_n}]_{\mathcal{V}}\subseteq\mathcal{M}_{\mathcal{V},\mathbf{e}} F = [ D e n ] V ⊆ M V , e be the canonical diagonal subalgebras of two tracial matrix ultraproducts. The following hold. (1) Every isomorphism θ : M U , d → M V , e \theta:\mathcal{M}_{\mathcal{U},\mathbf{d}}\to\mathcal{M}_{\mathcal{V},\mathbf{e}} θ : M U , d → M V , e sends D D D to a unitarily conjugate of F F F . (2) If M U , d ≡ M V , e \mathcal{M}_{\mathcal{U},\mathbf{d}}\equiv\mathcal{M}_{\mathcal{V},\mathbf{e}} M U , d ≡ M V , e , then ( M U , d , D ) ≡ ( M V , e , F ) (\mathcal{M}_{\mathcal{U},\mathbf{d}},D)\equiv(\mathcal{M}_{\mathcal{V},\mathbf{e}},F) ( M U , d , D ) ≡ ( M V , e , F ) . (3) If M U , d ≡ M V , e \mathcal{M}_{\mathcal{U},\mathbf{d}}\equiv\mathcal{M}_{\mathcal{V},\mathbf{e}} M U , d ≡ M V , e , then
∏ U m e t Sym ( d n ) ≡ ∏ V m e t Sym ( e n ) , \prod_{\mathcal{U}}^{\mathrm{met}}\operatorname{Sym}(d_n)\equiv\prod_{\mathcal{V}}^{\mathrm{met}}\operatorname{Sym}(e_n), U ∏ met Sym ( d n ) ≡ V ∏ met Sym ( e n ) , where the symmetric groups carry the normalized Hamming metric.
To obtain the last assertion, permutation matrices identify the Weyl group with the metric ultraproduct of symmetric groups. Its Hamming metric is one half the square of the quotient 2 2 2 -metric. Definability lets us make this identification under elementary equivalence, and [3 Theorem 5.6 and Remark 5.7] then gives the matrix-algebra assertion of Theorem A. Reconstruction from unitary groups gives the other assertion. The argument that preserves the diagonal also extends to arbitrary internal algebras, giving the following structural conclusion.
Theorem D. Let θ : M U , d → M V , e \theta: \mathcal{M}_{\mathcal{U},d} \to\mathcal{M}_{\mathcal{V},e} θ : M U , d → M V , e be an isomorphism of tracial matrix ultraproducts. Then for every unital von Neumann subalgebra A ⊆ M U , d A \subseteq\mathcal{M}_{\mathcal{U},d} A ⊆ M U , d ,
A is internal ⟺ θ ( A ) is internal . A\text{ is internal}\quad\Longleftrightarrow\quad\theta(A)\text{ is internal}. A is internal ⟺ θ ( A ) is internal .
The constructions used here also give fixed finite distance formulas over four unitary parameters for internal abelian algebras and six for arbitrary internal algebras; see Proposition 5.2. They also show that, for commuting internal algebras A A A , B B B , both A ′ ∩ B ′ A' \cap B' A ′ ∩ B ′ and their relative join ( A ′ ∩ B ′ ) ′ (A' \cap B')' ( A ′ ∩ B ′ ) ′ are internal (Corollary 4.4 ).
We next use the recovered permutation groups to distinguish an explicit family of factors from their amplifications. For a I I 1 \mathrm{II}_{1} II 1 factor Q Q Q , let Q t = p M k ( Q ) p Q^{t}=pM_{k}(Q)p Q t = p M k ( Q ) p , with normalized trace, where k > t k>t k > t and ( τ k ⊗ τ ) ( p ) = t / k (\tau_{k}\otimes\tau)(p)=t/k ( τ k ⊗ τ ) ( p ) = t / k . Its isomorphism class depends only on t > 0 t>0 t > 0 . The first-order fundamental group
F f o ( Q ) = { t > 0 : Q t ≡ Q } \mathcal{F}_{\mathrm{fo}}(Q)=\{t>0:Q^{t}\equiv Q\} F fo ( Q ) = { t > 0 : Q t ≡ Q }
contains F ( Q ) = { t > 0 : Q t ≅ Q } \mathcal{F}(Q)=\{t>0:Q^{t}\cong Q\} F ( Q ) = { t > 0 : Q t ≅ Q } ; see [17 Section 4] . For the construction, put G r = PSL 2 ( 2 r ) G_{r}=\operatorname{PSL}_{2}(2^{r}) G r = PSL 2 ( 2 r ) , for r ≥ 2 r\geq2 r ≥ 2 , so that ∣ G r ∣ = 2 r ( 2 2 r − 1 ) |G_{r}|=2^{r}(2^{2r}-1) ∣ G r ∣ = 2 r ( 2 2 r − 1 ) .
Theorem E. Fix a nonprincipal ultrafilter U \mathcal{U} U on N \mathbb{N} N . For each set S S S of odd primes, put
r n ( S ) = 2 n ∏ p ∈ S p ≤ n p , Q S = ∏ U ( M ∣ G r n ( S ) ∣ ( C ) , τ ∣ G r n ( S ) ∣ ) . r_{n}(S)=2^{n}\prod_{\substack{p\in S\\p\leq n}}p,\qquad
Q_{S}=\prod_{\mathcal{U}}\left(M_{|G_{r_{n}(S)}|}(\mathbb{C}),\tau_{|G_{r_{n}(S)}|}\right). r n ( S ) = 2 n p ∈ S p ≤ n ∏ p , Q S = U ∏ ( M ∣ G r n ( S ) ∣ ( C ) , τ ∣ G r n ( S ) ∣ ) . Then F f o ( Q S ) = F ( Q S ) = { 1 } \mathcal{F}_{\mathrm{fo}}(Q_{S})=\mathcal{F}(Q_{S})=\{1\} F fo ( Q S ) = F ( Q S ) = { 1 } . Distinct sets S S S give non-elementarily equivalent factors.
To prove this theorem, we follow the left-right action method of [3 Section 5] and recover the regular actions of PSL 2 ( 2 r ) \operatorname{PSL}_{2}(2^{r}) PSL 2 ( 2 r ) from their elementary properties. The recovered actions force the ambient dimension to be asymptotic to a group order. We then use element orders to detect divisibility of r r r , which rules out nontrivial amplifications.
Popa asked in 2008 whether there are uncountably many nonisomorphic tracial matrix ultraproducts [11 Introduction] . For the related ultrapower problem, Ge and Hadwin proved that, under CH, all tracial ultrapowers of a fixed finite von Neumann algebra with separable predual are isomorphic [15 Theorem 3.2] . Li later suggested that matrix ultraproducts might be independent of the ultrafilter [27 p. 1287] . Under failure of CH, Farah, Hart and Sherman obtained nonisomorphic matrix ultraproducts [11 Proposition 3.3] , and Farah and Shelah obtained the maximal number 2 c 2^{\mathfrak{c}} 2 c [14 Proposition 8.3] . Under CH, matrix ultraproducts are isomorphic exactly when they are elementarily equivalent. Theorem A distinguishes their theories in ZFC and hence answers Popa’s question and gives continuum many classes in the classification problem of [25 Problem 1] . For the model-theoretic framework and the game formulation, see [12 , 13 ] and [9 Question 4.1] .
For universal sofic groups, Thomas obtained the maximal number of isomorphism types under failure of CH and realized algebraic ultraproducts of finite expander groups as centralizers [33 Theorems 1.1 and 2.2] . The authors obtained continuum many theories in ZFC in [3 ] and proposed the diagonal-normalizer route to matrix ultraproducts [3 Remark 5.8] . Jekel discussed this route in [22 Section 1.1] . Since full types include sentence values, nonconvergence also obstructs convergence of full types of random matrix tuples. It answers the corresponding ultrafilter-independence question for noncommutative stochastic processes as well; see [21 Question 5.6] , [22 Section 1.1] , and [23 Remark 1.3 and Proposition D(1)] .
Our internality arguments use Popa’s coordinate commutant identities [30 Section 1.6] and conjugacy of coordinate MASAs [29 Section 2.1(a)] . The diagonal is a MASA, but its normalizer generates a proper subalgebra: tracial matrix ultraproducts have no Cartan subalgebras [29 Proposition 2.3] ; see also [28 Theorem 7.3] for ultrapowers of factors.
While preparing this manuscript, we learned of the independent work of Ando and Goldbring [4 ] . They prove that every tracial matrix ultraproduct has trivial first-order fundamental group, which also implies nonconvergence.
The paper is organized as follows: Section 2 supplies the internality results needed to recognize the diagonal. Using these, Section 3 recovers the diagonal and its normalizer under isomorphism, and Section 4 extends preservation to all internal algebras. To pass to elementary equivalence, we make the constructions definable in Section 5. Section 6 then recovers the left-right actions from admissible tuples and uses element orders to prove Theorem E.
Spectral gap and internal algebras
We use the normalized trace τ d = d − 1 Tr \tau_d = d^{-1}\operatorname{Tr} τ d = d − 1 Tr on M d = M d ( C ) M_d = M_d(\mathbb{C}) M d = M d ( C ) and its norm ∥ x ∥ 2 , d = τ d ( x ∗ x ) 1 / 2 \lVert x\rVert_{2,d} = \tau_d(x^*x)^{1/2} ∥ x ∥ 2 , d = τ d ( x ∗ x ) 1/2 . We write A ≤ 1 A_{\leq1} A ≤ 1 for the operator-norm unit ball of an algebra A A A . Bounded functional calculus lets us choose contractive coordinate representatives for every unit-ball element of an internal algebra.
We use primes for commutants in the specified ambient matrix ultraproduct. For a finite tuple u \mathbf{u} u , we write { u } ′ ∩ M U , d \{\mathbf{u}\}' \cap\mathcal{M}_{U,d} { u } ′ ∩ M U , d for its commutant, and we define the relative join of two subalgebras by
A ∨ r e l B = ( A ′ ∩ B ′ ) ′ . (2) A \mathbin{\vee_{\mathrm{rel}}} B = (A' \cap B')'. \tag*{(2)} A ∨ rel B = ( A ′ ∩ B ′ ) ′ . (2)
Thus the relative join is the smallest algebra containing A A A and B B B that equals its relative bicommutant.
Internal subalgebras, commutants and spectral gap
The following identities let us pass between coordinate models and commutants in the ultraproduct. They also reduce unitary conjugacy of internal MASAs to the matrix case.
Lemma 2.1. Let M U , d \mathcal{M}_{U,d} M U , d be a tracial matrix ultraproduct, and let A n ⊆ M d n ( C ) A_n \subseteq M_{d_n}(\mathbb{C}) A n ⊆ M d n ( C ) be unital ∗ * ∗ -subalgebras. Put A = [ A n ] U A = [A_n]_U A = [ A n ] U . Then
A ′ = [ A n ′ ∩ M d n ] U , Z ( A ) = [ Z ( A n ) ] U , A ′ ′ = A . A' = [A_n' \cap M_{d_n}]_U,\qquad Z(A) = [Z(A_n)]_U,\qquad A'' = A. A ′ = [ A n ′ ∩ M d n ] U , Z ( A ) = [ Z ( A n ) ] U , A ′′ = A . The coordinate conditional expectations E A n : M d n ( C ) → A n E_{A_n}: M_{d_n}(\mathbb{C}) \to A_n E A n : M d n ( C ) → A n induce the conditional expectation E A : M U , d → A E_A: \mathcal{M}_{U,d} \to A E A : M U , d → A . Every internal abelian algebra has abelian coordinate models and is contained in an internal MASA. Every internal MASA is a MASA in the ambient ultraproduct and is unitarily conjugate to the canonical diagonal [ D d n ] U [D_{d_n}]_U [ D d n ] U .
Proof. We first identify the conditional expectation induced by the coordinate maps. Since each E A n E_{A_n} E A n is a contraction in both the operator norm and the 2-norm, these maps induce a trace-preserving projection onto [ A n ] U [A_n]_U [ A n ] U . The trace-pairing characterization of conditional expectation then identifies this projection with E A E_A E A .
The commutant identity A ′ = [ A n ′ ∩ M d n ] U A' = [A_n' \cap M_{d_n}]_U A ′ = [ A n ′ ∩ M d n ] U follows from [30 Section 1.6] . To obtain the center identity, we apply the same result with ambient coordinate algebras A n A_n A n , which gives Z ( A ) = [ Z ( A n ) ] U Z(A) = [Z(A_n)]_U Z ( A ) = [ Z ( A n ) ] U . Applying the commutant identity twice in M d n M_{d_n} M d n and using the finite-dimensional bicommutant theorem then gives A ′ ′ = A A'' = A A ′′ = A .
Suppose now that A A A is abelian. The center identity allows us to replace its coordinate models by Z ( A n ) Z(A_n) Z ( A n ) . We can therefore extend each model to a MASA D n 0 ⊆ M d n D_n^0 \subseteq M_{d_n} D n 0 ⊆ M d n , and the commutant identity shows that [ D n 0 ] U [D_n^0]_U [ D n 0 ] U is a MASA containing A A A . If A A A is maximal among internal abelian algebras, this inclusion forces A = [ D n 0 ] U A = [D_n^0]_U A = [ D n 0 ] U , so it is also maximal abelian in the ambient ultraproduct. Conversely, an internal algebra that is maximal abelian is maximal among internal abelian algebras. Finally, coordinate unitary conjugacies identify each of these MASAs with the canonical diagonal [ D d n ] U [D_{d_n}]_U [ D d n ] U [29 Section 2.1(a)] . Its maximal abelianness can also be read directly from the commutant identity:
[ D d n ] U ′ ∩ M U , d = [ D d n ′ ∩ M d n ] U = [ D d n ] U . [D_{d_n}]_U' \cap\mathcal{M}_{U,d} = [D_{d_n}' \cap M_{d_n}]_U = [D_{d_n}]_U. [ D d n ] U ′ ∩ M U , d = [ D d n ′ ∩ M d n ] U = [ D d n ] U .
□ \square □
To recognize an algebra given as a commutant, we will use spectral gap. Let P \mathcal{P} P be a tracial von Neumann algebra and let u = ( u 1 , … , u h ) ∈ U ( P ) h \mathbf{u} = (u_1,\ldots,u_h) \in U(\mathcal{P})^h u = ( u 1 , … , u h ) ∈ U ( P ) h , with h ≥ 1 h \geq1 h ≥ 1 . Put C = P ∩ { u 1 , … , u h } ′ \mathcal{C} = \mathcal{P} \cap\{u_1,\ldots,u_h\}' C = P ∩ { u 1 , … , u h } ′ . We say that u \mathbf{u} u has spectral gap in P \mathcal{P} P if there is 0 < L < ∞ 0 < L < \infty 0 < L < ∞ such that
∥ x − E C x ∥ 2 2 ≤ L ∑ j = 1 h ∥ [ u j , x ] ∥ 2 2 ( x ∈ P ) . (3) \lVert x - E_{\mathcal{C}}x\rVert_2^2 \leq L \sum_{j=1}^{h} \lVert[u_j,x]\rVert_2^2 \qquad(x \in\mathcal{P}). \tag*{(3)} ∥ x − E C x ∥ 2 2 ≤ L j = 1 ∑ h ∥[ u j , x ] ∥ 2 2 ( x ∈ P ) . (3)
We call any such L L L a spectral gap constant for the tuple. The distance in this estimate is measured to the commutant within the specified ambient algebra.
Our main input, the spectral gap internality theorem [ALT26, Theorem B], turns this estimate into coordinate models for the commutant. We use its relative form so that we can also work inside an internal subalgebra.
Theorem 2.2. Let P = [ P n ] U P=[P_n]_U P = [ P n ] U be an internal subalgebra of a tracial matrix ultraproduct M U , d \mathcal{M}_{U,\mathbf{d}} M U , d , and let u 1 , … , u h ∈ U ( P ) u_1,\ldots,u_h\in\mathcal{U}(P) u 1 , … , u h ∈ U ( P ) have spectral gap in P P P . Then their relative commutant C = P ∩ { u 1 , … , u h } ′ C=P\cap\{u_1,\ldots,u_h\}' C = P ∩ { u 1 , … , u h } ′ has the form C = [ C n ] U C=[C_n]_U C = [ C n ] U for unital ∗ * ∗ -subalgebras C n ⊆ P n C_n\subseteq P_n C n ⊆ P n .
Proof. We will approximate the conditional expectation onto the relative commutant by expectations on the coordinate summands, where the matrix version of the theorem applies. Let L L L be a spectral gap constant, and decompose the coordinate algebras with their inherited traces as
( P n , τ d n ∣ P n ) = ⨁ s ( M a n , s , w n , s τ a n , s ) , w n , s > 0 , ∑ s w n , s = 1. (P_n,\tau_{d_n}|_{P_n})=\bigoplus_s(M_{a_{n,s}},w_{n,s}\tau_{a_{n,s}}),\qquad w_{n,s}>0,\qquad\sum_s w_{n,s}=1. ( P n , τ d n ∣ P n ) = s ⨁ ( M a n , s , w n , s τ a n , s ) , w n , s > 0 , s ∑ w n , s = 1.
We choose unitary lifts u j , n ∈ P n u_{j,n}\in P_n u j , n ∈ P n and form their lazy conjugation averages
T n ( x ) = 1 2 x + 1 4 h ∑ j = 1 h ( u j , n x u j , n ∗ + u j , n ∗ x u j , n ) . T_n(x)=\frac{1}{2}x+\frac{1}{4h}\sum_{j=1}^{h}(u_{j,n}xu_{j,n}^*+u_{j,n}^*xu_{j,n}). T n ( x ) = 2 1 x + 4 h 1 j = 1 ∑ h ( u j , n x u j , n ∗ + u j , n ∗ x u j , n ) .
The diagonal argument of [ALT26, Proposition 2.2] applies to finite-dimensional tracial algebras. It therefore gives powers F n = T n k n F_n=T_n^{k_n} F n = T n k n , with k n → U ∞ k_n\to_U\infty k n → U ∞ , that induce E C E_C E C . Since these maps preserve every summand, we may work with their restrictions T n , s , F n , s T_{n,s},F_{n,s} T n , s , F n , s . We use the corresponding restrictions for the unitary entries as well.
To find suitable expectations on the summands, fix ε > 0 \varepsilon>0 ε > 0 and call a summand bad if F n , s F_{n,s} F n , s has ∞ \infty ∞ -to-2 2 2 distance at least ε \varepsilon ε from every expectation onto a unital subalgebra of M a n , s M_{a_{n,s}} M a n , s . We claim that the total weight of the bad summands tends to zero. To measure the failure of the spectral gap estimate on each summand, we set
η n , s = sup ∥ x ∥ ≤ 1 ( ∥ x − F n , s x ∥ 2 2 − L ∑ j = 1 h ∥ [ u j , n , s , x ] ∥ 2 2 ) ≥ 0. \eta_{n,s}=\sup_{\lVert x\rVert\leq1}\left(\lVert x-F_{n,s}x\rVert_2^2-L\sum_{j=1}^{h}\lVert[u_{j,n,s},x]\rVert_2^2\right)\geq0. η n , s = ∥ x ∥ ≤ 1 sup ( ∥ x − F n , s x ∥ 2 2 − L j = 1 ∑ h ∥[ u j , n , s , x ] ∥ 2 2 ) ≥ 0.
where we use the normalized trace on that summand. Choosing maximizing contractions in all the summands and applying spectral gap in P P P gives ∑ s w n , s η n , s → U 0 \sum_s w_{n,s}\eta_{n,s}\to_U0 ∑ s w n , s η n , s → U 0 . If the total bad weight did not tend to zero, Markov’s inequality would then allow us to choose bad summands s ( n ) s(n) s ( n ) for which η n , s ( n ) → U 0 \eta_{n,s(n)}\to_U0 η n , s ( n ) → U 0 .
We show that this choice contradicts the matrix version of the theorem. In the auxiliary tracial ultraproduct Q = ∏ U M a n , s ( n ) Q=\prod_U M_{a_{n,s(n)}} Q = ∏ U M a n , s ( n ) , put v j = [ u j , n , s ( n ) ] U v_j=[u_{j,n,s(n)}]_U v j = [ u j , n , s ( n ) ] U and D = Q ∩ { v 1 , … , v h } ′ D=Q\cap\{v_1,\ldots,v_h\}' D = Q ∩ { v 1 , … , v h } ′ . The errors on the selected summands tend to zero, so the induced maps T , F T,F T , F satisfy
∥ x − F x ∥ 2 2 ≤ L ∑ j = 1 h ∥ [ v j , x ] ∥ 2 2 . \lVert x-Fx\rVert_2^2\leq L\sum_{j=1}^{h}\lVert[v_j,x]\rVert_2^2. ∥ x − F x ∥ 2 2 ≤ L j = 1 ∑ h ∥[ v j , x ] ∥ 2 2 .
We also have T F = F TF=F T F = F : indeed, 0 ≤ T n , s ≤ I 0\leq T_{n,s}\leq I 0 ≤ T n , s ≤ I on L 2 L^2 L 2 gives the bound ∥ ( I − T n , s ) T n , s k n ∥ 2 → 2 ≤ 1 / ( k n + 1 ) \lVert(I-T_{n,s})T_{n,s}^{k_n}\rVert_{2\to2}\leq1/(k_n+1) ∥( I − T n , s ) T n , s k n ∥ 2 → 2 ≤ 1/ ( k n + 1 ) . It follows that F ( Q ) ⊆ D F(Q)\subseteq D F ( Q ) ⊆ D , while the displayed inequality shows that F F F fixes D D D pointwise. Since the induced map is self-adjoint, we obtain F = E D F=E_D F = E D . Thus the same inequality is a spectral gap estimate, and we can apply [ALT26, Theorem B], whose statement also allows bounded coordinate dimensions, to obtain D = [ D n ] U D=[D_n]_U D = [ D n ] U . The coordinate expectations and the selected powers now induce the same map. Hence [ALT26, Lemma 2.1] gives ∥ F n , s ( n ) − E D n ∥ ∞ → 2 → U 0 \lVert F_{n,s(n)}-E_{D_n}\rVert_{\infty\to2}\to_U0 ∥ F n , s ( n ) − E D n ∥ ∞ → 2 → U 0 , contrary to the choice of bad summands.
We have proved that the bad weight tends to zero for every ε > 0 \varepsilon>0 ε > 0 . In each summand, choose an expectation within 1 / n 1/n 1/ n of the best approximation to F n , s F_{n,s} F n , s , and let C n ⊆ P n C_n\subseteq P_n C n ⊆ P n be the direct sum of the resulting ranges. All errors are bounded by 2 2 2 , so separating the good and bad summands gives
∥ F n − E C n ∥ ∞ → 2 2 ≤ ( ε + 1 / n ) 2 + 4 ∑ s b a d w n , s . \lVert F_n-E_{C_n}\rVert_{\infty\to2}^2\leq(\varepsilon+1/n)^2+4\sum_{s\ {\rm bad}}w_{n,s}. ∥ F n − E C n ∥ ∞ → 2 2 ≤ ( ε + 1/ n ) 2 + 4 s bad ∑ w n , s .
Taking the ultralimit and then letting ε \varepsilon ε tend to zero shows that the total error vanishes. Thus these coordinate expectations induce E C E_C E C , and we conclude that C = [ C n ] U C=[C_n]_U C = [ C n ] U . □ \square □
Commuting internal abelian algebras
To combine commuting internal abelian algebras, we need their relative join to remain internal. Their coordinate models need not commute exactly, so we use the abelian stability theorem [1 Theorem 1.6] . It states that for every δ > 0 \delta> 0 δ > 0 there is ε > 0 \varepsilon> 0 ε > 0 , independent of the countable abelian group G G G and the dimension d d d , such that a map f : G → U ( d ) f : G \to U(d) f : G → U ( d ) with uniform multiplicative defect less than ε \varepsilon ε in normalized Hilbert–Schmidt norm is uniformly within δ \delta δ of a unitary representation on C d \mathbb{C}^{d} C d .
Proposition 2.3. Let A , B A, B A , B be commuting internal abelian subalgebras of a tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},d} M U , d . Then A ′ ∩ B ′ A' \cap B' A ′ ∩ B ′ and A ∨ r e l B A \mathbin{\vee_{\mathrm{rel}}} B A ∨ rel B are internal, and the latter is abelian.
Proof. We choose abelian models A = [ A n ] U A = [A_n]_{\mathcal{U}} A = [ A n ] U and B = [ B n ] U B = [B_n]_{\mathcal{U}} B = [ B n ] U and first record what commutation means in the coordinates. Since the two internal algebras commute, a coordinate selection argument gives
ε n : = sup a ∈ U ( A n ) , b ∈ U ( B n ) ∥ [ a , b ] ∥ 2 , d n ⟶ U 0. \varepsilon_n := \sup_{\substack{a \in U(A_n),\\ b \in U(B_n)}} \lVert[a,b] \rVert_{2,d_n} \longrightarrow_{\mathcal{U}} 0. ε n := a ∈ U ( A n ) , b ∈ U ( B n ) sup ∥[ a , b ] ∥ 2 , d n ⟶ U 0. To replace this approximate commutation by an exact abelian action, we choose countable dense subgroups G n ⊆ U ( A n ) G_n \subseteq U(A_n) G n ⊆ U ( A n ) and H n ⊆ U ( B n ) H_n \subseteq U(B_n) H n ⊆ U ( B n ) . The map f n : G n × H n → U ( d n ) f_n : G_n \times H_n \to U(d_n) f n : G n × H n → U ( d n ) defined by f n ( a , b ) = a b f_n(a,b) = ab f n ( a , b ) = ab then has multiplicative defect at most ε n \varepsilon_n ε n , since its defect on ( a , b ) , ( c , e ) (a,b),(c,e) ( a , b ) , ( c , e ) is a [ b , c ] e a[b,c]e a [ b , c ] e . Uniform abelian stability therefore gives representations π n \pi_n π n in the same dimensions satisfying
η n : = sup a ∈ G n , b ∈ H n ∥ π n ( a , b ) − a b ∥ 2 , d n ⟶ U 0. \eta_n := \sup_{\substack{a \in G_n,\\ b \in H_n}} \lVert\pi_n(a,b) - ab \rVert_{2,d_n} \longrightarrow_{\mathcal{U}} 0. η n := a ∈ G n , b ∈ H n sup ∥ π n ( a , b ) − ab ∥ 2 , d n ⟶ U 0. We use their commutants as coordinate models by setting C n = π n ( G n × H n ) ′ ∩ M d n C_n = \pi_n(G_n \times H_n)' \cap M_{d_n} C n = π n ( G n × H n ) ′ ∩ M d n . To identify the resulting ultraproduct, take x = [ x n ] U ∈ A ′ ∩ B ′ x = [x_n]_{\mathcal{U}} \in A' \cap B' x = [ x n ] U ∈ A ′ ∩ B ′ with bounded representatives. A coordinate selection argument shows that these representatives commute asymptotically and uniformly with G n G_n G n and H n H_n H n . By the uniform approximation just obtained, they also commute asymptotically and uniformly with π n ( G n × H n ) \pi_n(G_n \times H_n) π n ( G n × H n ) . Averaging over the compact closure of this image with respect to Haar measure therefore gives
∥ x n − E C n x n ∥ 2 , d n ≤ sup g ∈ G n × H n ∥ [ x n , π n ( g ) ] ∥ 2 , d n ⟶ U 0. \lVert x_n - E_{C_n}x_n \rVert_{2,d_n} \leq\sup_{g \in G_n \times H_n} \lVert[x_n,\pi_n(g)] \rVert_{2,d_n} \longrightarrow_{\mathcal{U}} 0. ∥ x n − E C n x n ∥ 2 , d n ≤ g ∈ G n × H n sup ∥[ x n , π n ( g )] ∥ 2 , d n ⟶ U 0. Thus these representatives can be replaced by elements of the coordinate commutants without changing their ultraproduct class.
Conversely, representatives in C n C_n C n commute asymptotically and uniformly with G n G_n G n and H n H_n H n by comparison with π n ( a , 1 ) \pi_n(a,1) π n ( a , 1 ) and π n ( 1 , b ) \pi_n(1,b) π n ( 1 , b ) . Density then shows that their ultraproduct classes commute with A A A and B B B . We have proved A ′ ∩ B ′ = [ C n ] U A' \cap B' = [C_n]_{\mathcal{U}} A ′ ∩ B ′ = [ C n ] U , and Lemma 2.1 shows that its commutant is internal as well. To check that this relative join is abelian, put S = A ∪ B S = A \cup B S = A ∪ B . Since the original algebras are abelian and commute, we have S ⊆ S ′ S \subseteq S' S ⊆ S ′ , and taking commutants gives S ′ ′ ⊆ S ′ = ( S ′ ′ ) ′ S'' \subseteq S' = (S'')' S ′′ ⊆ S ′ = ( S ′′ ) ′ . This is precisely the required abelianness.
Corollary 2.4. Every finite family of pairwise commuting internal abelian subalgebras of a tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},d} M U , d is contained in a common internal MASA.
Proof. We apply Proposition 2.3 successively to form the relative joins of the given algebras. At each step, the next algebra commutes with the relative bicommutant of the preceding family, so the proposition applies again and the resulting join remains internal and abelian. The final join contains the whole family, and Lemma 2.1 places it in an internal MASA.
Recovering the diagonal and its normalizer
We construct two tuples with spectral gap whose abelian commutants have the canonical diagonal as their joint commutant. The internality results then identify its image under an isomorphism. Passing to the Weyl group recovers the associated universal sofic group and gives the isomorphism conclusions.
Explicit quantum expanders
We need unitaries whose spectral gap constants do not depend on dimension. The following explicit pairs give such estimates for scalar commutants and for sums of scalar blocks of distinct sizes.
Lemma 3.1. For k ≥ 1 k \ge1 k ≥ 1 , let ρ k = Sym k − 1 ( C 2 ) \rho_k = \operatorname{Sym}^{k-1}(\mathbb{C}^{2}) ρ k = Sym k − 1 ( C 2 ) be the k k k -dimensional irreducible unitary representation of SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) , written in the normalized symmetric-tensor basis, and put
g 1 = 1 3 ( 1 + i 1 − 1 1 − i ) , g 2 = 1 3 ( 1 + i − 1 1 1 − i ) , g_1 = \frac{1}{\sqrt{3}}\begin{pmatrix} 1+i & 1 \\ -1 & 1-i \end{pmatrix}, \qquad g_2 = \frac{1}{\sqrt{3}}\begin{pmatrix} 1+i & -1 \\ 1 & 1-i \end{pmatrix}, g 1 = 3 1 ( 1 + i − 1 1 1 − i ) , g 2 = 3 1 ( 1 + i 1 − 1 1 − i ) , v j , k = ρ k ( g j ) ( j = 1 , 2 ) . v_{j,k} = \rho_k(g_j) \qquad(j=1,2). v j , k = ρ k ( g j ) ( j = 1 , 2 ) . Then the following hold.
(1) For every k ≥ 1 k \ge1 k ≥ 1 , the pair ( v 1 , k , v 2 , k ) (v_{1,k},v_{2,k}) ( v 1 , k , v 2 , k ) has commutant C 1 \mathbb{C}1 C 1 in M k M_k M k and spectral gap with constant 2 2 2 :
∥ z − τ k ( z ) 1 ∥ 2 , k 2 ≤ 2 ∑ j = 1 2 ∥ [ z , v j , k ] ∥ 2 , k 2 ( z ∈ M k ) . \lVert z-\tau_k(z)1\rVert_{2,k}^{2} \le2\sum_{j=1}^{2}\lVert[z,v_{j,k}]\rVert_{2,k}^{2}\qquad(z\in M_k). ∥ z − τ k ( z ) 1 ∥ 2 , k 2 ≤ 2 j = 1 ∑ 2 ∥[ z , v j , k ] ∥ 2 , k 2 ( z ∈ M k ) . (2) Let m 1 , … , m r m_1,\ldots,m_r m 1 , … , m r be pairwise distinct positive integers, and put
C = ⨁ i = 1 r C 1 m i ⊆ M d , d = ∑ i m i , C=\bigoplus_{i=1}^{r}\mathbb{C}1_{m_i}\subseteq M_d,\qquad d=\sum_i m_i, C = i = 1 ⨁ r C 1 m i ⊆ M d , d = i ∑ m i , v ^ j = ⨁ i = 1 r v j , m i ( j = 1 , 2 ) . \widehat{v}_j=\bigoplus_{i=1}^{r}v_{j,m_i}\qquad(j=1,2). v j = i = 1 ⨁ r v j , m i ( j = 1 , 2 ) . Then ( v ^ 1 , v ^ 2 ) (\widehat{v}_1,\widehat{v}_2) ( v 1 , v 2 ) has commutant C C C in M d M_d M d and spectral gap with constant 8 8 8 :
∥ x − E C x ∥ 2 2 ≤ 8 ∑ j = 1 2 ∥ [ x , v ^ j ] ∥ 2 2 ( x ∈ M d ) . \lVert x-E_Cx\rVert_2^2\le8\sum_{j=1}^{2}\lVert[x,\widehat{v}_j]\rVert_2^2\qquad(x\in M_d). ∥ x − E C x ∥ 2 2 ≤ 8 j = 1 ∑ 2 ∥[ x , v j ] ∥ 2 2 ( x ∈ M d ) .
Proof. To prove (1), we relate the given unitaries to the Ramanujan construction. Under the usual identification of unit quaternions with SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) , they correspond to g 1 = ( 1 + i + j ) / 3 g_1=(1+i+j)/\sqrt{3} g 1 = ( 1 + i + j ) / 3 and g 2 = ( 1 + i − j ) / 3 g_2=(1+i-j)/\sqrt{3} g 2 = ( 1 + i − j ) / 3 . For this comparison, put s ε , η = ( k + ε i + η j ) / 3 s_{\varepsilon,\eta}=(k+\varepsilon i+\eta j)/\sqrt{3} s ε , η = ( k + ε i + η j ) / 3 , for ε , η ∈ { − 1 , 1 } \varepsilon,\eta\in\{-1,1\} ε , η ∈ { − 1 , 1 } . These are the four half-turns in the p = 3 p=3 p = 3 construction of [20 Appendix C] . Since
{ k s ε , η : ε , η ∈ { − 1 , 1 } } = { − g 1 , − g 1 − 1 , − g 2 , − g 2 − 1 } , \{ks_{\varepsilon,\eta}:\varepsilon,\eta\in\{-1,1\}\}=\{-g_1,-g_1^{-1},-g_2,-g_2^{-1}\}, { k s ε , η : ε , η ∈ { − 1 , 1 }} = { − g 1 , − g 1 − 1 , − g 2 , − g 2 − 1 } ,
we can apply the Ramanujan estimate in [20 Section 6.1, (93)] to obtain
∥ T k ∣ ker τ k ∥ 2 → 2 ≤ 2 3 , T k = ∑ j = 1 2 ( Ad ( v j , k ) + Ad ( v j , k ∗ ) ) . \left\|T_k\big|_{\ker\tau_k}\right\|_{2\to2}\le2\sqrt{3},\qquad T_k=\sum_{j=1}^{2}\left(\operatorname{Ad}(v_{j,k})+\operatorname{Ad}(v_{j,k}^{*})\right). T k k e r τ k 2 → 2 ≤ 2 3 , T k = j = 1 ∑ 2 ( Ad ( v j , k ) + Ad ( v j , k ∗ ) ) .
Indeed, multiplication by Ad ( ρ k ( k ) ) \operatorname{Ad}(\rho_k(k)) Ad ( ρ k ( k )) preserves the norm, while the central signs disappear under conjugation. To obtain the required commutator estimate, take z ∈ M k z\in M_k z ∈ M k and z 0 = z − τ k ( z ) 1 z_0=z-\tau_k(z)1 z 0 = z − τ k ( z ) 1 . The preceding bound yields
∑ j = 1 2 ∥ [ z , v j , k ] ∥ 2 , k 2 = ⟨ z 0 , ( 4 I − T k ) z 0 ⟩ ≥ ( 4 − 2 3 ) ∥ z 0 ∥ 2 , k 2 ≥ 1 2 ∥ z 0 ∥ 2 , k 2 . \sum_{j=1}^{2}\lVert[z,v_{j,k}]\rVert_{2,k}^{2}=\langle z_0,(4I-T_k)z_0\rangle\ge(4-2\sqrt{3})\lVert z_0\rVert_{2,k}^{2}\ge\frac{1}{2}\lVert z_0\rVert_{2,k}^{2}. j = 1 ∑ 2 ∥[ z , v j , k ] ∥ 2 , k 2 = ⟨ z 0 , ( 4 I − T k ) z 0 ⟩ ≥ ( 4 − 2 3 ) ∥ z 0 ∥ 2 , k 2 ≥ 2 1 ∥ z 0 ∥ 2 , k 2 .
This gives the spectral gap estimate with constant 2 2 2 . It also shows that every element of the commutant lies in C 1 \mathbb{C}1 C 1 , completing the proof of (1).
For (2), we first use (1) to obtain a gap for arbitrary representations. Given a unit vector ξ \xi ξ in a nontrivial irreducible representation ρ m \rho_m ρ m , we apply the estimate to the rank-one projection p p p onto C ξ \mathbb{C}\xi C ξ . Writing ∥ x ∥ H S = Tr ( x ∗ x ) 1 / 2 \lVert x\rVert_{\mathrm{HS}}=\operatorname{Tr}(x^{*}x)^{1/2} ∥ x ∥ HS = Tr ( x ∗ x ) 1/2 for the unnormalized Hilbert–Schmidt norm, we obtain
1 2 ≤ 1 − 1 m = ∥ p − 1 m 1 ∥ H S 2 ≤ 2 ∑ j = 1 2 ∥ [ p , ρ m ( g j ) ] ∥ H S 2 ≤ 4 ∑ j = 1 2 ∥ ρ m ( g j ) ξ − ξ ∥ 2 . \frac{1}{2}\le1-\frac{1}{m}=\left\lVert p-\frac{1}{m}1\right\rVert_{\mathrm{HS}}^{2}\le2\sum_{j=1}^{2}\lVert[p,\rho_m(g_j)]\rVert_{\mathrm{HS}}^{2}\le4\sum_{j=1}^{2}\lVert\rho_m(g_j)\xi-\xi\rVert^{2}. 2 1 ≤ 1 − m 1 = p − m 1 1 HS 2 ≤ 2 j = 1 ∑ 2 ∥[ p , ρ m ( g j )] ∥ HS 2 ≤ 4 j = 1 ∑ 2 ∥ ρ m ( g j ) ξ − ξ ∥ 2 .
Here the first inequality uses m ≥ 2 m\ge2 m ≥ 2 , and the last follows from the rank-one calculation
∥ [ p , v ] ∥ H S 2 = 2 ( 1 − ∣ ⟨ v ξ , ξ ⟩ ∣ 2 ) ≤ 2 ∥ v ξ − ξ ∥ 2 . \lVert[p,v]\rVert_{\mathrm{HS}}^{2}=2(1-\lvert\langle v\xi,\xi\rangle\rvert^{2})\le2\lVert v\xi-\xi\rVert^{2}. ∥[ p , v ] ∥ HS 2 = 2 ( 1 − ∣⟨ v ξ , ξ ⟩ ∣ 2 ) ≤ 2 ∥ v ξ − ξ ∥ 2 .
for a unitary v v v . The resulting bound is independent of the irreducible representation. By homogeneity and complete reducibility, it follows that every finite-dimensional continuous unitary representation π \pi π of SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) satisfies
∥ ξ − p π ξ ∥ 2 2 ≤ 8 ∑ j = 1 2 ∥ π ( g j ) ξ − ξ ∥ 2 , (4) \lVert\xi-p_{\pi}\xi\rVert_{2}^{2}\le8\sum_{j=1}^{2}\lVert\pi(g_j)\xi-\xi\rVert^{2}, \tag*{(4)} ∥ ξ − p π ξ ∥ 2 2 ≤ 8 j = 1 ∑ 2 ∥ π ( g j ) ξ − ξ ∥ 2 , (4)
where p π p_{\pi} p π projects onto the invariant vectors.
We now apply (4) to the conjugation representation associated with ρ = ⨁ i ρ m i \rho=\bigoplus_i\rho_{m_i} ρ = ⨁ i ρ m i on L 2 ( M d ) L^2(M_d) L 2 ( M d ) . The summands ρ m i \rho_{m_i} ρ m i are pairwise inequivalent, so Schur’s lemma identifies the commutant as ρ ( SU ( 2 ) ) ′ = C \rho(\operatorname{SU}(2))'= \mathbb{C} ρ ( SU ( 2 ) ) ′ = C . Thus averaging the conjugation action gives E C E_C E C , and the representation gap becomes precisely the estimate in (2). Since vanishing of its right-hand side forces membership in this algebra, it also identifies the commutant of v ^ 1 , v ^ 2 \hat{v}_1,\hat{v}_2 v ^ 1 , v ^ 2 with C \mathbb{C} C . □
Recovering the diagonal
The scalar estimate of Lemma 3.1 lets us obtain row and column algebras as commutants with spectral gap. Their joint commutant is the diagonal. The block estimate will be needed for general internal algebras in Section 4.
Lemma 3.2. For every k ≥ 1 k\ge1 k ≥ 1 , there are two unitary triples ( z i 1 , k , z i 2 , k , z i 3 , k ) (z_{i1,k},z_{i2,k},z_{i3,k}) ( z i 1 , k , z i 2 , k , z i 3 , k ) , i = 1 , 2 i=1,2 i = 1 , 2 , in M k 2 = M k ⊗ M k M_{k^2}=M_k\otimes M_k M k 2 = M k ⊗ M k , with commutants
B 1 , k = D k ⊗ 1 , B 2 , k = 1 ⊗ D k , B_{1,k}=D_k\otimes1,\qquad B_{2,k}=1\otimes D_k, B 1 , k = D k ⊗ 1 , B 2 , k = 1 ⊗ D k , and spectral gap with constant 10 10 10 :
∥ x − E B i , k x ∥ 2 2 ≤ 10 ∑ j = 1 3 ∥ [ x , z i j , k ] ∥ 2 2 ( x ∈ M k 2 , i = 1 , 2 ) . \lVert x-E_{B_{i,k}}x\rVert_2^2\le10\sum_{j=1}^{3}\lVert[x,z_{ij,k}]\rVert_2^2\qquad(x\in M_{k^2},\ i=1,2). ∥ x − E B i , k x ∥ 2 2 ≤ 10 j = 1 ∑ 3 ∥[ x , z ij , k ] ∥ 2 2 ( x ∈ M k 2 , i = 1 , 2 ) . For every tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},d} M U , d , these triples yield six unitaries z i j z_{ij} z ij whose commutants B i = { z i 1 , z i 2 , z i 3 } ′ B_i=\{z_{i1},z_{i2},z_{i3}\}' B i = { z i 1 , z i 2 , z i 3 } ′ are commuting internal abelian algebras, have the same spectral gap estimates, and satisfy
B 1 ′ ∩ B 2 ′ = [ D d n ] U . B'_1\cap B'_2=[D_{d_n}]_{\mathcal{U}}. B 1 ′ ∩ B 2 ′ = [ D d n ] U .
Proof. We construct the first triple by using the expander pair to leave one matrix factor in the commutant, and a third unitary to reduce that factor to its diagonal. Index the standard basis of C k \mathbb{C}^k C k by 0 , … , k − 1 0,\ldots,k-1 0 , … , k − 1 , let S S S be its cyclic shift, and set
W = ∑ r = 0 k − 1 e r r ⊗ S r , ( z 11 , k , z 12 , k , z 13 , k ) = ( 1 ⊗ v 1 , k , 1 ⊗ v 2 , k , W ) . W=\sum_{r=0}^{k-1}e_{rr}\otimes S^r,\qquad(z_{11,k},z_{12,k},z_{13,k})=(1\otimes v_{1,k},1\otimes v_{2,k},W). W = r = 0 ∑ k − 1 e r r ⊗ S r , ( z 11 , k , z 12 , k , z 13 , k ) = ( 1 ⊗ v 1 , k , 1 ⊗ v 2 , k , W ) .
We take the tensor flip for the second triple, so it suffices to verify the first. Put Q = M k ⊗ 1 Q=M_k\otimes1 Q = M k ⊗ 1 and a = ∥ x − E Q x ∥ 2 a=\lVert x-E_Qx\rVert_2 a = ∥ x − E Q x ∥ 2 . Tensoring Lemma 3.1(1) with the identity gives a 2 ≤ 2 ∑ j = 1 2 ∥ [ x , z 1 j , k ] ∥ 2 2 a^2\le2\sum_{j=1}^{2}\lVert[x,z_{1j,k}]\rVert_2^2 a 2 ≤ 2 ∑ j = 1 2 ∥[ x , z 1 j , k ] ∥ 2 2 , which controls the distance to the first matrix factor. Within that factor, the orthonormality of the powers of S S S for the normalized trace gives
∥ [ y , W ] ∥ 2 2 = 2 ∥ y − E B 1 , k y ∥ 2 2 ( y ∈ Q ) . \lVert[y,W]\rVert_2^2=2\lVert y-E_{B_{1,k}}y\rVert_2^2\qquad(y\in Q). ∥[ y , W ] ∥ 2 2 = 2 ∥ y − E B 1 , k y ∥ 2 2 ( y ∈ Q ) .
We combine these two estimates by writing c = ∥ [ x , W ] ∥ 2 c=\lVert[x,W]\rVert_2 c = ∥[ x , W ] ∥ 2 and applying Pythagoras:
∥ x − E B 1 , k x ∥ 2 2 ≤ a 2 + 1 2 ( c + 2 a ) 2 ≤ 5 a 2 + c 2 ≤ 10 ∑ j = 1 3 ∥ [ x , z 1 j , k ] ∥ 2 2 . \lVert x-E_{B_{1,k}}x\rVert_2^2\le a^2+\frac{1}{2}(c+2a)^2\le5a^2+c^2\le10\sum_{j=1}^{3}\lVert[x,z_{1j,k}]\rVert_2^2. ∥ x − E B 1 , k x ∥ 2 2 ≤ a 2 + 2 1 ( c + 2 a ) 2 ≤ 5 a 2 + c 2 ≤ 10 j = 1 ∑ 3 ∥[ x , z 1 j , k ] ∥ 2 2 .
This proves the asserted gap. Since the triple commutes with B 1 , k B_{1,k} B 1 , k , the same estimate identifies its commutant, and the tensor flip gives the corresponding conclusions for the second triple.
To pass to arbitrary d n d_n d n , we put k n = ⌊ d n ⌋ k_n=\lfloor\sqrt{d_n}\rfloor k n = ⌊ d n ⌋ and extend the six unitaries by the identity on the complement of the standard k n 2 k_n^2 k n 2 -dimensional corner p n p_n p n . The relative dimension of this corner tends to one, so
∥ x n − p n x n p n ∥ 2 , d n ≤ 2 ∥ x n ∥ 1 − k n 2 / d n ⟶ 0 \lVert x_n-p_nx_np_n\rVert_{2,d_n}\le2\lVert x_n\rVert\sqrt{1-k_n^2/d_n}\longrightarrow0 ∥ x n − p n x n p n ∥ 2 , d n ≤ 2 ∥ x n ∥ 1 − k n 2 / d n ⟶ 0
for every bounded sequence. Hence corner inclusion identifies the ultraproduct of the square matrices with M U , d \mathcal{M}_{\mathcal{U},d} M U , d . This identification preserves the canonical diagonals and transfers the gap estimates.
In particular, the resulting B i B_i B i are the ultraproducts of the row and column algebras, extended by scalars on the negligible complement.
It remains to identify their joint commutant. In each square corner, the expectations onto B 1 , k ′ = D k ⊗ M k B'_{1,k}=D_k\otimes M_k B 1 , k ′ = D k ⊗ M k and B 2 , k ′ = M k ⊗ D k B'_{2,k}=M_k\otimes D_k B 2 , k ′ = M k ⊗ D k commute and have product E D k 2 E_{D_k^2} E D k 2 . Their product estimate therefore passes to the ultraproduct, where Lemma 2.1 identifies the two commutants and gives the asserted intersection. □
The next lemma recognizes an internal diagonal from these spectral gap and commutation conditions. This will let us transport the construction under isomorphisms and, later, by transferring its defining conditions.
Lemma 3.3. Let C 1 C_1 C 1 , C 2 C_2 C 2 be commuting abelian subalgebras of a tracial matrix ultraproduct. Suppose that each C i C_i C i is the commutant of a finite unitary tuple with spectral gap, and that D = C 1 ′ ∩ C 2 ′ D=C'_1\cap C'_2 D = C 1 ′ ∩ C 2 ′ is abelian. Then D D D is an internal MASA, and hence is unitarily conjugate to the canonical diagonal.
Proof. The spectral gap hypothesis makes both C i C_i C i internal by Theorem 2.2. We first identify their joint commutant as a MASA: since the algebras commute and are abelian, C i ⊆ D C_i\subseteq D C i ⊆ D , whence D ′ ⊆ D D'\subseteq D D ′ ⊆ D . Together with the assumed abelianness, this shows that D D D is a MASA. To prove its internality, apply Corollary 2.4 to place C 1 , C 2 C_1,C_2 C 1 , C 2 in a common internal MASA E E E . Then E ⊆ D E\subseteq D E ⊆ D , and maximal abelianness forces E = D E=D E = D . The required conjugacy now follows from Lemma 2.1. □
Proof of Theorem C(1). Let θ : M U , d → M V , e \theta:\mathcal M_{\mathcal U,d}\to\mathcal M_{\mathcal V,e} θ : M U , d → M V , e be an isomorphism, and let D , F D,F D , F be the canonical diagonals. To apply the recognition lemma to the image diagonal, we use Lemma 3.2 to write D = B 1 ′ ∩ B 2 ′ D=B'_1\cap B'_2 D = B 1 ′ ∩ B 2 ′ . Since the isomorphism preserves the trace and commutants, it also preserves the spectral gap estimates. Consequently θ ( B 1 ) , θ ( B 2 ) \theta(B_1),\theta(B_2) θ ( B 1 ) , θ ( B 2 ) satisfy the hypotheses of Lemma 3.3, with joint commutant θ ( D ) \theta(D) θ ( D ) . We conclude from that lemma that θ ( D ) \theta(D) θ ( D ) is an internal MASA and hence is unitarily conjugate to F F F . □
Weyl groups and isomorphisms
Having recovered the diagonal up to conjugacy, we use its normalizer to recover the universal sofic group. For an internal MASA D ⊆ M U , d D\subseteq\mathcal M_{\mathcal U,d} D ⊆ M U , d , write
N M U , d ( D ) = { u ∈ U ( M U , d ) : u D u ∗ = D } , N_{\mathcal M_{\mathcal U,d}}(D)=\{u\in U(\mathcal M_{\mathcal U,d}):uDu^*=D\}, N M U , d ( D ) = { u ∈ U ( M U , d ) : u D u ∗ = D } ,
W ( M U , d , D ) = N M U , d ( D ) / U ( D ) . W(\mathcal M_{\mathcal U,d},D)=N_{\mathcal M_{\mathcal U,d}}(D)/U(D). W ( M U , d , D ) = N M U , d ( D ) / U ( D ) .
We call W ( M U , d , D ) W(\mathcal M_{\mathcal U,d},D) W ( M U , d , D ) the Weyl group . We identify permutations with their permutation matrices. In the coordinate algebras, every element of N M n ( D n ) N_{M_n}(D_n) N M n ( D n ) has the form σ w \sigma w σ w , where σ ∈ Sym ( n ) \sigma\in\operatorname{Sym}(n) σ ∈ Sym ( n ) and w ∈ U ( D n ) w\in U(D_n) w ∈ U ( D n ) . To carry this description to the ultraproduct, we need to represent each normalizer by coordinate normalizers. The next estimate gives the required approximation.
Popa, Sinclair and Smith proved quantitative estimates for the distance of a unitary to the normalizer of a MASA in a separably acting I I 1 \mathrm{II}_1 II 1 factor [31 Theorem 6.2] . For matrix diagonals, the following elementary argument gives the bound needed here.
Lemma 3.4. For u ∈ U ( n ) u\in U(n) u ∈ U ( n ) , put
r D n ( u ) = sup a ∈ ( D n ) 1 ∥ u a u ∗ − E D n ( u a u ∗ ) ∥ 2 . r_{D_n}(u)=\sup_{a\in(D_n)_1}\|uau^*-E_{D_n}(uau^*)\|_2. r D n ( u ) = a ∈ ( D n ) 1 sup ∥ u a u ∗ − E D n ( u a u ∗ ) ∥ 2 .
There is a unitary v ∈ N M n ( D n ) v\in N_{M_n}(D_n) v ∈ N M n ( D n ) such that
∥ u − v ∥ 2 ≤ 2 r D n ( u ) . (5) \|u-v\|_2\leq\sqrt{2}\,r_{D_n}(u). \tag*{(5)} ∥ u − v ∥ 2 ≤ 2 r D n ( u ) . (5)
The same estimate holds for an internal MASA in a tracial matrix ultraproduct M U , d \mathcal M_{\mathcal U,d} M U , d . After conjugating it to the canonical diagonal, each unitary normalizer has coordinate representatives that normalize the coordinate diagonals.
Proof. We use the doubly stochastic matrix P = ( ∣ u i j ∣ 2 ) i , j P=(|u_{ij}|^2)_{i,j} P = ( ∣ u ij ∣ 2 ) i , j to choose the support of a nearby normalizing unitary. To relate its entries to the normalizer defect, take independent uniform signs ε j \varepsilon_j ε j and set s = diag ( ε 1 , … , ε n ) s=\operatorname{diag}(\varepsilon_1,\ldots,\varepsilon_n) s = diag ( ε 1 , … , ε n ) . Averaging over these signs gives
E ∥ u s u ∗ − E D n ( u s u ∗ ) ∥ 2 2 = 1 − 1 n ∑ i , j P i j 2 . \mathbb E\|usu^*-E_{D_n}(usu^*)\|_2^2=1-\frac{1}{n}\sum_{i,j}P_{ij}^2. E ∥ u s u ∗ − E D n ( u s u ∗ ) ∥ 2 2 = 1 − n 1 i , j ∑ P ij 2 .
Thus, with r = r D n ( u ) r=r_{D_n}(u) r = r D n ( u ) , we obtain ∑ i , j P i j 2 ≥ n ( 1 − r 2 ) \sum_{i,j}P_{ij}^{2}\ge n(1-r^{2}) ∑ i , j P ij 2 ≥ n ( 1 − r 2 ) . By Birkhoff’s theorem, P P P is a convex combination of permutation matrices, so one of the corresponding permutations, say σ \sigma σ , satisfies
∑ i P i , σ ( i ) ≥ ∑ i , j P i j 2 . \sum_i P_{i,\sigma(i)}\ge\sum_{i,j}P_{ij}^{2}. i ∑ P i , σ ( i ) ≥ i , j ∑ P ij 2 .
We choose v ∈ N M n ( D n ) v\in N_{M_n}(D_n) v ∈ N M n ( D n ) supported on these positions and give its nonzero entries the phases of the corresponding entries of u u u . At a zero entry of u u u , the phase can be chosen arbitrarily. With this choice, the preceding bound yields
∥ u − v ∥ 2 2 = 2 − 2 n ∑ i ∣ u i , σ ( i ) ∣ ≤ 2 − 2 n ∑ i P i , σ ( i ) ≤ 2 r 2 . \lVert u-v\rVert_2^2=2-\frac{2}{n}\sum_i\lvert u_{i,\sigma(i)}\rvert\le2-\frac{2}{n}\sum_iP_{i,\sigma(i)}\le2r^2. ∥ u − v ∥ 2 2 = 2 − n 2 i ∑ ∣ u i , σ ( i ) ∣ ≤ 2 − n 2 i ∑ P i , σ ( i ) ≤ 2 r 2 .
This proves the matrix estimate.
We next pass the estimate to a tracial matrix ultraproduct. Polar decomposition allows us to lift its unitaries to unitary sequences. For D = [ D d n ] U D=[D_{d_n}]_{\mathcal U} D = [ D d n ] U , the coordinate expectations induce E D E_D E D , and the normalizer defects satisfy
r D ( [ u n ] U ) = lim U r D d n ( u n ) . r_D([u_n]_{\mathcal U})=\lim_{\mathcal U}r_{D_{d_n}}(u_n). r D ([ u n ] U ) = U lim r D d n ( u n ) .
Indeed, the coordinate suprema give one inequality, and choosing a maximizer in each coordinate gives the reverse inequality. The matrix estimate therefore passes to the ultraproduct. If u u u normalizes D D D , the defect vanishes, so the resulting coordinate diagonal normalizers represent u u u . Finally, unitary conjugacy transfers both conclusions to every internal MASA. □
For the canonical diagonal, Lemma 3.4 now lets us identify the Weyl group with the associated universal sofic group:
W ( M U , d , [ D d n ] U ) ≅ ∏ U m e t Sym ( d n ) . (6) \mathcal{W}(\mathcal{M}_{\mathcal U,d},[D_{d_n}]_{\mathcal U})\cong\prod_{\mathcal U}^{\mathrm{met}}\operatorname{Sym}(d_n). \tag*{(6)} W ( M U , d , [ D d n ] U ) ≅ U ∏ met Sym ( d n ) . (6)
Indeed, every normalizer has coordinate representatives σ n w n \sigma_nw_n σ n w n , and the diagonal factors disappear in the quotient. To identify the metric as well, we observe that for permutation matrices,
1 2 inf w ∈ U ( D n ) ∥ σ − ρ w ∥ 2 2 = 1 − # { j : σ ( j ) = ρ ( j ) } n = d ( σ , ρ ) . (7) \frac{1}{2}\inf_{w\in U(D_n)}\lVert\sigma-\rho w\rVert_2^2
=1-\frac{\#\{j:\sigma(j)=\rho(j)\}}{n}=d(\sigma,\rho). \tag*{(7)} 2 1 w ∈ U ( D n ) inf ∥ σ − ρw ∥ 2 2 = 1 − n # { j : σ ( j ) = ρ ( j )} = d ( σ , ρ ) . (7)
Indeed, we match phases at the common nonzero entries, while the other entries have disjoint supports. We equip W \mathcal{W} W with the metric
d ( u ˉ , v ˉ ) = 1 2 inf w ∈ U ( D ) ∥ u − v w ∥ 2 2 . d(\bar{u},\bar{v})=\frac{1}{2}\inf_{w\in U(D)}\lVert u-vw\rVert_2^2. d ( u ˉ , v ˉ ) = 2 1 w ∈ U ( D ) inf ∥ u − v w ∥ 2 2 .
Since coordinate minima pass to ultralimits, this is the normalized Hamming metric under the identification above.
Theorem C′. Let D = [ D d n ] U ⊆ M U , d D=[D_{d_n}]_{\mathcal U}\subseteq\mathcal{M}_{\mathcal U,d} D = [ D d n ] U ⊆ M U , d and F = [ D e n ] V ⊆ M V , e F=[D_{e_n}]_{\mathcal V}\subseteq\mathcal{M}_{\mathcal V,e} F = [ D e n ] V ⊆ M V , e be the canonical diagonals of two tracial matrix ultraproducts.
(1) Every isomorphism θ : M U , d → M V , e \theta:\mathcal{M}_{\mathcal U,d}\to\mathcal{M}_{\mathcal V,e} θ : M U , d → M V , e sends D D D to a unitary conjugate of F F F .
(2) If M U , d ≅ M V , e \mathcal{M}_{\mathcal U,d}\cong\mathcal{M}_{\mathcal V,e} M U , d ≅ M V , e , then ( M U , d , D ) ≅ ( M V , e , F ) (\mathcal{M}_{\mathcal U,d},D)\cong(\mathcal{M}_{\mathcal V,e},F) ( M U , d , D ) ≅ ( M V , e , F ) .
(3) If M U , d ≅ M V , e \mathcal{M}_{\mathcal U,d}\cong\mathcal{M}_{\mathcal V,e} M U , d ≅ M V , e , then
∏ U m e t Sym ( d n ) ≅ ∏ V m e t Sym ( e n ) \prod_{\mathcal U}^{\mathrm{met}}\operatorname{Sym}(d_n)\cong\prod_{\mathcal V}^{\mathrm{met}}\operatorname{Sym}(e_n) U ∏ met Sym ( d n ) ≅ V ∏ met Sym ( e n ) as groups with the normalized Hamming metric.
Proof. By part (1), already proved above, we can compose any isomorphism with an inner automorphism so that it maps D D D onto F F F . This proves (2). The resulting isomorphism also identifies the Weyl groups and preserves d d d . Using (3.3) to identify these groups and (3.4) to identify their metrics, we obtain (3). □
Theorem A′. There are 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 pairwise nonisomorphic tracial matrix ultraproducts of the sequence ( M n ( C ) , τ n ) (M_n(\mathbb{C}),\tau_n) ( M n ( C ) , τ n ) , and 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 pairwise nonisomorphic metric ultraproducts of ( U ( n ) , d 2 ) (U(n),d_2) ( U ( n ) , d 2 ) , where d 2 ( u , v ) = ∥ u − v ∥ 2 d_2(u,v)=\lVert u-v\rVert_2 d 2 ( u , v ) = ∥ u − v ∥ 2 .
Proof. Apply [3 Theorem 5.6 and Remark 5.7] to choose continuum many nonisomorphic universal sofic groups. By Theorem C′(3), the corresponding matrix ultraproducts must also be pairwise nonisomorphic.
To obtain the assertion for unitary groups, suppose that two such groups were isomorphic. Dye’s reconstruction theorem would then give a linear or conjugate-linear ∗ * ∗ -isomorphism of the factors; see [19 p. 104] . Since entrywise complex conjugation is a conjugate-linear ∗ * ∗ -automorphism of a tracial matrix ultraproduct, either case gives a complex ∗ * ∗ -isomorphism, contradicting the first assertion. Finally, polar decomposition identifies U ( ∏ U M n ) U(\prod_{\mathcal U} M_n) U ( ∏ U M n ) with ∏ U m e t ( U ( n ) , d 2 ) \prod_{\mathcal U}^{\mathrm{met}}(U(n),d_2) ∏ U met ( U ( n ) , d 2 ) , so these unitary groups are the required metric ultraproducts. □
In Section 5 we will express these constructions by fixed formulas. The diagonal and its normalizer will then be uniformly definable, and the Weyl group uniformly interpretable. This will let us apply the same reduction to elementary equivalence.
Internal algebras and isomorphisms
To extend diagonal preservation to every internal subalgebra, we first treat the abelian case. We arrange its coordinate atoms in rows and columns with distinct total ranks, so that Lemma 3.1(2) gives two pairs of unitaries whose commutants recover the algebra. An arbitrary internal algebra can then be recovered inside the commutant of its center using two further unitaries.
Row and column encoding
Proposition 4.1. Every internal abelian subalgebra B B B of a tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},d} M U , d contains commuting internal abelian algebras B 1 , B 2 B_1,B_2 B 1 , B 2 and admits four unitaries u 1 , u 2 , v 1 , v 2 u_1,u_2,v_1,v_2 u 1 , u 2 , v 1 , v 2 such that
B 1 = { u 1 , u 2 } ′ , B 2 = { v 1 , v 2 } ′ , B ′ = B 1 ′ ∩ B 2 ′ , B = ( B 1 ′ ∩ B 2 ′ ) ′ . (8) \begin{aligned}
B_1&=\{u_1,u_2\}', \qquad B_2=\{v_1,v_2\}',\\
B'&=B_1'\cap B_2', \qquad B=(B_1'\cap B_2')'. \tag*{(8)}
\end{aligned} B 1 B ′ = { u 1 , u 2 } ′ , B 2 = { v 1 , v 2 } ′ , = B 1 ′ ∩ B 2 ′ , B = ( B 1 ′ ∩ B 2 ′ ) ′ . (8)
Both pairs ( u 1 , u 2 ) (u_1,u_2) ( u 1 , u 2 ) and ( v 1 , v 2 ) (v_1,v_2) ( v 1 , v 2 ) have spectral gap in M U , d \mathcal{M}_{\mathcal{U},d} M U , d with constant 8:
∥ x − E B 1 x ∥ 2 2 ≤ 8 ∑ j = 1 2 ∥ [ x , u j ] ∥ 2 2 , ( x ∈ M U , d ) . \lVert x-E_{B_1}x\rVert_2^2\le8\sum_{j=1}^{2}\lVert[x,u_j]\rVert_2^2,
\qquad(x\in\mathcal{M}_{\mathcal{U},d}). ∥ x − E B 1 x ∥ 2 2 ≤ 8 j = 1 ∑ 2 ∥[ x , u j ] ∥ 2 2 , ( x ∈ M U , d ) .
∥ x − E B 2 x ∥ 2 2 ≤ 8 ∑ j = 1 2 ∥ [ x , v j ] ∥ 2 2 . \lVert x-E_{B_2}x\rVert_2^2\le8\sum_{j=1}^{2}\lVert[x,v_j]\rVert_2^2. ∥ x − E B 2 x ∥ 2 2 ≤ 8 j = 1 ∑ 2 ∥[ x , v j ] ∥ 2 2 .
If B B B is a MASA, then B = B 1 ′ ∩ B 2 ′ B=B_1'\cap B_2' B = B 1 ′ ∩ B 2 ′ .
Proof. We will group the coordinate atoms into rows and columns so that the block spectral gap estimates apply to both families. By Lemma 2.1, we can choose abelian models B n ⊆ M d n B_n\subseteq M_{d_n} B n ⊆ M d n . We list their atoms e i , n e_{i,n} e i , n in decreasing order of rank and put m i , n = rank ( e i , n ) m_{i,n}=\operatorname{rank}(e_{i,n}) m i , n = rank ( e i , n ) , so that
e 1 , n , … , e r n , n , m 1 , n ≥ ⋯ ≥ m r n , n > 0. e_{1,n},\ldots,e_{r_n,n},\qquad m_{1,n}\ge\cdots\ge m_{r_n,n}>0. e 1 , n , … , e r n , n , m 1 , n ≥ ⋯ ≥ m r n , n > 0.
Our first task is to find projections p n ∈ B n ′ p_n\in B_n' p n ∈ B n ′ with τ d n ( 1 − p n ) → U 0 \tau_{d_n}(1-p_n)\mathrel{\to_{\mathcal U}}0 τ d n ( 1 − p n ) → U 0 and two abelian algebras C 1 , n , C 2 , n ⊆ p n B n p n C_{1,n},C_{2,n}\subseteq p_nB_np_n C 1 , n , C 2 , n ⊆ p n B n p n , each with pairwise distinct atom ranks, that together generate p n B n p n p_nB_np_n p n B n p n . The small discarded complements will not affect the resulting internal algebras.
Suppose first that r n → U ∞ r_n\mathrel{\to_{\mathcal U}}\infty r n → U ∞ . To arrange the atoms in a triangular array, let t n = k n ( k n + 1 ) / 2 ≤ r n t_n=k_n(k_n+1)/2\le r_n t n = k n ( k n + 1 ) /2 ≤ r n be the largest triangular number below r n r_n r n and retain the first t n t_n t n atoms. We set p n = ∑ i ≤ t n e i , n p_n=\sum_{i\le t_n}e_{i,n} p n = ∑ i ≤ t n e i , n . Since the discarded atoms have the smallest ranks, their total trace satisfies
τ d n ( 1 − p n ) ≤ r n − t n r n ≤ k n r n ≤ 2 r n → U 0. \tau_{d_n}(1-p_n)\le\frac{r_n-t_n}{r_n}\le\frac{k_n}{r_n}\le\sqrt{\frac{2}{r_n}}\mathrel{\to_{\mathcal U}}0. τ d n ( 1 − p n ) ≤ r n r n − t n ≤ r n k n ≤ r n 2 → U 0.
Thus this choice discards a negligible complement. We arrange the retained atoms, in their given order, in an array with row lengths k n , k n − 1 , … , 1 k_n,k_n-1,\ldots,1 k n , k n − 1 , … , 1 , filling one row at a time. Suppressing n n n , we denote them by e i j e_{ij} e ij , where 1 ≤ i ≤ k 1\le i\le k 1 ≤ i ≤ k and 1 ≤ j ≤ k + 1 − i 1\le j\le k+1-i 1 ≤ j ≤ k + 1 − i , and put m i j = rank ( e i j ) m_{ij}=\operatorname{rank}(e_{ij}) m ij = rank ( e ij ) . The ordering ensures that these ranks are nonincreasing along both rows and columns.
We now group the atoms by rows and by columns, defining
E i = ∑ j = 1 k + 1 − i e i j , F j = ∑ i = 1 k + 1 − j e i j , C 1 = ⨁ i = 1 k C E i , C 2 = ⨁ j = 1 k C F j . E_i=\sum_{j=1}^{k+1-i}e_{ij},\qquad F_j=\sum_{i=1}^{k+1-j}e_{ij},\qquad C_1=\bigoplus_{i=1}^{k}CE_i,\qquad C_2=\bigoplus_{j=1}^{k}CF_j. E i = j = 1 ∑ k + 1 − i e ij , F j = i = 1 ∑ k + 1 − j e ij , C 1 = i = 1 ⨁ k C E i , C 2 = j = 1 ⨁ k C F j .
To see why their ranks are distinct, write R i = rank ( E i ) R_i=\operatorname{rank}(E_i) R i = rank ( E i ) and S j = rank ( F j ) S_j=\operatorname{rank}(F_j) S j = rank ( F j ) . For 1 ≤ i < k 1\le i<k 1 ≤ i < k , the difference of adjacent row ranks is
R i − R i + 1 = m i , k + 1 − i + ∑ j = 1 k − i ( m i j − m i + 1 , j ) > 0 , R_i-R_{i+1}=m_{i,k+1-i}+\sum_{j=1}^{k-i}(m_{ij}-m_{i+1,j})>0, R i − R i + 1 = m i , k + 1 − i + j = 1 ∑ k − i ( m ij − m i + 1 , j ) > 0 ,
and the analogous comparison of adjacent columns gives S j > S j + 1 S_j>S_{j+1} S j > S j + 1 for 1 ≤ j < k 1\le j<k 1 ≤ j < k . Hence both families have the required distinct ranks. Moreover, E i F j = e i j E_iF_j=e_{ij} E i F j = e ij at every occupied cell, so the row and column algebras together recover p B n p pB_np p B n p .
If r n r_n r n is instead bounded on a set in U \mathcal{U} U , we can make the atom ranks distinct by removing only a bounded number of dimensions. After restricting to that set, say r n ≤ R r_n\le R r n ≤ R , we choose e i , n ′ ≤ e i , n e'_{i,n}\le e_{i,n} e i , n ′ ≤ e i , n with
rank ( e i , n ′ ) = max { m i , n − i , 0 } , p n = ∑ i e i , n ′ . \operatorname{rank}(e'_{i,n})=\max\{m_{i,n}-i,0\},\qquad p_n=\sum_i e'_{i,n}. rank ( e i , n ′ ) = max { m i , n − i , 0 } , p n = i ∑ e i , n ′ .
The positive retained ranks are now strictly decreasing, and at most R ( R + 1 ) / 2 R(R+1)/2 R ( R + 1 ) /2 dimensions have been discarded. Consequently τ d n ( 1 − p n ) → U 0 \tau_{d_n}(1-p_n)\to_{\mathcal{U}}0 τ d n ( 1 − p n ) → U 0 , and in this case we may simply take C 1 , n = C 2 , n = p n B n p n C_{1,n}=C_{2,n}=p_nB_np_n C 1 , n = C 2 , n = p n B n p n .
In either case we obtain internal abelian algebras by setting
B i = [ C i , n ⊕ C ( 1 − p n ) ] U ( i = 1 , 2 ) . B_i=[C_{i,n}\oplus\mathbb{C}(1-p_n)]_{\mathcal{U}}\qquad(i=1,2). B i = [ C i , n ⊕ C ( 1 − p n ) ] U ( i = 1 , 2 ) .
For B 1 B_1 B 1 , the distinct atom ranks allow us to apply Lemma 3.1(2) to C 1 , n C_{1,n} C 1 , n in p n M d n p n p_nM_{d_n}p_n p n M d n p n . We extend the resulting unitaries by the identity on 1 − p n 1-p_n 1 − p n , denote the extensions by u 1 , n , u 2 , n u_{1,n},u_{2,n} u 1 , n , u 2 , n , and put u j = [ u j , n ] U u_j=[u_{j,n}]_{\mathcal{U}} u j = [ u j , n ] U for j = 1 , 2 j=1,2 j = 1 , 2 . If p n = 0 p_n=0 p n = 0 , we choose arbitrary coordinate unitaries; these choices are irrelevant in the ultraproduct because τ d n ( p n ) → U 1 \tau_{d_n}(p_n)\to_{\mathcal{U}}1 τ d n ( p n ) → U 1 .
It remains to pass the corner estimate to the ambient algebra. Multiplying it by τ d n ( p n ) \tau_{d_n}(p_n) τ d n ( p n ) expresses it in the ambient 2-norm. Moreover, compression changes a bounded x n x_n x n by at most 2 ∥ x n ∥ τ d n ( 1 − p n ) 2\lVert x_n\rVert\sqrt{\tau_{d_n}(1-p_n)} 2 ∥ x n ∥ τ d n ( 1 − p n ) in 2-norm. Since the unitaries reduce p n p_n p n and the compression error tends to zero, the estimate from Lemma 3.1(2) passes to the ultraproduct. We obtain B 1 = { u 1 , u 2 } ′ B_1=\{u_1,u_2\}' B 1 = { u 1 , u 2 } ′ , and ( u 1 , u 2 ) (u_1,u_2) ( u 1 , u 2 ) has spectral gap with constant 8. Applying the same construction to C 2 , n C_{2,n} C 2 , n gives v 1 , v 2 v_1,v_2 v 1 , v 2 with commutant B 2 B_2 B 2 and the same spectral gap constant.
We finally verify that these two internal algebras recover the original one. The unit balls of H n = p n B n p n ⊕ C ( 1 − p n ) H_n=p_nB_np_n\oplus\mathbb{C}(1-p_n) H n = p n B n p n ⊕ C ( 1 − p n ) and B n B_n B n have Hausdorff 2-distance at most 2 τ d n ( 1 − p n ) 2\sqrt{\tau_{d_n}(1-p_n)} 2 τ d n ( 1 − p n ) . Indeed, compression sends a contraction in B n B_n B n to a nearby contraction in H n H_n H n . In the other direction, we extend the scalar on each retained part of an atom to the whole atom, choosing zero on discarded atoms. Since the discarded trace tends to zero, we have [ H n ] U = B [H_n]_{\mathcal{U}}=B [ H n ] U = B and B 1 , B 2 ⊆ B B_1,B_2\subseteq B B 1 , B 2 ⊆ B .
The two coordinate algebras commute and generate H n H_n H n . To pass this generation statement to their commutants, let P i , n = E ( C i , n ⊕ C ( 1 − p n ) ) ′ P_{i,n}=E(C_{i,n}\oplus\mathbb{C}(1-p_n))' P i , n = E ( C i , n ⊕ C ( 1 − p n ) ) ′ for i = 1 , 2 i=1,2 i = 1 , 2 , with commutants taken in M d n M_{d_n} M d n . Since the coordinate algebras commute, these expectations commute and have product E H n ′ E_{H'_n} E H n ′ . Therefore Lemma 2.1, together with the estimate
∥ x n − P 1 , n P 2 , n x n ∥ 2 ≤ ∥ x n − P 1 , n x n ∥ 2 + ∥ x n − P 2 , n x n ∥ 2 \lVert x_n-P_{1,n}P_{2,n}x_n\rVert_2\le\lVert x_n-P_{1,n}x_n\rVert_2+\lVert x_n-P_{2,n}x_n\rVert_2 ∥ x n − P 1 , n P 2 , n x n ∥ 2 ≤ ∥ x n − P 1 , n x n ∥ 2 + ∥ x n − P 2 , n x n ∥ 2
gives B 1 ′ ∩ B 2 ′ = B ′ B'_1\cap B'_2=B' B 1 ′ ∩ B 2 ′ = B ′ . Taking commutants and using B ′ ′ = B B''=B B ′′ = B proves (4.1). When the original algebra is a MASA, the final assertion follows from B ′ = B B'=B B ′ = B .
Preservation of internal algebras
The encoding reduces preservation in the abelian case to spectral gap internality and relative joins of commuting internal abelian algebras.
Proposition 4.2. Every isomorphism θ : M U , d → M V , e \theta:M_{\mathcal{U},d}\to M_{\mathcal{V},e} θ : M U , d → M V , e of tracial matrix ultraproducts maps internal abelian algebras to internal abelian algebras.
Proof. Let θ \theta θ be such an isomorphism and let B B B be internal and abelian. Since the ambient algebras are finite factors, θ \theta θ preserves the trace and hence the spectral gap estimates. We choose B 1 B_1 B 1 , B 2 B_2 B 2 and u 1 u_1 u 1 , u 2 u_2 u 2 , v 1 v_1 v 1 , v 2 v_2 v 2 as in Proposition 4.1 so that we can transfer their commutant descriptions to the target. The pair ( θ ( u 1 ) , θ ( u 2 ) ) (\theta(u_1),\theta(u_2)) ( θ ( u 1 ) , θ ( u 2 )) then has spectral gap there, with commutant θ ( B 1 ) \theta(B_1) θ ( B 1 ) . Theorem 2.2 therefore shows that θ ( B 1 ) \theta(B_1) θ ( B 1 ) is internal, and the same argument applies to θ ( B 2 ) \theta(B_2) θ ( B 2 ) . The two images are abelian and still commute, so Proposition 2.3 makes their relative join internal. By (4.1), this join is exactly θ ( B ) \theta(B) θ ( B ) , which proves the assertion.
For a general internal algebra, we apply the abelian result to its center. The following relative commutant description then recovers the algebra inside the commutant of that center.
Lemma 4.3. Let A A A be an internal subalgebra of a tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},d} M U , d , and put Z = Z ( A ) Z=Z(A) Z = Z ( A ) . Then there are u 1 , u 2 ∈ U ( A ′ ) ⊆ U ( Z ′ ) u_1,u_2 \in U(A') \subseteq U(Z') u 1 , u 2 ∈ U ( A ′ ) ⊆ U ( Z ′ ) such that
A = Z ′ ∩ { u 1 , u 2 } ′ . A=Z' \cap\{u_1,u_2\}'. A = Z ′ ∩ { u 1 , u 2 } ′ . and ( u 1 , u 2 ) (u_1,u_2) ( u 1 , u 2 ) has spectral gap in Z ′ Z' Z ′ with constant 2.
Proof. We write A = [ A n ] U A=[A_n]_{\mathcal{U}} A = [ A n ] U and use the multiplicity spaces in the coordinate decomposition to choose the required unitaries:
A n = ⨁ i ( M a i , n ⊗ 1 b i , n ) , A_n=\bigoplus_i(M_{a_{i,n}}\otimes1_{b_{i,n}}), A n = i ⨁ ( M a i , n ⊗ 1 b i , n ) ,
u j , n = ⨁ i ( 1 a i , n ⊗ v j , b i , n ) . u_{j,n}=\bigoplus_i(1_{a_{i,n}}\otimes v_{j,b_{i,n}}). u j , n = i ⨁ ( 1 a i , n ⊗ v j , b i , n ) .
On Z ( A n ) ′ = ⨁ i M a i , n b i , n Z(A_n)'=\bigoplus_i M_{a_{i,n}b_{i,n}} Z ( A n ) ′ = ⨁ i M a i , n b i , n , we can apply Lemma 3.1(1) in each multiplicity factor. Tensoring its estimate with the identity on the first factors and summing with the inherited trace weights gives the coordinate spectral gap estimate. By Lemma 2.1, the central commutants have ultraproduct Z ′ Z' Z ′ , while the coordinate expectations onto A n A_n A n induce E A E_A E A . Thus both sides of the estimate pass to the ultraproduct, giving the asserted spectral gap for u j = [ u j , n ] U ∈ U ( A ′ ) u_j=[u_{j,n}]_{\mathcal{U}}\in U(A') u j = [ u j , n ] U ∈ U ( A ′ ) .
Proof of Theorem D. Let θ : M U , d → M V , e \theta:\mathcal{M}_{\mathcal{U},d}\to\mathcal{M}_{\mathcal{V},e} θ : M U , d → M V , e be an isomorphism and let A ⊆ M U , d A\subseteq\mathcal{M}_{\mathcal{U},d} A ⊆ M U , d be internal. We first apply the abelian case to its center Z = Z ( A ) Z=Z(A) Z = Z ( A ) , which is internal by Lemma 2.1. Proposition 4.2 then shows that θ ( Z ) \theta(Z) θ ( Z ) is internal, and consequently its commutant θ ( Z ) ′ \theta(Z)' θ ( Z ) ′ is internal as well. We can now transfer the description in Lemma 4.3 to this internal algebra. It expresses θ ( A ) \theta(A) θ ( A ) as the relative commutant of two unitaries whose spectral gap estimate is preserved by the isomorphism. Theorem 2.2 therefore implies that θ ( A ) \theta(A) θ ( A ) is internal. Applying the same argument to θ − 1 \theta^{-1} θ − 1 proves the converse.
The same relative description also extends the result on commuting internal algebras beyond the abelian case.
Corollary 4.4. If A , B A,B A , B are commuting internal subalgebras of a tracial matrix ultraproduct M U , d \mathcal{M}_{\mathcal{U},d} M U , d , then A ′ ∩ B ′ A'\cap B' A ′ ∩ B ′ and A ∨ r e l B A\vee_{\mathrm{rel}}B A ∨ rel B are internal.
Proof. We first put the two commutant descriptions in a common internal algebra. Since the centers are commuting internal abelian algebras, P 0 = Z ( A ) ′ ∩ Z ( B ) ′ P_0=Z(A)'\cap Z(B)' P 0 = Z ( A ) ′ ∩ Z ( B ) ′ is internal by Proposition 2.3. Applying Lemma 4.3 to A ′ A' A ′ and B ′ B' B ′ , and using Z ( A ′ ) = Z ( A ) Z(A')=Z(A) Z ( A ′ ) = Z ( A ) and Z ( B ′ ) = Z ( B ) Z(B')=Z(B) Z ( B ′ ) = Z ( B ) , gives u 1 , u 2 ∈ U ( A ) u_1,u_2\in U(A) u 1 , u 2 ∈ U ( A ) and v 1 , v 2 ∈ U ( B ) v_1,v_2\in U(B) v 1 , v 2 ∈ U ( B ) that describe A ′ A' A ′ and B ′ B' B ′ inside their respective central commutants, with spectral gap constant 2. All four unitaries belong to P 0 P_0 P 0 , where both spectral gap estimates apply.
To do this, consider the expectations E A ′ E_{A'} E A ′ and E B ′ E_{B'} E B ′ . They commute and preserve P 0 P_0 P 0 , since they project onto the invariant vectors of the commuting conjugation actions of U ( A ) U(A) U ( A ) and U ( B ) U(B) U ( B ) . Their product is therefore the expectation onto C = A ′ ∩ B ′ C=A'\cap B' C = A ′ ∩ B ′ . Combining this fact with the two spectral gap estimates gives, for x ∈ P 0 x\in P_0 x ∈ P 0 ,
∥ x − E C x ∥ 2 2 ≤ ∥ x − E A ′ x ∥ 2 2 + ∥ x − E B ′ x ∥ 2 2 ≤ 2 ∑ j = 1 2 ( ∥ [ x , u j ] ∥ 2 2 + ∥ [ x , v j ] ∥ 2 2 ) . \lVert x-E_Cx\rVert_2^2\leq\lVert x-E_{A'}x\rVert_2^2+\lVert x-E_{B'}x\rVert_2^2\leq2\sum_{j=1}^{2}\left(\lVert[x,u_j]\rVert_2^2+\lVert[x,v_j]\rVert_2^2\right). ∥ x − E C x ∥ 2 2 ≤ ∥ x − E A ′ x ∥ 2 2 + ∥ x − E B ′ x ∥ 2 2 ≤ 2 j = 1 ∑ 2 ( ∥[ x , u j ] ∥ 2 2 + ∥[ x , v j ] ∥ 2 2 ) .
It follows that u 1 , u 2 , v 1 , v 2 u_1,u_2,v_1,v_2 u 1 , u 2 , v 1 , v 2 have spectral gap in P 0 P_0 P 0 , with commutant C = P 0 ∩ { u 1 , u 2 , v 1 , v 2 } ′ C=P_0\cap\{u_1,u_2,v_1,v_2\}' C = P 0 ∩ { u 1 , u 2 , v 1 , v 2 } ′ . We can therefore apply Theorem 2.2 to conclude that C C C is internal. Lemma 2.1 then gives internality of its commutant C ′ = A ∨ r e l B C'=A\mathbin{\vee_{\mathrm{rel}}}B C ′ = A ∨ rel B as well.
Definability of the diagonal and its normalizer
To pass from isomorphisms to elementary equivalence, we express the constructions of § 3 and § 4 by uniform distance formulas. Spectral gap defines the required commutants, and the normalizer estimate then allows us to interpret the Weyl group. Internality and conjugacy ensure that the diagonal recovered in this way can be replaced by the canonical one.
Distance formulas and internal algebras
Write δ H ( x ) = dist 2 ( x , H ≤ 1 ) = ∥ x − E H x ∥ 2 \delta_H(x)=\operatorname{dist}_2(x,H_{\leq1})=\lVert x-E_Hx\rVert_2 δ H ( x ) = dist 2 ( x , H ≤ 1 ) = ∥ x − E H x ∥ 2 for a unital von Neumann subalgebra H H H of a tracial algebra M M M . All algebra variables below range over M ≤ 1 M_{\leq1} M ≤ 1 . To quantify over a definable set, we replace the restriction by a distance penalty. For an L L L -Lipschitz function f f f and a nonempty subset S S S of this ball, we have
inf s ∈ S f ( s ) = inf ∥ y ∥ ≤ 1 ( f ( y ) + L dist 2 ( y , S ) ) , sup s ∈ S f ( s ) = sup ∥ y ∥ ≤ 1 ( f ( y ) − L dist 2 ( y , S ) ) . \inf_{s\in S}f(s)=\inf_{\lVert y\rVert\leq1}\left(f(y)+L\operatorname{dist}_2(y,S)\right),\qquad\sup_{s\in S}f(s)=\sup_{\lVert y\rVert\leq1}\left(f(y)-L\operatorname{dist}_2(y,S)\right). s ∈ S inf f ( s ) = ∥ y ∥ ≤ 1 inf ( f ( y ) + L dist 2 ( y , S ) ) , s ∈ S sup f ( s ) = ∥ y ∥ ≤ 1 sup ( f ( y ) − L dist 2 ( y , S ) ) .
Both identities follow by taking approximate nearest points. Thus a formula for the distance to a set lets us replace quantification over that set by quantification over the whole ball [10 Section 3.2] .
Lemma 5.1. The following distance identities hold in a tracial von Neumann algebra M M M .
(1) Let P ⊆ M P\subseteq M P ⊆ M be a unital von Neumann subalgebra, and let u = ( u 1 , … , u h ) ∈ U ( P ) h \mathbf{u}=(u_1,\ldots,u_h)\in U(P)^h u = ( u 1 , … , u h ) ∈ U ( P ) h have spectral gap in P P P with constant L L L . For C = P ∩ { u } ′ C=P\cap\{\mathbf{u}\}' C = P ∩ { u } ′ ,
δ C ( x ) = inf ∥ y ∥ ≤ 1 ( ∥ x − y ∥ 2 + δ P ( y ) + L ( ∑ j = 1 h ∥ [ u j , y ] ∥ 2 2 ) 1 / 2 ) . \delta_C(x)=\inf_{\lVert y\rVert\leq1}\left(\lVert x-y\rVert_2+\delta_P(y)+\sqrt{L}\left(\sum_{j=1}^{h}\lVert[u_j,y]\rVert_2^2\right)^{1/2}\right). δ C ( x ) = ∥ y ∥ ≤ 1 inf ∥ x − y ∥ 2 + δ P ( y ) + L ( j = 1 ∑ h ∥[ u j , y ] ∥ 2 2 ) 1/2 . (2) If H 1 , … , H r ⊆ M H_1,\ldots,H_r\subseteq M H 1 , … , H r ⊆ M are commuting unital von Neumann subalgebras and K = ⋂ i H i ′ K=\bigcap_i H_i' K = ⋂ i H i ′ , then
σ i ( x ) = sup ∥ y ∥ ≤ 1 ( ∥ [ x , y ] ∥ 2 2 − 2 δ H i ( y ) ) , δ K ( x ) = inf ∥ y ∥ ≤ 1 ( ∥ x − y ∥ 2 + ∑ i σ i ( y ) ) . \sigma_i(x)=\sup_{\lVert y\rVert\leq1}\left(\lVert[x,y]\rVert_2^2-2\delta_{H_i}(y)\right),\qquad\delta_K(x)=\inf_{\lVert y\rVert\leq1}\left(\lVert x-y\rVert_2+\sum_i\sigma_i(y)\right). σ i ( x ) = ∥ y ∥ ≤ 1 sup ( ∥[ x , y ] ∥ 2 2 − 2 δ H i ( y ) ) , δ K ( x ) = ∥ y ∥ ≤ 1 inf ( ∥ x − y ∥ 2 + i ∑ σ i ( y ) ) . In both cases, finite formulas for the input distances give a finite formula for the resulting distance over the same parameters and, in (1), the tuple u \mathbf{u} u .
Proof. For (1), we first bound the distance to the relative commutant. Bimodularity of E P E_P E P allows us to apply spectral gap inside the subalgebra, and Pythagoras then gives δ C ( x ) 2 ≤ δ P ( x ) 2 + L ∑ j ∥ [ u j , x ] ∥ 2 2 \delta_C(x)^2\leq\delta_P(x)^2+L\sum_j\lVert[u_j,x]\rVert_2^2 δ C ( x ) 2 ≤ δ P ( x ) 2 + L ∑ j ∥[ u j , x ] ∥ 2 2 . Combining this estimate with the triangle inequality shows that every term in the displayed infimum is at least δ C ( x ) \delta_C(x) δ C ( x ) . Conversely, the choice y = E C x y=E_Cx y = E C x attains this value, proving the formula; compare [18 Lemma 4.1] .
For (2), the penalty identity first restricts each supremum to the given algebra, yielding σ i ( x ) = sup y ∈ ( H i ) ≤ 1 ∥ [ x , y ] ∥ 2 2 \sigma_i(x)=\sup_{y\in(H_i)_{\leq1}}\lVert[x,y]\rVert_2^2 σ i ( x ) = sup y ∈ ( H i ) ≤ 1 ∥[ x , y ] ∥ 2 2 . Averaging unitary conjugates then gives ∥ x − E H i ′ x ∥ 2 ≤ σ i ( x ) \lVert x-E_{H_i'}x\rVert_2\leq\sigma_i(x) ∥ x − E H i ′ x ∥ 2 ≤ σ i ( x ) , as in [16 Corollary 5.12] . Since the algebras commute, the expectations onto their commutants commute and have product E K E_K E K . Telescoping this product therefore gives ∥ x − E K x ∥ 2 ≤ ∑ i σ i ( x ) \lVert x-E_Kx\rVert_2\leq\sum_i\sigma_i(x) ∥ x − E K x ∥ 2 ≤ ∑ i σ i ( x ) . We now argue as in (1): the triangle inequality gives the lower bound for the infimum, and conditional expectation onto the joint commutant gives equality.
Proposition 5.2. Every internal abelian subalgebra of a tracial matrix ultraproduct M u , d \mathcal{M}_{\mathbf{u},d} M u , d is definable over at most four unitary parameters, and every internal subalgebra is definable over at most six unitary parameters. Their unit-ball distance predicates are given by fixed finite formulas. The formulas are independent of the algebra, dimensions, and ultrafilter.
Proof. For an internal abelian algebra B B B , Proposition 4.1 writes B = ( B 1 ′ ∩ B 2 ′ ) ′ B=(B_1'\cap B_2')' B = ( B 1 ′ ∩ B 2 ′ ) ′ , where the commuting abelian algebras B i B_i B i are commutants of unitary pairs with spectral gap constant 8 8 8 . Lemma 5.1 (1) therefore gives their distance formulas over the four parameters. Applying (2) first to B 1 , B 2 B_1,B_2 B 1 , B 2 and then to their joint commutant gives the distance to B B B without introducing further parameters. For arbitrary internal A A A , we apply the abelian construction to Z ( A ) Z(A) Z ( A ) . Part (2) then gives the distance to P = Z ( A ) ′ P = Z(A)' P = Z ( A ) ′ over the same four parameters. By Lemma 4.3 , A A A is the relative commutant in P P P of two further unitaries with gap constant 2, so part (1) gives its distance formula over all six parameters. Since each step uses fixed formulas and constants, the resulting formulas have the asserted uniformity.
The diagonal code
To transfer a diagonal under elementary equivalence, we need a condition on the parameters which guarantees that they define an internal MASA. We obtain it by expressing the conditions of Lemma 3.3 by finite formulas. The row and column construction supplies tuples satisfying these conditions.
Lemma 5.3. There are fixed nonnegative finite formulas Δ ( z ) \Delta(\mathbf{z}) Δ ( z ) and δ ( x ; z ) \delta(x;\mathbf{z}) δ ( x ; z ) , with z = ( z i j ) 1 ≤ i ≤ 2 , 1 ≤ j ≤ 3 \mathbf{z} = (z_{ij})_{1\leq i\leq2,\,1\leq j\leq3} z = ( z ij ) 1 ≤ i ≤ 2 , 1 ≤ j ≤ 3 , such that in a tracial matrix ultraproduct M M M , a tuple z ∈ M ≤ 1 6 \mathbf{z} \in M_{\leq1}^{6} z ∈ M ≤ 1 6 satisfies Δ ( z ) = 0 \Delta(\mathbf{z}) = 0 Δ ( z ) = 0 if and only if:
(1) the z i j z_{ij} z ij are unitary and their commutants C i = { z i 1 , z i 2 , z i 3 } ′ ∩ M C_i = \{z_{i1},z_{i2},z_{i3}\}' \cap M C i = { z i 1 , z i 2 , z i 3 } ′ ∩ M satisfy
dist 2 ( x , ( C i ) ≤ 1 ) ≤ 4 ∑ j = 1 3 ∥ [ z i j , x ] ∥ 2 ( x ∈ M ≤ 1 , i = 1 , 2 ) ; \operatorname{dist}_2(x,(C_i)_{\leq1}) \leq4\sum_{j=1}^{3}\lVert[z_{ij},x]\rVert_2 \qquad(x\in M_{\leq1},\ i=1,2); dist 2 ( x , ( C i ) ≤ 1 ) ≤ 4 j = 1 ∑ 3 ∥[ z ij , x ] ∥ 2 ( x ∈ M ≤ 1 , i = 1 , 2 ) ; (2) C 1 , C 2 C_1,C_2 C 1 , C 2 are abelian and commute, and D = C 1 ′ ∩ C 2 ′ D = C_1' \cap C_2' D = C 1 ′ ∩ C 2 ′ is abelian.
For every such zero, D D D is an internal MASA and δ ( x ; z ) = dist 2 ( x , D ≤ 1 ) \delta(x;\mathbf{z}) = \operatorname{dist}_2(x,D_{\leq1}) δ ( x ; z ) = dist 2 ( x , D ≤ 1 ) . In M d ( C ) M_d(\mathbb{C}) M d ( C ) , the same zero-set characterization and distance identity hold, with D D D unitarily conjugate to D d D_d D d .
Proof. We encode the distance estimates and the commutation conditions separately. For i = 1 , 2 i=1,2 i = 1 , 2 , define the following formulas on contractions:
η ( z ) = max i , j ∥ z i j ∗ z i j − 1 ∥ 2 , ϵ i ( x ) = ∑ j = 1 3 ∥ [ z i j , x ] ∥ 2 , \eta(\mathbf{z})=\max_{i,j}\lVert z_{ij}^{*}z_{ij}-1\rVert_2,
\qquad
\epsilon_i(x)=\sum_{j=1}^{3}\lVert[z_{ij},x]\rVert_2, η ( z ) = i , j max ∥ z ij ∗ z ij − 1 ∥ 2 , ϵ i ( x ) = j = 1 ∑ 3 ∥[ z ij , x ] ∥ 2 ,
\delta_i(x)=\inf_{\lVert y\rVert\leq1}\bigl(\lVert x-y\rVert_2+4\epsilon_i(y)\bigr),
\qquad
\sigma_i(x)=\sup_{\lVert y\rVert\leq1}\bigl(\lVert[x,y]\rVert_2\mathbin{\dotminus}2\delta_i(y)\bigr),
\gamma_i(\mathbf{z})=\sup_{\lVert x\rVert\leq1}\inf_{\lVert y\rVert\leq1}\max\{\epsilon_i(y),\lVert x-y\rVert_2\mathbin{\dotminus}4\epsilon_i(x)\},
\qquad
\kappa_{ij}(\mathbf{z})=\sup_{\lVert x\rVert\leq1}\bigl(\sigma_i(x)\mathbin{\dotminus}2\delta_j(x)\bigr).
Here κ i j \kappa_{ij} κ ij is indexed by 1 ≤ i , j ≤ 2 1\leq i,j\leq2 1 ≤ i , j ≤ 2 . To recover the joint commutant and impose its abelianness, put
δ ( x ; z ) = inf ∥ y ∥ ≤ 1 ( ∥ x − y ∥ 2 + σ 1 ( y ) + σ 2 ( y ) ) , \delta(x;\mathbf{z})=\inf_{\lVert y\rVert\leq1}\bigl(\lVert x-y\rVert_2+\sigma_1(y)+\sigma_2(y)\bigr), δ ( x ; z ) = ∥ y ∥ ≤ 1 inf ( ∥ x − y ∥ 2 + σ 1 ( y ) + σ 2 ( y ) ) ,
\beta(\mathbf{z})=\sup_{\lVert x\rVert,\lVert y\rVert\leq1}\bigl(\lVert[x,y]\rVert_2\mathbin{\dotminus}2(\delta(x;\mathbf{z})+\delta(y;\mathbf{z}))\bigr),
Δ ( z ) = 1 2 max { η ( z ) , max i γ i ( z ) , max i , j κ i j ( z ) , β ( z ) } . \Delta(\mathbf{z})=\frac{1}{2}\max\{\eta(\mathbf{z}),\max_i\gamma_i(\mathbf{z}),\max_{i,j}\kappa_{ij}(\mathbf{z}),\beta(\mathbf{z})\}. Δ ( z ) = 2 1 max { η ( z ) , i max γ i ( z ) , i , j max κ ij ( z ) , β ( z )} .
We verify the zero-set characterization in the order of the construction. The condition η = 0 \eta=0 η = 0 gives unitarity, while compactness in a matrix algebra, or countable saturation in an ultraproduct, makes γ i = 0 \gamma_i=0 γ i = 0 equivalent to the distance estimate in (1). Under this estimate, the triangle inequality gives δ i = δ C i \delta_i=\delta_{C_i} δ i = δ C i ; squaring the estimate also gives spectral gap with constant 48. We can therefore use the penalty identities to identify κ i j = 0 \kappa_{ij}=0 κ ij = 0 with commutation of C i , C j C_i,C_j C i , C j . Lemma 5.1 (2) then gives the distance formula for the joint commutant, so β = 0 \beta=0 β = 0 is exactly abelianness of D D D . These identifications prove both directions of the characterization.
It remains to identify the resulting algebra. By Lemma 3.3 , D D D is an internal MASA in the ultraproduct case. In a matrix algebra, C i ⊆ D C_i\subseteq D C i ⊆ D implies D ′ ⊆ D D'\subseteq D D ′ ⊆ D ; since the joint commutant is abelian, it is a MASA, and simultaneous diagonalization gives the asserted conjugacy. □
Proof of Theorem B. To show that the zero-set is nonempty, use Lemma 3.2 to choose two unitary triples with spectral gap constant 10, whose commutants B 1 , B 2 B_1,B_2 B 1 , B 2 satisfy B 1 ′ ∩ B 2 ′ = D B_1'\cap B_2'=D B 1 ′ ∩ B 2 ′ = D for the canonical diagonal D D D . Their spectral gap estimates imply the required distance bounds, and hence the triples give a zero of Δ \Delta Δ . The characterization in Lemma 5.3 applies to every zero, while the conjugacy assertion follows from Lemma 2.1 . Together these give the remaining claims. □
The Weyl group
The normalizer estimate turns the distance to the diagonal into the distance to its normalizer. The same diagonal predicate recovers the Hamming metric on the quotient, giving a uniform interpretation of the Weyl group and its action.
Proposition 5.4. Let D D D be an internal MASA in a tracial matrix ultraproduct M M M , and put δ D ( x ) = dist 2 ( x , D ≤ 1 ) \delta_D(x)=\operatorname{dist}_2(x,D_{\leq1}) δ D ( x ) = dist 2 ( x , D ≤ 1 ) . The formulas
ξ D ( v ) = sup ∥ y ∥ ≤ 1 ( δ D ( v y v ∗ ) − δ D ( y ) ) , \xi_D(v)=\sup_{\lVert y\rVert\leq1}\left(\delta_D(vyv^*)-\delta_D(y)\right), ξ D ( v ) = ∥ y ∥ ≤ 1 sup ( δ D ( v y v ∗ ) − δ D ( y ) ) , ν ( x ) = inf ∥ v ∥ ≤ 1 ( ∥ x − v ∥ 2 + 2 ξ D ( v ) + ( 1 + 2 2 ) ∥ v ∗ v − 1 ∥ 2 ) , \nu(x)=\inf_{\lVert v\rVert\leq1}\left(\lVert x-v\rVert_2+\sqrt{2}\xi_D(v)+(1+2\sqrt{2})\lVert v^*v-1\rVert_2\right), ν ( x ) = ∥ v ∥ ≤ 1 inf ( ∥ x − v ∥ 2 + 2 ξ D ( v ) + ( 1 + 2 2 ) ∥ v ∗ v − 1 ∥ 2 ) , d ( u ˉ , v ˉ ) = δ D ( u ∗ v ) 2 ( u , v ∈ N M ( D ) ) d(\bar{u},\bar{v})=\delta_D(u^*v)^2 \qquad(u,v\in N_M(D)) d ( u ˉ , v ˉ ) = δ D ( u ∗ v ) 2 ( u , v ∈ N M ( D )) give the distance ν ( x ) = dist 2 ( x , N M ( D ) ) \nu(x)=\operatorname{dist}_2(x,N_M(D)) ν ( x ) = dist 2 ( x , N M ( D )) and the normalized Hamming metric on N M ( D ) / U ( D ) N_M(D)/U(D) N M ( D ) / U ( D ) . Thus this group and its action on D D D are uniformly interpretable in ( M , D ) (M,D) ( M , D ) . The same assertions hold for D d ⊆ M d D_d\subseteq M_d D d ⊆ M d .
Proof. We first recover the normalizer distance. The penalty identity restricts the supremum to the diagonal and gives ξ D ( v ) = sup y ∈ D ≤ 1 δ D ( v y v ∗ ) \xi_D(v)=\sup_{y\in D_{\leq1}}\delta_D(vyv^*) ξ D ( v ) = sup y ∈ D ≤ 1 δ D ( v y v ∗ ) . For unitary v v v , this is the normalizer defect in Lemma 3.4, so its distance to N M ( D ) N_M(D) N M ( D ) is at most 2 ξ D ( v ) \sqrt{2}\xi_D(v) 2 ξ D ( v ) . Since the defect vanishes on the normalizer, the infimum over unitaries of ∥ x − v ∥ 2 + 2 ξ D ( v ) \lVert x-v\rVert_2+\sqrt{2}\xi_D(v) ∥ x − v ∥ 2 + 2 ξ D ( v ) is precisely the required distance.
To remove the unitary restriction, note that this integrand is ( 1 + 2 2 ) (1+2\sqrt{2}) ( 1 + 2 2 ) -Lipschitz in v v v , while polar decomposition gives dist 2 ( v , U ( M ) ) ≤ ∥ v ∗ v − 1 ∥ 2 \operatorname{dist}_2(v,U(M))\leq\lVert v^*v-1\rVert_2 dist 2 ( v , U ( M )) ≤ ∥ v ∗ v − 1 ∥ 2 for contractions. Thus the penalty bounds the change in the integrand when rounding to a unitary, which proves the formula for ν \nu ν .
To identify the quotient metric, consider a coordinate normalizer w = σ z w=\sigma z w = σ z , with z z z diagonal unitary. Its diagonal entries give δ D d ( w ) 2 = 1 − ∥ E D d w ∥ 2 2 = d ( σ , 1 ) \delta_{D_d}(w)^2=1-\lVert E_{D_d}w\rVert_2^2=d(\sigma,1) δ D d ( w ) 2 = 1 − ∥ E D d w ∥ 2 2 = d ( σ , 1 ) . By Lemma 3.4, coordinate normalizers represent every normalizer in the ultraproduct, so this identity passes to the ultraproduct and gives the formula for d d d . Distance penalties now interpret the quotient as an imaginary sort [7 Section 11] . Its conjugation action on D D D is well defined because U ( D ) U(D) U ( D ) acts trivially there. □
Upgrading to elementary equivalence
We can now transfer the type of a diagonal code between elementarily equivalent matrix ultraproducts.
Proof of Theorem C(2). Let M = M U , d M=\mathcal{M}_{U,d} M = M U , d and P = M V , e P=\mathcal{M}_{V,e} P = M V , e be elementarily equivalent, with canonical diagonals D , F D,F D , F . We recover the target diagonal from the type of a tuple coding the source diagonal. By Theorem B and unitary conjugacy, we can choose z z z in M M M with Δ ( z ) = 0 \Delta(z)=0 Δ ( z ) = 0 such that δ ( ⋅ ; z ) \delta(\cdot;z) δ ( ⋅ ; z ) is the distance to D ≤ 1 D_{\leq1} D ≤ 1 . Elementary equivalence and countable saturation then give a tuple z ~ \widetilde{z} z in P P P with the same complete type as z z z .
In particular, Δ ( z ~ ) = 0 \Delta(\widetilde{z})=0 Δ ( z ) = 0 , so Theorem B identifies δ ( ⋅ ; z ~ ) \delta(\cdot;\widetilde{z}) δ ( ⋅ ; z ) as the distance to an internal MASA D ~ ⊆ P \widetilde{D}\subseteq P D ⊆ P , unitarily conjugate to F F F . We can now compare sentences in the expanded pairs: replacing the diagonal predicate by δ ( ⋅ ; z ) \delta(\cdot;z) δ ( ⋅ ; z ) turns each such sentence into a formula in the coding tuple. Equality of the tuple types therefore gives ( M , D ) ≡ ( P , D ~ ) (M,D)\equiv(P,\widetilde{D}) ( M , D ) ≡ ( P , D ) , and conjugating the target diagonal gives ( M , D ) ≡ ( P , F ) (M,D)\equiv(P,F) ( M , D ) ≡ ( P , F ) , as required. □
Proof of Theorem C(3). By part (2), the expanded diagonal pairs are elementarily equivalent. The uniform interpretation in Proposition 5.4 therefore gives elementary equivalence of their Weyl groups with the metric d d d . Under the identification (6) , these are the associated universal sofic groups, proving the assertion. □
Proof of Theorem A. Consider the full sequence Sym ( n ) \operatorname{Sym}(n) Sym ( n ) . By [3 Theorem 5.6 and Remark 5.7] , we can choose continuum many ultrafilters for which the metric ultraproducts have pairwise distinct theories. By Theorem C(3), elementary equivalence of two corresponding tracial matrix ultraproducts would imply elementary equivalence of their universal sofic groups. Hence the matrix ultraproducts have pairwise distinct theories.
We pass to unitary groups using the reconstruction argument and the Keisler–Shelah theorem: for tracial matrix ultraproducts, elementary equivalence of the unitary groups as metric groups implies elementary equivalence of the algebras; see [19 p. 104] . Combining this implication with the polar-decomposition identification used in Theorem A′, we obtain continuum many distinct theories of metric ultraproducts of ( U ( n ) , d 2 ) (U(n),d_2) ( U ( n ) , d 2 ) . Since both languages are separable, continuum is also the maximum possible number of complete theories. Finally, if either sequence of theories converged, Łoś’ theorem would make all its nonprincipal ultraproducts elementarily equivalent, contradicting what we have just proved. Thus neither sequence converges. □
Recovering left-right actions
Following the method of [3 Section 5] , we realize algebraic ultraproducts of PSL 2 ( 2 r ) \operatorname{PSL}_2(2^r) PSL 2 ( 2 r ) as definable centralizers in the Weyl group. Transferring their defining conditions recovers finite groups whose orders are asymptotic to the matrix dimensions. Element orders then distinguish an explicit family and rule out its nontrivial amplifications, proving Theorem E .
Admissible tuples and their transfer
Recall that G r = PSL 2 ( 2 r ) G_r=\operatorname{PSL}_2(2^r) G r = PSL 2 ( 2 r ) and ∣ G r ∣ = 2 r ( 2 2 r − 1 ) \lvert G_r\rvert=2^r(2^{2r}-1) ∣ G r ∣ = 2 r ( 2 2 r − 1 ) , for r ≥ 2 r\ge2 r ≥ 2 . For a tracial matrix ultraproduct M M M with canonical diagonal D D D , write W = N M ( D ) / U ( D ) W=N_M(D)/U(D) W = N M ( D ) / U ( D ) , with normalized Hamming metric d d d and its conjugation action on D D D . For a = ( a 1 , a 2 ) ∈ W 2 a=(a_1,a_2)\in W^2 a = ( a 1 , a 2 ) ∈ W 2 , put
C ( a ) = { x ∈ W : [ x , a 1 ] = [ x , a 2 ] = 1 } , ϵ a ( x ) = max j = 1 , 2 d ( x a j , a j x ) . C(a)=\{x\in W:[x,a_1]=[x,a_2]=1\},\qquad\epsilon_a(x)=\max_{j=1,2}d(xa_j,a_jx). C ( a ) = { x ∈ W : [ x , a 1 ] = [ x , a 2 ] = 1 } , ϵ a ( x ) = j = 1 , 2 max d ( x a j , a j x ) .
Here group commutators mean [ x , y ] = x y x − 1 y − 1 [x,y]=xyx^{-1}y^{-1} [ x , y ] = x y x − 1 y − 1 . A group is a C A CA C A -group if the centralizer of every nonidentity element is abelian; a subgroup is 1 1 1 -discrete if distinct elements have distance one.
We impose a distance estimate so that the centralizers admit restricted quantification, and spectral gap so that they are discrete and the reconstructed action has a large orbit. Commutation and the group-theoretic conditions will identify the two regular actions.
Definition 6.1. For constants K , L ≥ 1 K,L\ge1 K , L ≥ 1 , a tuple ( a , t ) ∈ W 2 × W (a,t)\in W^2\times W ( a , t ) ∈ W 2 × W is admissible if, putting b j = t a j t − 1 b_j=ta_jt^{-1} b j = t a j t − 1 and c = [ b 1 , b 2 ] c=[b_1,b_2] c = [ b 1 , b 2 ] , the following hold:
(1) dist ( x , C ( a ) ) ≤ L ϵ a ( x ) \operatorname{dist}(x,C(a))\le L\epsilon_a(x) dist ( x , C ( a )) ≤ L ϵ a ( x ) for every x ∈ W x\in W x ∈ W ;
(2) ∥ f − τ ( f ) 1 ∥ 2 2 ≤ K ∑ j = 1 2 ∥ a j f a j ∗ − f ∥ 2 2 \lVert f-\tau(f)1\rVert_2^2\le K\sum_{j=1}^{2}\lVert a_jfa_j^*-f\rVert_2^2 ∥ f − τ ( f ) 1 ∥ 2 2 ≤ K ∑ j = 1 2 ∥ a j f a j ∗ − f ∥ 2 2 for every f ∈ D f\in D f ∈ D ;
(3) [ a i , b j ] = 1 [a_i,b_j]=1 [ a i , b j ] = 1 for 1 ≤ i , j ≤ 2 1\le i,j\le2 1 ≤ i , j ≤ 2 , and [ C ( a ) , C ( b ) ] = 1 [C(a),C(b)]=1 [ C ( a ) , C ( b )] = 1 ;
(4) C ( a ) C(a) C ( a ) is a C A CA C A -group and max i = 1 , 2 d ( [ c , b i c b i − 1 ] , 1 ) = 1 \max_{i=1,2}d([c,b_icb_i^{-1}],1)=1 max i = 1 , 2 d ([ c , b i c b i − 1 ] , 1 ) = 1 .
The action in (2) is well defined, since any normalizer representatives of a j a_j a j induce the same action on D D D . Conjugation by t t t gives (1) and (2) for b b b , so the two tuples satisfy the same estimates. Condition (3) also makes their centralizers mutual: b j ∈ C ( a ) b_j\in C(a) b j ∈ C ( a ) gives C W ( C ( a ) ) ⊆ C ( b ) C_W(C(a))\subseteq C(b) C W ( C ( a )) ⊆ C ( b ) , and commutation gives the reverse inclusion.
Expanding generators in the regular actions give such tuples with constants independent of the dimensions.
Lemma 6.2. There are constants K , L K,L K , L such that, for every sequence r j ≥ 2 r_j\ge2 r j ≥ 2 tending to infinity, the canonical diagonal pair
M = ∏ U M ∣ G r j ∣ , D = [ ℓ ∞ ( G r j ) ] U M=\prod_{\mathcal U}M_{\lvert G_{r_j}\rvert},\qquad D=\left[\ell^\infty(G_{r_j})\right]_{\mathcal U} M = U ∏ M ∣ G r j ∣ , D = [ ℓ ∞ ( G r j ) ] U has an admissible tuple with
C ( a ) = ∏ U a l g ρ j ( G r j ) , C ( b ) = ∏ U a l g λ j ( G r j ) . C(a)=\prod_{\mathcal U}^{\mathrm{alg}}\rho_j(G_{r_j}),\qquad C(b)=\prod_{\mathcal U}^{\mathrm{alg}}\lambda_j(G_{r_j}). C ( a ) = U ∏ alg ρ j ( G r j ) , C ( b ) = U ∏ alg λ j ( G r j ) . Here λ ( g ) x = g x \lambda(g)x=gx λ ( g ) x = g x and ρ ( g ) x = x g − 1 \rho(g)x=xg^{-1} ρ ( g ) x = x g − 1 are the regular actions.
Proof. The construction in [24 Section 2] gives expanding generating pairs ( s 1 , r , s 2 , r ) (s_{1,r},s_{2,r}) ( s 1 , r , s 2 , r ) with a uniform spectral gap constant K K K . Indeed, its generators are c c c , h c h − 1 hch^{-1} h c h − 1 , h − 1 c h h^{-1}ch h − 1 c h , with h 3 = 1 h^3=1 h 3 = 1 and second averaging eigenvalue at most 2 2 / 3 2\sqrt{2}/3 2 2 /3 . The displacement under either conjugate of c c c is at most 2 ∥ λ ( h ) f − f ∥ 2 + ∥ λ ( c ) f − f ∥ 2 2\lVert\lambda(h)f-f\rVert_2+\lVert\lambda(c)f-f\rVert_2 2 ∥ λ ( h ) f − f ∥ 2 + ∥ λ ( c ) f − f ∥ 2 , so ( h , c ) (h,c) ( h , c ) has uniform spectral gap. We take a i = [ λ j ( s i , r j ) ] U a_i=[\lambda_j(s_{i,r_j})]_{\mathcal U} a i = [ λ j ( s i , r j ) ] U and let t t t be induced by inversion; then b i = [ ρ j ( s i , r j ) ] U b_i=[\rho_j(s_{i,r_j})]_{\mathcal U} b i = [ ρ j ( s i , r j ) ] U . To control the distance to the centralizer, take u ∈ Sym ( G r ) u \in\operatorname{Sym}(G_r) u ∈ Sym ( G r ) , put a ( r ) = ( λ ( s 1 , r ) , λ ( s 2 , r ) ) a^{(r)}=(\lambda(s_{1,r}),\lambda(s_{2,r})) a ( r ) = ( λ ( s 1 , r ) , λ ( s 2 , r )) and partition G r G_r G r into E v = { g : g − 1 u ( g ) = v } E_v=\{g:g^{-1}u(g)=v\} E v = { g : g − 1 u ( g ) = v } . Writing p v = ∣ E v ∣ / ∣ G r ∣ p_v=|E_v|/|G_r| p v = ∣ E v ∣/∣ G r ∣ , we apply spectral gap to their indicator functions and obtain
dist ( u , ρ ( G r ) ) = 1 − max v p v ≤ ∑ v p v ( 1 − p v ) ≤ 2 K ∑ i = 1 2 d ( u λ ( s i , r ) , λ ( s i , r ) u ) ≤ 4 K ϵ a ( r ) ( u ) . \operatorname{dist}(u,\rho(G_r))=1-\max_v p_v\leq\sum_v p_v(1-p_v)\leq2K\sum_{i=1}^{2}d(u\lambda(s_{i,r}),\lambda(s_{i,r})u)\leq4K\epsilon_{a^{(r)}}(u). dist ( u , ρ ( G r )) = 1 − v max p v ≤ v ∑ p v ( 1 − p v ) ≤ 2 K i = 1 ∑ 2 d ( u λ ( s i , r ) , λ ( s i , r ) u ) ≤ 4 K ϵ a ( r ) ( u ) .
Thus we may take L = 4 K L=4K L = 4 K . The estimate shows that permutations which asymptotically commute with the left generators are asymptotically equal to right translations. It therefore identifies the ultraproduct centralizer with the right translations, and conjugacy gives the left translations. Together with spectral gap, this proves (1)–(3).
To verify the final condition, recall that the G r G_r G r are nonabelian simple CA-groups [ACMM25, Proposition 12.2]. For generators g 1 , g 2 g_1,g_2 g 1 , g 2 , put c = [ g 1 , g 2 ] ≠ 1 c=[g_1,g_2]\ne1 c = [ g 1 , g 2 ] = 1 . If both [ c , g i c g i − 1 ] = 1 [c,g_i c g_i^{-1}]=1 [ c , g i c g i − 1 ] = 1 , the CA property makes both generators normalize C G r ( c ) C_{G_r}(c) C G r ( c ) , which would then be a nontrivial proper normal subgroup. Since this contradicts simplicity, one commutator is nontrivial, and its regular translation has distance one from the identity, proving (4). □
Fix these constants K , L K,L K , L from now on. To transfer the construction, we express the centralizer distance estimate and the distance itself by the group formulas
γ ( a ) = sup x inf y max { ϵ a ( y ) , d ( x , y ) − L ϵ a ( x ) } , δ a ( x ) = inf y ( d ( x , y ) + L ϵ a ( y ) ) (9) \gamma(a)=\sup_x\inf_y\max\{\epsilon_a(y),d(x,y)-L\epsilon_a(x)\},\qquad\delta_a(x)=\inf_y(d(x,y)+L\epsilon_a(y)) \tag*{(9)} γ ( a ) = x sup y inf max { ϵ a ( y ) , d ( x , y ) − L ϵ a ( x )} , δ a ( x ) = y inf ( d ( x , y ) + L ϵ a ( y )) (9)
whose quantifiers range over W W W . Vanishing of the first formula expresses (1); under this condition, the second gives the distance to the centralizer. Together with discreteness, this permits transfer of the centralizer and its ordinary group theory along with the tuple.
Lemma 6.3. The centralizers of an admissible tuple are 1-discrete and uniformly definable, with distance formula δ a \delta_a δ a for C ( a ) C(a) C ( a ) . If ( M , D ) ≡ ( P , F ) (M,D)\equiv(P,F) ( M , D ) ≡ ( P , F ) are canonical diagonal pairs of tracial matrix ultraproducts, every admissible tuple in the first has a counterpart of the same type in the second, and their centralizers are elementarily equivalent as ordinary groups.
Proof. We prove discreteness by examining fixed points. For u = [ u j ] ∈ C ( a ) u=[u_j]\in C(a) u = [ u j ] ∈ C ( a ) , let p j p_j p j be the diagonal projection onto the fixed points of a permutation representative u j u_j u j . A fixed point can leave this set only where the permutations fail to commute, so permutation lifts of the a i a_i a i satisfy
∥ a i , j p j a i , j ∗ − p j ∥ 2 2 ≤ 2 d ( a i , j u j , u j a i , j ) ⟶ U 0. \lVert a_{i,j}p_ja_{i,j}^{*}-p_j\rVert_2^2\leq2d(a_{i,j}u_j,u_ja_{i,j})\longrightarrow_{\mathcal U}0. ∥ a i , j p j a i , j ∗ − p j ∥ 2 2 ≤ 2 d ( a i , j u j , u j a i , j ) ⟶ U 0.
Thus p = [ p j ] p=[p_j] p = [ p j ] is fixed by a a a , and condition (2) gives τ ( p ) ( 1 − τ ( p ) ) = 0 \tau(p)(1-\tau(p))=0 τ ( p ) ( 1 − τ ( p )) = 0 , so d ( u , 1 ) = 1 − τ ( p ) ∈ { 0 , 1 } d(u,1)=1-\tau(p)\in\{0,1\} d ( u , 1 ) = 1 − τ ( p ) ∈ { 0 , 1 } . Conjugacy proves the same assertion for C ( b ) C(b) C ( b ) .
For definability, the distance estimate (1) and the triangle inequality identify δ a \delta_a δ a with distance to C ( a ) C(a) C ( a ) . Conversely, in these countably saturated groups, γ ( a ) = 0 \gamma(a)=0 γ ( a ) = 0 expresses (1): it supplies a commuting u u u within L ϵ a ( x ) L\epsilon_a(x) L ϵ a ( x ) of each x x x . Thus this condition and the definitions of the centralizers transfer with the tuple type.
By Proposition 5.4, the Weyl group and its action on the diagonal are uniformly interpreted in the pair. Since canonical diagonal pairs are coordinate ultraproducts, they are countably saturated. We may therefore transfer the complete type of normalizer representatives of a 1 , a 2 , t a_1,a_2,t a 1 , a 2 , t to ( P , F ) (P,F) ( P , F ) . Condition (2) transfers as a universal condition on D ≤ 1 D_{\leq1} D ≤ 1 , and the fixed-point argument again gives 1-discreteness. Distance penalties also permit quantification over the centralizers. Since on a 1-discrete group ordinary group formulas are continuous formulas using distance for equality, their theories, including the CA condition, transfer as well. The same holds for the remaining commutation and nonabelian conditions in (3)–(4). □
Finite centralizers and reconstruction
The cluster argument of [3 Section 3] turns a definable, discrete centralizer into an algebraic ultraproduct of finite groups. It uses only discreteness and the centralizer distance estimate. We record this step before using spectral gap on the diagonal to recover the group orders.
Lemma 6.4. Let a = ( a 1 , a 2 ) ∈ ( ∏ u m e t Sym ( d j ) ) 2 a=(a_1,a_2)\in\left(\prod_{u}^{\mathrm{met}}\operatorname{Sym}(d_j)\right)^2 a = ( a 1 , a 2 ) ∈ ( ∏ u met Sym ( d j ) ) 2 , with permutation representatives a j = ( a 1 , j , a 2 , j ) a_j=(a_{1,j},a_{2,j}) a j = ( a 1 , j , a 2 , j ) . Suppose that dist ( x , C ( a ) ) ≤ L ϵ a ( x ) \operatorname{dist}(x,C(a))\le L\epsilon_a(x) dist ( x , C ( a )) ≤ L ϵ a ( x ) for every x x x , and that C ( a ) C(a) C ( a ) is 1 1 1 -discrete. There are finite groups H j H_j H j and maps α j : H j → Sym ( d j ) \alpha_j:H_j\to\operatorname{Sym}(d_j) α j : H j → Sym ( d j ) , with α j ( 1 ) = 1 \alpha_j(1)=1 α j ( 1 ) = 1 , such that
sup g ∈ H j ϵ a j ( α j ( g ) ) ⟶ U 0 , \sup_{g\in H_j}\epsilon_{a_j}(\alpha_j(g))\mathrel{\longrightarrow_U}0, g ∈ H j sup ϵ a j ( α j ( g )) ⟶ U 0 , sup g , h ∈ H j d ( α j ( g h ) , α j ( g ) α j ( h ) ) ⟶ U 0 , \sup_{g,h\in H_j}d(\alpha_j(gh),\alpha_j(g)\alpha_j(h))\mathrel{\longrightarrow_U}0, g , h ∈ H j sup d ( α j ( g h ) , α j ( g ) α j ( h )) ⟶ U 0 , inf g , h ∈ H j g ≠ h d ( α j ( g ) , α j ( h ) ) ⟶ U 1. \inf_{\substack{g,h\in H_j\\g\ne h}}d(\alpha_j(g),\alpha_j(h))\mathrel{\longrightarrow_U}1. g , h ∈ H j g = h inf d ( α j ( g ) , α j ( h )) ⟶ U 1. The last infimum is set to one for the trivial group. The map [ g j ] ↦ [ α j ( g j ) ] [g_j]\mapsto[\alpha_j(g_j)] [ g j ] ↦ [ α j ( g j )] identifies ∏ U a l g H j \prod_U^{\mathrm{alg}}H_j ∏ U alg H j with C ( a ) C(a) C ( a ) .
Proof. We express the hypotheses by formulas that can be evaluated in the coordinates. Besides γ ( a ) = 0 \gamma(a)=0 γ ( a ) = 0 , discreteness gives
ζ ( a ) : = sup x ( min { d ( x , 1 ) , 1 − d ( x , 1 ) } − δ ˙ a ( x ) ) = 0. \zeta(a):=\sup_x\left(\min\{d(x,1),1-d(x,1)\}-\dot{\delta}_a(x)\right)=0. ζ ( a ) := x sup ( min { d ( x , 1 ) , 1 − d ( x , 1 )} − δ ˙ a ( x ) ) = 0.
We put η j = γ ( a j ) + 1 / j \eta_j=\gamma(a_j)+1/j η j = γ ( a j ) + 1/ j and v j = ζ ( a j ) v_j=\zeta(a_j) v j = ζ ( a j ) , which both tend to zero along U U U . These coordinate defects control correction and separation: for every permutation x x x , the definitions supply y y y with
ϵ a j ( y ) ≤ η j , d ( x , y ) ≤ L ϵ a j ( x ) + η j (10) \epsilon_{a_j}(y)\le\eta_j,\qquad d(x,y)\le L\epsilon_{a_j}(x)+\eta_j \tag*{(10)} ϵ a j ( y ) ≤ η j , d ( x , y ) ≤ L ϵ a j ( x ) + η j (10)
and, since δ a j ( x ) ≤ L ϵ a j ( x ) \delta_{a_j}(x)\le L\epsilon_{a_j}(x) δ a j ( x ) ≤ L ϵ a j ( x ) ,
min { d ( x , 1 ) , 1 − d ( x , 1 ) } ≤ L ϵ a j ( x ) + v j . (11) \min\{d(x,1),1-d(x,1)\}\le L\epsilon_{a_j}(x)+v_j. \tag*{(11)} min { d ( x , 1 ) , 1 − d ( x , 1 )} ≤ L ϵ a j ( x ) + v j . (11)
To form the clusters, set r = 1 / ( 100 L ) r=1/(100L) r = 1/ ( 100 L ) and V j = { x : ϵ a j ( x ) ≤ r } V_j=\{x:\epsilon_{a_j}(x)\le r\} V j = { x : ϵ a j ( x ) ≤ r } . We work on a set in U U U where v j < 1 / 100 v_j<1/100 v j < 1/100 and η j < r / 10 \eta_j<r/10 η j < r /10 . Inversion invariance and subadditivity of the defect then give, for x , y ∈ V j x,y\in V_j x , y ∈ V j ,
d ( x , y ) ≤ 3 / 100 or d ( x , y ) ≥ 97 / 100. d(x,y)\le3/100\quad\text{or}\quad d(x,y)\ge97/100. d ( x , y ) ≤ 3/100 or d ( x , y ) ≥ 97/100.
Every product of two elements of V j V_j V j can be corrected into V j V_j V j within 2 L r + η j < 1 / 5 2Lr+\eta_j<1/5 2 L r + η j < 1/5 , by (10) . As in [26 Lemma 4.2] , closeness within 1 / 5 1/5 1/5 defines an equivalence relation on V j V_j V j , whose classes form a finite group H j H_j H j by composing representatives and correcting the result. Indeed, the distance gap makes this multiplication independent of the choices. The two corrected triple products lie within 4 / 5 4/5 4/5 of each other, so the same gap places them in one class and gives associativity. Identity and inversion are inherited from the symmetric group.
To obtain the required uniform estimates, we correct each representative once more and choose α j ( g ) \alpha_j(g) α j ( g ) with defect at most η j \eta_j η j , taking α j ( 1 ) = 1 \alpha_j(1)=1 α j ( 1 ) = 1 . Their products belong to V j V_j V j and represent the corresponding product classes. We may therefore apply (11) , using the fact that elements of one class are close and elements of distinct classes are far apart, to obtain
d ( α j ( g h ) , α j ( g ) α j ( h ) ) ≤ 3 L η j + v j , d ( α j ( g ) , α j ( h ) ) ≥ 1 − 2 L η j − v j ( g ≠ h ) . d(\alpha_j(gh),\alpha_j(g)\alpha_j(h))\le3L\eta_j+v_j,\qquad d(\alpha_j(g),\alpha_j(h))\ge1-2L\eta_j-v_j\quad(g\ne h). d ( α j ( g h ) , α j ( g ) α j ( h )) ≤ 3 L η j + v j , d ( α j ( g ) , α j ( h )) ≥ 1 − 2 L η j − v j ( g = h ) .
These estimates give the three limits and show that the induced homomorphism is injective. To prove surjectivity, let x = [ x j ] ∈ C ( a ) x=[x_j]\in C(a) x = [ x j ] ∈ C ( a ) . Then x j ∈ V j x_j\in V_j x j ∈ V j on a set in U U U ; its class g j g_j g j satisfies d ( x j , α j ( g j ) ) ≤ L ( ϵ a j ( x j ) + η j ) + v j ⟶ U 0 d(x_j,\alpha_j(g_j))\le L(\epsilon_{a_j}(x_j)+\eta_j)+v_j\mathrel{\longrightarrow_U}0 d ( x j , α j ( g j )) ≤ L ( ϵ a j ( x j ) + η j ) + v j ⟶ U 0 . Thus the chosen representatives also give every element of the centralizer, proving surjectivity. At the unused coordinates we take trivial groups. □ \square □
For an admissible tuple, the group-theoretic conditions identify the finite groups obtained from the cluster lemma. Spectral gap gives an orbit containing asymptotically all points. The two commuting faithful actions are then regular there, which recovers the matrix dimensions.
Proposition 6.5. Let M = M u , d M=\mathcal{M}_{u,d} M = M u , d have canonical diagonal D D D . If its Weyl group has an admissible tuple ( a , t ) (a,t) ( a , t ) , then there are integers r j ≥ 2 r_j\ge2 r j ≥ 2 such that
∣ G r j ∣ d j ⟶ U 1 , C ( a ) ≅ ∏ U a l g G r j . \frac{|G_{r_j}|}{d_j}\mathrel{\longrightarrow_U}1,\qquad C(a)\cong\prod_U^{\mathrm{alg}}G_{r_j}. d j ∣ G r j ∣ ⟶ U 1 , C ( a ) ≅ U ∏ alg G r j .
Proof. Lemmas 6.3 and 6.4 give finite groups H j H_j H j and uniformly separated approximate representations α j \alpha_j α j realizing C ( a ) C(a) C ( a ) . By admissibility, this group is CA and nonmetabelian: the commutator in (4) witnesses noncommutation in its derived subgroup. Since these are first-order conditions, they hold in H j H_j H j on a set in U \mathcal{U} U . The classification of finite CA-groups identifies the nonmetabelian ones with G r G_r G r , r ≥ 2 r \ge2 r ≥ 2 [ACMM25, Proposition 12.2]; see [Wei25, BSW58, Suz57]. After changing unused coordinates, we may therefore write H j = G r j H_j=G_{r_j} H j = G r j .
To recover the two commuting actions, choose permutation lifts a i , j a_{i,j} a i , j , t j t_j t j , and put β j ( g ) = t j α j ( g ) t j − 1 \beta_j(g)=t_j\alpha_j(g)t_j^{-1} β j ( g ) = t j α j ( g ) t j − 1 . Their images represent C ( b ) C(b) C ( b ) and commute asymptotically uniformly with those of α j \alpha_j α j : otherwise maximizing witnesses would contradict (3) in the ultraproduct. Consequently ( g , h ) ↦ α j ( g ) β j ( h ) (g,h)\mapsto\alpha_j(g)\beta_j(h) ( g , h ) ↦ α j ( g ) β j ( h ) is a uniform approximate action of G r j × G r j G_{r_j}\times G_{r_j} G r j × G r j . We may therefore apply uniform flexible permutation stability [6 Theorem 1.2] to obtain genuine actions
π j : G r j × G r j → Sym ( m j ) , m j ≥ d j , m j / d j → U 1 , \pi_j:G_{r_j}\times G_{r_j}\to\operatorname{Sym}(m_j),\qquad m_j\ge d_j,\qquad m_j/d_j\to_{\mathcal{U}}1, π j : G r j × G r j → Sym ( m j ) , m j ≥ d j , m j / d j → U 1 ,
uniformly close to these maps extended by the identity. By the uniform separation of the original maps, every nonidentity element of either factor still moves a proportion of points tending uniformly to one.
To find a large orbit, we use the original generators. Since a i ∈ C ( b ) a_i\in C(b) a i ∈ C ( b ) , choose h i , j ∈ G r j h_{i,j}\in G_{r_j} h i , j ∈ G r j such that π j ( 1 , h i , j ) \pi_j(1,h_{i,j}) π j ( 1 , h i , j ) approximates a i , j a_{i,j} a i , j , extended by the identity, and put H j ′ = ⟨ h 1 , j , h 2 , j ⟩ H'_j=\langle h_{1,j},h_{2,j}\rangle H j ′ = ⟨ h 1 , j , h 2 , j ⟩ . We claim that its action has an orbit Y j Y_j Y j with ∣ Y j ∣ / m j → U 1 |Y_j|/m_j\to_{\mathcal{U}}1 ∣ Y j ∣/ m j → U 1 . Indeed, otherwise unions of orbits give invariant sets with relative sizes in [ ε / 2 , 1 − ε / 2 ] [\varepsilon/2,1-\varepsilon/2] [ ε /2 , 1 − ε /2 ] for some 0 < ε ≤ 1 / 2 0<\varepsilon\le1/2 0 < ε ≤ 1/2 : take an orbit in this interval, or accumulate smaller ones. Intersecting these sets with the original d j d_j d j points changes their relative sizes by a quantity tending to zero. Their diagonal projections are asymptotically fixed by a 1 , a 2 a_1,a_2 a 1 , a 2 , since the corrected generators preserve the sets and approximate the original generators. The resulting ultraproduct projection has trace bounded away from zero and one, contradicting (2).
Since the first factor commutes with this action, it preserves the unique largest orbit Y j Y_j Y j . Its action there is faithful, since a nonidentity element trivial on Y j Y_j Y j would move only o U ( m j ) o_{\mathcal{U}}(m_j) o U ( m j ) points. The centralizer of a transitive action on Y j Y_j Y j has at most ∣ Y j ∣ |Y_j| ∣ Y j ∣ elements, as a commuting permutation is determined by the image of one point. Faithfulness therefore gives the first inequality below, while transitivity gives the second and subgroup inclusion gives the last:
∣ G r j ∣ ≤ ∣ Y j ∣ ≤ ∣ H j ′ ∣ ≤ ∣ G r j ∣ . |G_{r_j}|\le|Y_j|\le|H'_j|\le|G_{r_j}|. ∣ G r j ∣ ≤ ∣ Y j ∣ ≤ ∣ H j ′ ∣ ≤ ∣ G r j ∣.
Equality throughout makes both factor actions regular on Y j Y_j Y j , and gives ∣ G r j ∣ / d j → U 1 |G_{r_j}|/d_j\to_{\mathcal{U}}1 ∣ G r j ∣/ d j → U 1 , as required. □ \square □
Corollary 6.6. Let r j ≥ 2 r_j\ge2 r j ≥ 2 , with r j → ∞ r_j\to\infty r j → ∞ , and suppose that
∏ U M ∣ G r j ∣ ≡ M V , e . \prod_{\mathcal{U}} M_{|G_{r_j}|}\equiv M_{\mathcal{V},e}. U ∏ M ∣ G r j ∣ ≡ M V , e . Then there are s j ≥ 2 s_j\ge2 s j ≥ 2 with ∣ G s j ∣ / e j → 1 |G_{s_j}|/e_j\to1 ∣ G s j ∣/ e j → 1 , and ∏ U a l g G r j \prod_{\mathcal{U}}^{\mathrm{alg}} G_{r_j} ∏ U alg G r j and ∏ V a l g G s j \prod_{\mathcal{V}}^{\mathrm{alg}} G_{s_j} ∏ V alg G s j are elementarily equivalent in the ordinary group language.
Proof. Theorem C(2) gives elementary equivalence of the canonical diagonal pairs. We may therefore transfer the tuple from Lemma 6.2 using Lemma 6.3, and then apply Proposition 6.5 in the target to obtain the required finite groups and dimension ratios. □ \square □
Element orders and amplifications
Corollary 6.6 gives both a dimension ratio and elementary equivalence of the recovered group ultraproducts. We now use element orders to extract divisibility information from their theories.
Lemma 6.7. For m ≥ 3 m\ge3 m ≥ 3 and r ≥ 2 r\ge2 r ≥ 2 , the group G r G_r G r contains an element of order 2 m − 1 2^m-1 2 m − 1 if and only if m ∣ r m\mid r m ∣ r .
Proof. Since the field has characteristic two, G r = SL 2 ( 2 r ) G_r=\operatorname{SL}_2(2^r) G r = SL 2 ( 2 r ) , and an element of odd order is diagonalizable over an algebraic closure, with eigenvalues λ , λ − 1 \lambda,\lambda^{-1} λ , λ − 1 . Frobenius either fixes or exchanges them, so the order of the element divides 2 r − 1 2^r-1 2 r − 1 or 2 r + 1 2^r+1 2 r + 1 . To determine when 2 m − 1 2^m-1 2 m − 1 divides either number, write r = q m + j r=qm+j r = q m + j , where 0 ≤ j < m 0\leq j<m 0 ≤ j < m . Modulo 2 m − 1 2^m-1 2 m − 1 , these two integers are 2 j − 1 2^j-1 2 j − 1 and 2 j + 1 2^j+1 2 j + 1 . The first is divisible by 2 m − 1 2^m-1 2 m − 1 precisely when j = 0 j=0 j = 0 , while 0 < 2 j + 1 ≤ 2 m − 1 + 1 < 2 m − 1 0<2^j+1\leq2^{m-1}+1<2^m-1 0 < 2 j + 1 ≤ 2 m − 1 + 1 < 2 m − 1 .
Conversely, if m ∣ r m\mid r m ∣ r , the cyclic group F 2 r \mathbb{F}_{2^r} F 2 r contains an element λ \lambda λ of order 2 m − 1 2^m-1 2 m − 1 , and diag ( λ , λ − 1 ) ∈ G r \operatorname{diag}(\lambda,\lambda^{-1})\in G_r diag ( λ , λ − 1 ) ∈ G r has that order. □
Existence of an element of any fixed finite order is a first-order condition in the discrete group language. Thus the lemma and Łoś’ theorem read divisibility information from the theory of an algebraic ultraproduct of the G r G_r G r : they determine whether the indices r r r are divisible by each fixed m ≥ 3 m\geq3 m ≥ 3 along the ultrafilter.
To apply this information to amplifications, recall that Q t Q^t Q t is the normalized corner p M k ( Q ) p pM_k(Q)p p M k ( Q ) p with ( τ k ⊗ τ ) ( p ) = t / k (\tau_k\otimes\tau)(p)=t/k ( τ k ⊗ τ ) ( p ) = t / k , and that F f o ( Q ) = { t > 0 : Q t ≡ Q } \mathcal{F}_{\mathrm{fo}}(Q)=\{t>0:Q^t\equiv Q\} F fo ( Q ) = { t > 0 : Q t ≡ Q } . For matrix ultraproducts, this compression can be carried out coordinatewise.
Lemma 6.8. Let Q = M U , d Q=\mathcal{M}_{\mathcal{U},d} Q = M U , d be a tracial matrix ultraproduct, and let t > 0 t>0 t > 0 . Then
Q t ≅ ∏ U M e i ( C ) , e i = max { 1 , ⌊ t d i ⌋ } . Q^t\cong\prod_{\mathcal{U}}M_{e_i}(\mathbb{C}),\qquad e_i=\max\{1,\lfloor td_i\rfloor\}. Q t ≅ U ∏ M e i ( C ) , e i = max { 1 , ⌊ t d i ⌋} . Proof. To realize the amplification coordinatewise, choose an integer k > t k>t k > t and projections p i ∈ M k d i ( C ) p_i\in M_{kd_i}(\mathbb{C}) p i ∈ M k d i ( C ) of rank e i e_i e i . Under the entrywise identification M k ( Q ) ≅ ∏ U M k d i ( C ) M_k(Q)\cong\prod_{\mathcal{U}}M_{kd_i}(\mathbb{C}) M k ( Q ) ≅ ∏ U M k d i ( C ) , the projection p = [ p i ] U p=[p_i]_{\mathcal{U}} p = [ p i ] U has normalized trace t / k t/k t / k , so its normalized corner represents the amplification. Coordinate compression identifies this corner with ∏ U p i M k d i ( C ) p i ≅ ∏ U M e i ( C ) \prod_{\mathcal{U}}p_iM_{kd_i}(\mathbb{C})p_i\cong\prod_{\mathcal{U}}M_{e_i}(\mathbb{C}) ∏ U p i M k d i ( C ) p i ≅ ∏ U M e i ( C ) . Since e i / ( k d i ) → U t / k > 0 e_i/(kd_i)\to_{\mathcal{U}}t/k>0 e i / ( k d i ) → U t / k > 0 , normalizing the corner traces does not change the ideal of sequences with vanishing 2 2 2 -norm. □
The dimension ratio permits only a bounded change in the group indices. An unbounded set of fixed divisors along the ultrafilter forces that change to vanish.
Proposition 6.9. Let r i ≥ 2 r_i\geq2 r i ≥ 2 , with r i → ∞ r_i\to\infty r i → ∞ . Suppose that the integers m ≥ 3 m\geq3 m ≥ 3 satisfying { i : m ∣ r i } ∈ U \{i:m\mid r_i\}\in\mathcal{U} { i : m ∣ r i } ∈ U form an unbounded set. Then
F f o ( ∏ U M ∣ G r i ∣ ( C ) ) = F ( ∏ U M ∣ G r i ∣ ( C ) ) = { 1 } . \mathcal{F}_{\mathrm{fo}}\left(\prod_{\mathcal{U}}M_{|G_{r_i}|}(\mathbb{C})\right)=\mathcal{F}\left(\prod_{\mathcal{U}}M_{|G_{r_i}|}(\mathbb{C})\right)=\{1\}. F fo ( U ∏ M ∣ G r i ∣ ( C ) ) = F ( U ∏ M ∣ G r i ∣ ( C ) ) = { 1 } . Proof. Put Q = ∏ U M ∣ G r i ∣ ( C ) Q=\prod_{\mathcal{U}}M_{|G_{r_i}|}(\mathbb{C}) Q = ∏ U M ∣ G r i ∣ ( C ) and suppose that Q t ≡ Q Q^t\equiv Q Q t ≡ Q . By Lemma 6.8, Q t ≅ ∏ U M e i ( C ) Q^t\cong\prod_{\mathcal{U}}M_{e_i}(\mathbb{C}) Q t ≅ ∏ U M e i ( C ) , where e i = max { 1 , ⌊ t ∣ G r i ∣ ⌋ } e_i=\max\{1,\lfloor t|G_{r_i}|\rfloor\} e i = max { 1 , ⌊ t ∣ G r i ∣ ⌋} . We can therefore apply Corollary 6.6 to obtain integers s i ≥ 2 s_i\geq2 s i ≥ 2 such that
∣ G s i ∣ e i ⟶ U 1 , ∏ U G s i a l g ≡ ∏ U G r i a l g . \frac{|G_{s_i}|}{e_i}\longrightarrow_{\mathcal{U}}1,\qquad\prod_{\mathcal{U}}G_{s_i}^{\mathrm{alg}}\equiv\prod_{\mathcal{U}}G_{r_i}^{\mathrm{alg}}. e i ∣ G s i ∣ ⟶ U 1 , U ∏ G s i alg ≡ U ∏ G r i alg . The first limit implies that s i → U ∞ s_i\to_{\mathcal{U}}\infty s i → U ∞ , and comparison with the original dimensions gives
∣ G s i ∣ ∣ G r i ∣ = 8 s i − r i 1 − 4 − s i 1 − 4 − r i ⟶ U t . \frac{|G_{s_i}|}{|G_{r_i}|}=8^{s_i-r_i}\frac{1-4^{-s_i}}{1-4^{-r_i}}\longrightarrow_{\mathcal{U}}t. ∣ G r i ∣ ∣ G s i ∣ = 8 s i − r i 1 − 4 − r i 1 − 4 − s i ⟶ U t . Hence 8 s i − r i → U t 8^{s_i-r_i}\to_{\mathcal{U}}t 8 s i − r i → U t . Since t > 0 t>0 t > 0 , the integers s i − r i s_i-r_i s i − r i are bounded on a set in U \mathcal{U} U , so the ultrafilter selects one of their finitely many values, say k k k . Thus s i − r i = k s_i-r_i=k s i − r i = k on a set in U \mathcal{U} U and t = 8 k t=8^k t = 8 k .
To show that k = 0 k=0 k = 0 , fix m ≥ 3 m\geq3 m ≥ 3 such that { i : m ∣ r i } ∈ U \{i:m\mid r_i\}\in\mathcal{U} { i : m ∣ r i } ∈ U . By Lemma 6.7 and Łoś’ theorem, ∏ U G r i a l g \prod_{\mathcal{U}}G_{r_i}^{\mathrm{alg}} ∏ U G r i alg contains an element of order 2 m − 1 2^m-1 2 m − 1 . Elementary equivalence gives such an element in ∏ U G s i a l g \prod_{\mathcal{U}}G_{s_i}^{\mathrm{alg}} ∏ U G s i alg , so { i : m ∣ s i } ∈ U \{i:m\mid s_i\}\in\mathcal{U} { i : m ∣ s i } ∈ U . Intersecting these two sets with { i : s i − r i = k } \{i:s_i-r_i=k\} { i : s i − r i = k } shows that m ∣ k m\mid k m ∣ k . Such m m m are unbounded, so k = 0 k=0 k = 0 , and consequently t = 1 t=1 t = 1 . The claim for the ordinary fundamental group follows by inclusion. □
The indices defining the family in the theorem are eventually divisible by every fixed power of two. Their odd prime divisors distinguish the members of the family.
Proof of Theorem E. Recall that, for a set S S S of odd primes,
r i ( S ) = 2 i ∏ p ∈ S p ≤ i p , Q S = ∏ U M ∣ G r i ( S ) ∣ ( C ) . r_i(S)=2^i\prod_{\substack{p\in S\\p\leq i}}p,\qquad Q_S=\prod_{\mathcal{U}}M_{|G_{r_i(S)}|}(\mathbb{C}). r i ( S ) = 2 i p ∈ S p ≤ i ∏ p , Q S = U ∏ M ∣ G r i ( S ) ∣ ( C ) .
For every j ≥ 2 j \ge2 j ≥ 2 , the integer 2 j 2^j 2 j divides r i ( S ) r_i(S) r i ( S ) whenever i ≥ j i \ge j i ≥ j . Proposition 6.9 therefore gives F f o ( Q S ) = F ( Q S ) = { 1 } \mathcal{F}_{\mathrm{fo}}(Q_S)=\mathcal{F}(Q_S)=\{1\} F fo ( Q S ) = F ( Q S ) = { 1 } .
To distinguish the theories, suppose that Q S ≡ Q T Q_S \equiv Q_T Q S ≡ Q T . By Corollary 6.6, there are integers s i ≥ 2 s_i \ge2 s i ≥ 2 such that
∣ G s i ∣ ∣ G r i ( T ) ∣ ⟶ U 1 , ∏ u a l g G s i ≡ ∏ u a l g G r i ( S ) . \frac{\lvert G_{s_i}\rvert}{\lvert G_{r_i(T)}\rvert} \longrightarrow_{\mathcal{U}} 1,\qquad\prod_{u}^{\mathrm{alg}} G_{s_i} \equiv\prod_{u}^{\mathrm{alg}} G_{r_i(S)}. ∣ G r i ( T ) ∣ ∣ G s i ∣ ⟶ U 1 , u ∏ alg G s i ≡ u ∏ alg G r i ( S ) .
Since ∣ G r + 1 ∣ / ∣ G r ∣ > 8 \lvert G_{r+1}\rvert/\lvert G_r\rvert> 8 ∣ G r + 1 ∣ / ∣ G r ∣ > 8 , the first limit forces s i = r i ( T ) s_i=r_i(T) s i = r i ( T ) on a set in U \mathcal{U} U . For an odd prime p p p , Lemma 6.7 now gives
p ∈ S ⟺ { i : p ∣ r i ( S ) } ∈ U ⟺ { i : p ∣ r i ( T ) } ∈ U ⟺ p ∈ T . p \in S \Longleftrightarrow\{i:p\mid r_i(S)\}\in\mathcal{U} \Longleftrightarrow\{i:p\mid r_i(T)\}\in\mathcal{U} \Longleftrightarrow p\in T. p ∈ S ⟺ { i : p ∣ r i ( S )} ∈ U ⟺ { i : p ∣ r i ( T )} ∈ U ⟺ p ∈ T .
Thus S = T S=T S = T . Since there are 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 subsets of the odd primes, we obtain the asserted number of theories. □ \square □
Acknowledgments. We thank Ben Hayes, David Jekel, and Sorin Popa for discussions about tracial matrix ultraproducts, model theory, and random matrices. A comment by Jekel helped motivate the use of quantum expanders in the diagonal-recovery argument. We also are indebted to Ilijas Farah whose contributions helped to streamline some of the arguments.
AI assistance. Large language models were used to develop the proofs and write the manuscript. They suggested recovering the diagonal by combining commutants with spectral gap and helped with the subsequent arguments. The authors directed this work.
References
Institut für Geometrie, Technische Universität Dresden, 01062 Dresden, Germany
Email address: vadim.alekseev@tu-dresden.de, andreas.thom@tu-dresden.de