Group Approximation Properties and the Connes Embedding Problem

Intuition Beginner

The previous chapter sorted groups into two camps. A group is amenable when it carries a fair, shift-proof average, and the free group on two letters fails this test so badly that it can be cut up and reassembled into two copies of itself. That looks like a clean wall: amenable on one side, wild on the other. But the wall has rooms inside it. Many groups that fail the averaging test still behave well in weaker, more forgiving ways, and sorting those weaker good behaviours is what this unit is about.

The first relaxation keeps the averaging idea but lets the averages fade. Instead of demanding a perfect shift-proof average, you ask for a family of "almost averages" that spread out and thin to nothing across the group, the way ripples fade as they travel. The free group passes this softer test even though it failed the strict one. A group with such fading almost-averages is said to have the Haagerup property. Groups at the opposite extreme, the rigid property (T) groups, refuse even this and clamp every almost-average down.

The summit question of the chapter is a rebuilding question, like the one for nuclear algebras. Can every infinite symmetry built from a group be matched, to any tolerance, by shuffling a finite deck of cards? For decades the guess was yes. In 2020 the answer turned out to be no.

Visual Beginner

The picture is a ladder of rooms between the amenable floor and the wild ceiling, with the free group climbing partway up.

The dictionary reads: each rung up is a weaker requirement, so a group sitting on a high rung need not reach a lower one. The integers sit at the bottom (fully amenable); the free group fails the bottom rung but climbs several middle rungs; the rigid property (T) groups are locked out of even the second rung. The crossed-out top rung is the rebuilding question that 2020 settled in the negative.

Worked example Beginner

We check, by hand, that the integers carry the kind of fading almost-average that defines the second rung, and we watch a single fading function do the work.

Take the function on the integers given by $f(n)=r^{|n|}$ for a fixed number $r$ between zero and one, say $r=0.9$. So $f(0)=1$, $f(1)=f(-1)=0.9$, $f(2)=f(-2)=0.81$, and the values shrink toward zero as you walk away from the origin in either direction. This is a "fading" function: it equals one at the centre and dies off at the edges.

Step one: it fades. As $n$ grows, $0.9^{|n|}$ heads to zero, so only finitely many integers carry a value above any cutoff you fix. Pick the cutoff $0.1$; then $0.9^{|n|}>0.1$ only for $|n|$ up to $21$, a finite set. That is the fading-to-nothing requirement.

Step two: it is centred and positive in the right sense. The value at zero is one, the largest, and the function is symmetric, $f(n)=f(-n)$, matching the shape an almost-average should have.

Step three: shrink $r$ toward one. Taking $r=0.99$ makes the function fade much more slowly, spreading the almost-average over a wider stretch before it thins out, just as a larger averaging block did in the amenable picture.

What this tells us: the integers supply a whole family of centred functions that fade to nothing, one for each $r$. This fading family is the softened version of a shift-proof average, and having such a family is the Haagerup property. The free group has one too, which is why it climbs to this rung despite failing the amenable test.

Check your understanding Beginner

Formal definition Intermediate+

Throughout $G$ is a countable discrete group, $\lambda$ its left regular representation on ${\ell }^2(G)$, $C_r^{\ast }(G)$ the reduced group C*-algebra, and $L(G)=\lambda (G{)}^{\prime \prime }$ the group von Neumann algebra with canonical trace $\tau (x)=⟨x{\delta }_e,{\delta }_e⟩$. A function $\phi :G\to \mathbb{C}$ is positive-definite if the matrix $[\phi (s^{-1}t){]}_{s,t\in F}$ is positive for every finite $F\subseteq G$, and conditionally negative-definite (CND) if $\phi (e)=0$, $\phi (s^{-1})=\overline{\phi (s)}$, and ${\sum }_{s,t}{\bar{c}}_s{c}_t\phi (s^{-1}t)\le 0$ whenever ${\sum }_s{c}_s=0$. Schoenberg's correspondence ties them: $\psi$ is CND if and only if $e^{-r\psi }$ is positive-definite for every $r>0$.

Definition (Haagerup property / a-T-menability). $G$ has the Haagerup property if it admits a proper CND function $\psi :G\to [0,\infty )$ (proper: $\{s:\psi (s)\le R\}$ is finite for each $R$). Equivalently, $G$ admits a sequence ${\phi }_n$ of normalised positive-definite functions in $c_0(G)$ with ${\phi }_n\to 1$ pointwise ^{[Brown-Ozawa Ch. 12]}. Equivalently still, $G$ admits a proper affine isometric action on a real Hilbert space.

Definition (weak amenability and ${\Lambda }_{\mathrm{cb}}$). A Herz-Schur (completely bounded Fourier) multiplier is a function $\phi :G\to \mathbb{C}$ for which the map $m_{\phi }:\lambda (s)\mapsto \phi (s)\lambda (s)$ is completely bounded on $C_r^{\ast }(G)$; its cb-norm is $\Vert \phi {\Vert }_{\mathrm{cb}}$. The group $G$ is weakly amenable if there is a net ${\phi }_i\in c_{00}(G)$ (finitely supported) with ${\phi }_i\to 1$ pointwise and ${\mathrm{limsup}}_i\Vert {\phi }_i{\Vert }_{\mathrm{cb}}<\infty$. The Cowling-Haagerup constant is $$ \Lambda_{\mathrm{cb}}(G) = \inf \big{ \limsup_i |\varphi_i|{\mathrm{cb}} : \varphi_i \in c{00}(G),\ \varphi_i \to 1 \text{ pointwise} \big}, $$ with ${\Lambda }_{\mathrm{cb}}(G)=\infty$ when no such net exists ^{[Cowling-Haagerup 1989]}. Amenable groups have ${\Lambda }_{\mathrm{cb}}=1$; the free groups have ${\Lambda }_{\mathrm{cb}}(F_n)=1$ yet are non-amenable.

Definition (CBAP). $C_r^{\ast }(G)$ has the completely bounded approximation property (CBAP) if there is a net of finite-rank cb maps $T_i$ on $C_r^{\ast }(G)$ with $T_i\to \mathrm{id}$ in the point-norm topology and ${\mathrm{limsup}}_i\Vert T_i{\Vert }_{\mathrm{cb}}<\infty$; the least such bound is the Haagerup constant $\Lambda (C_r^{\ast }(G))={\Lambda }_{\mathrm{cb}}(G)$. Setting the constant to $1$ gives the metric (completely contractive) approximation property.

Definition (sofic and hyperlinear). $G$ is sofic if for every finite $F\subseteq G$ and $\epsilon >0$ there are $n$ and a map $\sigma :G\to \mathrm{Sym}(n)$ that is an $\epsilon$-approximate homomorphism and $\epsilon$-free in the normalised Hamming metric $d_H(\pi ,\rho )=\frac{1}{n}|\{k:\pi (k)\ne \rho (k)\}|$; $G$ is hyperlinear if the same holds with $\mathrm{Sym}(n)$ replaced by the unitary group $U(n)$ and Hamming by the normalised Hilbert-Schmidt metric $d_{\mathrm{HS}}(u,v)=\Vert u-v{\Vert }_2$ ^{[Elek-Szabó 2005]}.

Definition (Connes embedding property). Let $R$ be the hyperfinite ${\mathrm{II}}_1$ factor and $\omega$ a free ultrafilter on $\mathbb{N}$; the tracial ultrapower $R^{\omega }$ is ${\ell }^{\infty }(\mathbb{N},R)/\{(x_n):{\lim }_{\omega }\Vert x_n{\Vert }_2=0\}$, a ${\mathrm{II}}_1$ factor with trace ${\lim }_{\omega }\tau$. A separable ${\mathrm{II}}_1$ factor $M$ has the Connes embedding property if there is a trace-preserving embedding $M\hookrightarrow R^{\omega }$. The Connes Embedding Problem (CEP) asks whether every separable ${\mathrm{II}}_1$ factor has this property.

Counterexamples to common slips Intermediate+

• Haagerup is not amenability. The free group $F_2$ has the Haagerup property but is non-amenable; properness of the CND function, not boundedness of an averaging net, is the requirement, and $F_2$'s word length $\ell$ is itself CND and proper.
• Weak amenability is not the CBAP of $L(G)$ automatically. ${\Lambda }_{\mathrm{cb}}(G)<\infty$ gives the CBAP for $C_r^{\ast }(G)$ and $L(G)$, but the *value* matters: lattices in $\mathrm{Sp}(n,1)$ have ${\Lambda }_{\mathrm{cb}}=2n-1>1$, so they are weakly amenable yet lack the metric approximation property.
• Property (T) blocks Haagerup but not weak amenability in general. An infinite property (T) group never has the Haagerup property (the two are mutually exclusive for infinite groups), yet some property (T) groups are still weakly amenable while others (${\mathrm{SL}}_3(\mathbb{Z})$) have ${\Lambda }_{\mathrm{cb}}=\infty$.
• Sofic is not known to be everything. Every amenable and every residually finite group is sofic, and no non-sofic group is known; CEP-failure does not produce a non-hyperlinear group directly, because the dictionary runs through ${\mathrm{II}}_1$ factors, not the groups alone.

Key theorem with proof Intermediate+

Theorem (free groups have the Haagerup property; word length is CND). Let $F_S$ be the free group on a set $S$ with word-length function $\ell :F_S\to \mathbb{N}$. Then $\ell$ is a proper conditionally-negative-definite function, so $F_S$ has the Haagerup property; consequently $e^{-r\ell }$ is a normalised positive-definite function in $c_0(F_S)$ for every $r>0$, and ${\phi }_r=e^{-r\ell }\to 1$ pointwise as $r\to 0$. ^{[Haagerup 1979; Brown-Ozawa Ch. 12]}

Proof. Realise the Cayley graph of $F_S$ as a tree $T$ with vertex set $F_S$ and an edge between $g$ and $gs$ for $s\in S$. Orient each edge; let $E$ be the set of oriented edges and form the real Hilbert space $H={\ell }^2(E)$. For $g\in F_S$ let $[e,g]$ denote the unique reduced geodesic path from the identity $e$ to $g$, and define ${\xi }_g\in H$ to be the signed indicator of the edges traversed by that path (with $+1$ for an edge crossed along its orientation, $-1$ against). Then ${\xi }_e=0$ and, because $T$ is a tree, the path from $g$ to $h$ is the symmetric difference of the paths from $e$ to $g$ and from $e$ to $h$ along their shared initial segment, giving $$ |\xi_g - \xi_h|^2 = #{\text{edges on the geodesic } [g,h]} = \ell(g^{-1}h) = d(g,h). $$

The group acts on $T$ by left translation, permuting edges; this induces an orthogonal representation $\pi$ on $H$, and the assignment $g\mapsto {\xi }_g$ is a $1$-cocycle: ${\xi }_{gh}={\xi }_g+\pi (g){\xi }_h$, since travelling from $e$ to $gh$ goes from $e$ to $g$ and then along the $g$-translate of the path from $e$ to $h$. Define $\psi (g)=\Vert {\xi }_g{\Vert }^2=\ell (g)$. A function of the form $g\mapsto \Vert b(g){\Vert }^2$ for an affine isometric cocycle $b$ is conditionally negative-definite: for $\sum c_g=0$, $$ \sum_{g,h} \bar c_g c_h \psi(g^{-1}h) = \sum_{g,h} \bar c_g c_h |\xi_g - \xi_h|^2 = -2\Big| \sum_g c_g \xi_g \Big|^2 \le 0, $$ using $\sum c_g=0$ to discard the $\Vert {\xi }_g{\Vert }^2$ and $\Vert {\xi }_h{\Vert }^2$ diagonal terms. Thus $\ell$ is CND.

Properness is immediate: $\{g:\ell (g)\le R\}$ is the ball of radius $R$ in the tree, a finite set. By Schoenberg's correspondence $e^{-r\ell }$ is positive-definite for each $r>0$, it is normalised since $\ell (e)=0$, and it lies in $c_0(F_S)$ because $\ell$ is proper, so $e^{-r\ell }(g)\to 0$ as $\ell (g)\to \infty$. Finally $e^{-r\ell (g)}\to 1$ for each fixed $g$ as $r\to 0$. Hence $F_S$ has the Haagerup property. $\square$

Bridge. This theorem builds toward the entire approximation ladder of the chapter and it appears again in the von Neumann embedding theory of the master tier, where the free-group cocycle reappears as the model for a-T-menable actions. The foundational reason the free group escapes amenability yet keeps a usable approximation is exactly that its Cayley graph is a tree, so geodesics are unique and the word length is the squared length of a tree-valued cocycle; this is exactly the geometric content that the Følner picture of 39.05.06 lacks for $F_2$ and the analytic substitute that replaces it. The construction generalises: any group acting properly on a tree, or more generally on a space with measured walls, inherits a proper CND function, so the Haagerup property is the operator-algebraic shadow of negative-curvature geometry, and the bridge is that fading positive-definite functions are to the Haagerup property what Følner sets are to amenability — putting these together, weak amenability quantifies the same fading by a cb-norm budget, and the central insight is that amenability, the Haagerup property, and weak amenability form a strictly descending chain of approximation requirements, each the analytic relaxation of the one above, organising the rest of this unit and feeding the nuclearity comparison of 39.05.04.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none — Mathlib carries the C*-algebra layer and a developing amenability API but none of the approximation hierarchy or embedding theory here: it lacks positive-definite and conditionally-negative-definite functions on groups with Schoenberg's correspondence, the Haagerup property, completely bounded multipliers and the Cowling-Haagerup constant, the CBAP, the tracial ultrapower $R^{\omega }$ and the Connes embedding property, Kirchberg's QWEP equivalence, and sofic/hyperlinear approximation. The intended statement of the Key theorem reads schematically:

import Mathlib.Analysis.InnerProductSpace.l2Space
import Mathlib.GroupTheory.FreeGroup.Basic

variable {S : Type*}

/-- A function is conditionally negative-definite. -/
def IsCND (ψ : FreeGroup S → ℝ) : Prop :=
  ψ 1 = 0 ∧ (∀ g, ψ g⁻¹ = ψ g) ∧
    ∀ (F : Finset (FreeGroup S)) (c : FreeGroup S → ℝ),
      (∑ g ∈ F, c g = 0) → (∑ g ∈ F, ∑ h ∈ F, c g * c h * ψ (g⁻¹ * h) ≤ 0)

/-- The word length on a free group is a proper conditionally negative-definite
function; hence the free group has the Haagerup property. The cocycle is the
signed geodesic in the Cayley tree, with ‖ξ_g - ξ_h‖² = ℓ(g⁻¹h). -/
theorem freeGroup_wordLength_isCND :
    IsCND (fun g => (FreeGroup.norm g : ℝ)) :=
  sorry  -- tree cocycle b : g ↦ signed geodesic in ℓ²(edges);
         -- ψ(g)=‖b g‖², the cocycle identity b(gh)=b g + π g (b h),
         -- and ∑ c_g = 0 ⇒ ∑ c_g c_h ‖ξ_g-ξ_h‖² = -2‖∑ c_g ξ_g‖² ≤ 0.

Advanced results Master

The approximation hierarchy above amenability is a strictly descending chain, and the Connes embedding problem is its capstone, joining operator algebras to quantum information and computability.

The hierarchy and its separations. For a countable discrete group the implications $$ \text{amenable} ;\Rightarrow; \text{Haagerup} + \text{weakly amenable } (\Lambda_{\mathrm{cb}}=1) ;\Rightarrow; \text{exact} ;\Rightarrow; \text{hyperlinear} $$ all hold, and each is strict. The free groups $F_n$ separate the first arrow: non-amenable, yet Haagerup and weakly amenable with ${\Lambda }_{\mathrm{cb}}(F_n)=1$ ^{[Haagerup 1979]}. Lattices in $\mathrm{Sp}(n,1)$ separate weak amenability from the metric approximation property, carrying ${\Lambda }_{\mathrm{cb}}=2n-1$ ^{[Cowling-Haagerup 1989]}. Property (T) groups such as ${\mathrm{SL}}_3(\mathbb{Z})$ are not Haagerup and have ${\Lambda }_{\mathrm{cb}}=\infty$, hence lack the CBAP for $C_r^{\ast }(G)$, yet remain exact and hyperlinear. The Haagerup property and property (T) are mutually exclusive for infinite groups: the first demands an unbounded affine isometric action with no fixed point, the second forbids exactly that.

Weak amenability and the CBAP transfer. Weak amenability of $G$ with constant ${\Lambda }_{\mathrm{cb}}(G)$ is equivalent to the CBAP of $C_r^{\ast }(G)$, and to the weak* CBAP of $L(G)$, with the same constant; the finitely supported multipliers ${\phi }_i$ become the finite-rank cb maps $m_{{\phi }_i}$ on the algebra. This is the cb-quantified analogue of the Følner-driven CPAP of 39.05.04: there the approximating maps are completely positive contractions and the constant is $1$ (nuclearity), here they are merely completely bounded with a constant $\ge 1$. Haagerup's theorem that ${\Lambda }_{\mathrm{cb}}(F_n)=1$ is the first example of a non-nuclear C*-algebra with the metric approximation property, the historical origin of the whole circle of ideas ^{[Haagerup 1979]}.

Sofic, hyperlinear, and $L^2$-invariants. Sofic groups (Hamming-approximable in $\mathrm{Sym}(n)$) are hyperlinear (Hilbert-Schmidt-approximable in $U(n)$); the permutation-matrix functor sends an approximate action on a finite set to an approximate unitary representation, and an exact homomorphism becomes a trace-preserving embedding of $L(G)$ into $R^{\omega }$ ^{[Elek-Szabó 2005]}. Soficity has paid off in proving conjectures for sofic groups: Gottschalk's surjunctivity (Gromov, Weiss), the determinant conjecture and the algebraic eigenvalue conjecture, and the Connes embedding property for $L(G)$. No non-sofic group is known; whether all groups are sofic, or all ${\mathrm{II}}_1$ factors embeddable, were the group- and factor-level shadows of CEP.

The Connes embedding problem and its reformulations. Connes's 1976 classification of injective factors closed with the remark that it seemed hard to produce a ${\mathrm{II}}_1$ factor not embeddable in $R^{\omega }$ ^{[Connes 1976]}; this hardened into the CEP. Kirchberg recast it entirely in C*-language: CEP holds if and only if the QWEP conjecture holds (every C*-algebra is a quotient of one with Lance's weak expectation property), if and only if $C^{\ast }(F_{\infty })\otimes C^{\ast }(F_{\infty })$ carries a unique C*-norm, if and only if every finite von Neumann algebra is QWEP ^{[Kirchberg 1993]}. The same statement reappeared in quantum information as Tsirelson's problem: whether the commuting-operator model and the tensor-product model of quantum correlations yield the same closure of correlation sets ^{[Tsirelson 1993]}. Ozawa's survey made the chain CEP $\iff$ QWEP $\iff$ Kirchberg $\iff$ Tsirelson explicit.

The 2020 refutation. The complexity-theoretic identity ${\mathrm{MIP}}^{\ast }=\mathrm{RE}$ of Ji, Natarajan, Vidick, Wright, and Yuen shows that the class of languages decidable by a classical verifier interacting with two entangled provers equals the recursively enumerable sets — in particular it contains the halting problem ^{[JNVWY 2020]}. A perfect parallel-repetition and compression argument builds nonlocal games whose entangled value in the tensor-product model differs from the value in the commuting-operator model, refuting Tsirelson's problem; by Kirchberg's and Ozawa's equivalences this refutes the Connes embedding problem. There exists a separable ${\mathrm{II}}_1$ factor that does not embed in $R^{\omega }$. The proof is non-constructive in the operator-algebraic sense: it exhibits the failure through an undecidability reduction rather than a named factor.

Synthesis. The central insight organising this unit is that amenability is only the top of a descending tower of approximation properties, and the foundational reason the tower has many floors is that the single notion "approximate the identity by finite-dimensional data" splits, once you weaken completely-positive-contractive to completely-bounded and weaken global averaging to fading positive-definite functions, into the Haagerup property, weak amenability with its constant ${\Lambda }_{\mathrm{cb}}$, exactness, and finite-dimensional approximability of the trace. This is exactly the relaxation that lets the free group, locked out of amenability by its tree geometry in 39.05.06, re-enter every lower floor: its word-length cocycle is the proper CND function, its radial multipliers give ${\Lambda }_{\mathrm{cb}}=1$, and its permutation representations make it sofic. Putting these together, the Connes embedding problem is the summit — whether the tracial approximation by matrices that nuclearity supplies for amenable groups via 39.05.04 extends to every separable ${\mathrm{II}}_1$ factor — and it is dual on two faces, the QWEP/unique-norm face of Kirchberg and the nonlocal-games face of Tsirelson. The bridge is that a logical undecidability, ${\mathrm{MIP}}^{\ast }=\mathrm{RE}$, generalises a question about operator algebras into a question about the limits of computation and, by closing it negatively, shows the approximation tower has a genuine ceiling: not every II${}_1$ factor is matricially approximable, the first time a soft analytic embedding question was settled by hard complexity theory.

Full proof set Master

Proposition 1 (Schoenberg correspondence). A function $\psi :G\to \mathbb{R}$ with $\psi (e)=0$ and $\psi (g^{-1})=\psi (g)$ is conditionally negative-definite if and only if $e^{-r\psi }$ is positive-definite for every $r>0$. Proof: if $\psi$ is CND, write $\psi (g)=\Vert b(g){\Vert }^2$ for the Gelfand-pair cocycle $b$ into a real Hilbert space $H$ (existence of $b$ is the GNS-type construction from the CND kernel $K(g,h)=\frac{1}{2}(\psi (g^{-1}h)-\text{boundary terms})$). Then for $\sum c_g=0$ removed by working with differences, $$ \sum_{g,h} \bar c_g c_h e^{-r\psi(g^{-1}h)} = \sum_{g,h} \bar c_g c_h e^{-r|b(g)-b(h)|^2}, $$ and $e^{-r\Vert x-y{\Vert }^2}$ is a positive-definite kernel on $H$ (the Gaussian/Mehler kernel), so the sum is $\ge 0$; this gives positive-definiteness of $e^{-r\psi }$. Conversely, if $e^{-r\psi }$ is positive-definite for all $r>0$, then $\frac{1-e^{-r\psi }}{r}$ is conditionally negative-definite for each $r$, and letting $r\to 0^{+}$ the pointwise limit $\psi$ is CND as a limit of CND functions. $\square$

Proposition 2 (word length is proper CND on $F_S$). Restated from the Key theorem; the cocycle $g\mapsto {\xi }_g\in {\ell }^2(\text{edges})$ has $\Vert {\xi }_g-{\xi }_h{\Vert }^2=\ell (g^{-1}h)$. Proof: as in the Key theorem, using that the geodesic between two vertices of a tree is the non-shared part of their geodesics to the root, so the signed indicators subtract to the signed indicator of $[g,h]$, whose squared norm counts its edges, which is $d(g,h)=\ell (g^{-1}h)$. Properness is finiteness of balls in a locally finite tree. $\square$

Proposition 3 (Haagerup $\Rightarrow$ exact for the reduced algebra implication fails to reverse; weak amenability $\Rightarrow$ CBAP). If ${\Lambda }_{\mathrm{cb}}(G)<\infty$ then $C^r(G)$hastheCBAPwithconstant$\Lambda{\mathrm{cb}}(G)$.\ast Proof:let$\varphi_i \in c_{00}(G)$with$\varphi_i \to 1$pointwiseand$\limsup |\varphi_i|{\mathrm{cb}} = \Lambda{\mathrm{cb}}(G)$.EachHerz-Schurmultiplier$m_{\varphi_i} : \lambda(g) \mapsto \varphi_i(g)\lambda(g)$extendstoacbmapon$C^r(G)$with$|m{\varphi_i}|{\mathrm{cb}} = |\varphi_i|{\mathrm{cb}}$(Bo\dot{z}ejko-Fendler:theHerz-Schurandcb-multipliernormscoincide).Because$\varphi_i$isfinitelysupported,$m_{\varphi_i}$isfinite-rank.For$x = \sum_g x_g \lambda(g)$inthedense$$-subalgebra$\mathbb{C}[G]$,$m_{\varphi_i}(x) = \sum_g \varphi_i(g) x_g \lambda(g) \to x$innormsincethesumisfiniteand$\varphi_i(g) \to 1$;astandard$\varepsilon/3$argumentwiththeuniformcb-bound$\Lambda_{\mathrm{cb}}(G)$extendspoint-normconvergencetoallof$C^*r(G)$.Hence$(m{\varphi_i})$isaCBAPnetwith$\limsup |m_{\varphi_i}|{\mathrm{cb}} = \Lambda{\mathrm{cb}}(G)$.$\square$

Proposition 4 (sofic $\Rightarrow$ hyperlinear). Every sofic group has $L(G)\hookrightarrow R^{\omega }$. Proof: identical to Exercise 6 — convert Hamming-approximate homomorphisms into permutation unitaries, whose normalised-trace data realises the canonical trace of $L(G)$ in the matricial tracial ultraproduct ${\prod }_{\omega }(M_{k_n},\mathrm{tr})$, which embeds in $R^{\omega }$. Trace preservation forces injectivity of the induced $L(G)\to R^{\omega }$. $\square$

Proposition 5 (CEP $\iff$ unique C*-norm on $C^{\ast }(F_{\infty }{)}^{\otimes 2}$, Kirchberg). The Connes embedding problem holds if and only if $C^(F_\infty) \otimes_{\min} C^(F_\infty) = C^(F_\infty) \otimes_{\max} C^(F_\infty)$. Proof sketch: the forward direction is Exercise 7 — tracial states on the max tensor product are commuting pairs of representations, embeddable into $R^{\omega }$ under CEP and there approximable by tensor-split matricial pairs, collapsing max to min on tracial states; since $C^{\ast }(F_{\infty })$ is residually finite-dimensional its tensor norms are detected by such states. Conversely, given the unique norm, any separable ${\mathrm{II}}_1$ factor $M$ is generated by countably many unitaries, hence is a quotient of $C^{\ast }(F_{\infty })$ carrying a trace; the unique-norm hypothesis upgrades the trace to one continuous in the minimal norm, which is exactly the condition for the GNS construction to land inside an ultraproduct of matrix algebras, i.e. inside $R^{\omega }$. Thus the two statements are equivalent, and both equal QWEP for finite von Neumann algebras ^{[Kirchberg 1993]}. $\square$

Proposition 6 (refutation via ${\mathrm{MIP}}^{\ast }=\mathrm{RE}$). There exists a separable ${\mathrm{II}}_1$ factor with no trace-preserving embedding into $R^{\omega }$. Proof (as a reduction, after ^{[JNVWY 2020]}): ${\mathrm{MIP}}^{\ast }=\mathrm{RE}$ produces, for each Turing machine $T$, a nonlocal game ${\mathcal{G}}_T$ whose tensor-product entangled value ${\omega }_t({\mathcal{G}}_T)$ equals $1$ if $T$ halts and is $\le 1/2$ otherwise, while the commuting-operator value ${\omega }_{qc}({\mathcal{G}}_T)$ is always $1$. If the two models always agreed (Tsirelson's problem positive), the halting problem would be co-recursively-enumerable, contradicting its undecidability; hence ${\omega }_t\ne {\omega }_{qc}$ for some game, refuting Tsirelson's problem. By Proposition 5 and Ozawa's equivalence Tsirelson $\iff$ Kirchberg $\iff$ CEP, the Connes embedding problem fails: some separable ${\mathrm{II}}_1$ factor does not embed into $R^{\omega }$. $\square$

Connections Master

• Amenable groups: invariant means, the Følner condition, and paradoxical decompositions 39.05.06 — amenability is the top floor of the approximation tower this unit descends; the free group's failure of the Følner condition there, forced by its tree geometry, is exactly what the proper CND word-length cocycle here repairs, so the Haagerup property is the analytic substitute for the missing Følner sequence and weak amenability is its cb-quantified refinement.

• Nuclear C-algebras and the completely positive approximation property 39.05.04* — nuclearity is the CPAP with completely-positive-contractive maps and constant $1$; the CBAP of this unit is the same factorisation with merely completely-bounded maps and a constant ${\Lambda }_{\mathrm{cb}}\ge 1$, so weak amenability is to the CBAP what amenability is to nuclearity, and the Connes embedding problem asks whether the tracial matrix-approximability that nuclearity guarantees for amenable groups extends to every separable ${\mathrm{II}}_1$ factor.

• Exact C-algebras and the local lifting property 39.05.05* — exactness sits strictly between weak amenability and hyperlinearity in the tower; the QWEP class through which Kirchberg reformulates the Connes embedding problem is built on the weak expectation and local lifting properties that unit develops, so the embedding question is the von Neumann shadow of the exactness theory.

• Full and reduced group C-algebras 39.05.07* — the unique-C*-norm reformulation lives on $C^{\ast }(F_{\infty })\otimes C^{\ast }(F_{\infty })$, a statement about the gap between the full and reduced completions that unit constructs; Haagerup's multipliers and the Cowling-Haagerup constant are invariants of exactly that pair of completions.

• Property (T), Kazhdan pairs, and rigidity 39.04.03 — property (T) is the precise negation of the Haagerup property for infinite groups: the Delorme-Guichardet fixed-point characterisation forbids the unbounded affine isometric action that a proper CND function provides, so the two properties partition the rigid from the a-T-menable, and ${\mathrm{SL}}_3(\mathbb{Z})$'s ${\Lambda }_{\mathrm{cb}}=\infty$ is the operator-algebraic measure of its rigidity.

• The hyperfinite II${}_1$ factor and Connes' classification 39.03.05 — $R$ and its tracial ultrapower $R^{\omega }$ are the target of the embedding problem; Connes' identification of injective with hyperfinite is what makes $R^{\omega }$ the universal receptacle whose universality the 2020 theorem finally limits.

Historical & philosophical context Master

The relaxation of amenability begins with Uffe Haagerup's 1979 Inventiones paper, which proved that the reduced C*-algebra of the free group, though non-nuclear, has the metric approximation property, using the radial multipliers built from the word length and showing the word length is conditionally negative-definite ^{[Haagerup 1979]}. This single example fractured the apparent dichotomy of 39.05.06 into a hierarchy: a group could fail amenability yet keep finite-dimensional approximations of its algebra. Michael Cowling and Haagerup's 1989 Inventiones paper introduced the constant ${\Lambda }_{\mathrm{cb}}(G)$ and computed it for rank-one lattices, distinguishing weak amenability with constant one from larger constants and tying the invariant to Mostow rigidity ^{[Cowling-Haagerup 1989]}. The geometric reading — proper affine isometric actions on Hilbert spaces, the Gromov a-T-menability picture — recast the Haagerup property as the negation of Kazhdan's property (T).

The embedding question descends from Alain Connes's 1976 Annals classification of injective factors, which remarked in closing that producing a ${\mathrm{II}}_1$ factor not embeddable in $R^{\omega }$ appeared difficult ^{[Connes 1976]}. Eberhard Kirchberg's 1993 Inventiones paper turned this into hard C*-algebra theory, proving the equivalence with the QWEP conjecture and with the uniqueness of the C*-norm on $C^{\ast }(F_{\infty })\otimes C^{\ast }(F_{\infty })$ ^{[Kirchberg 1993]}. Independently, Boris Tsirelson's work on quantum correlations posed what became the same question in quantum information, the commuting-operator versus tensor-product models ^{[Tsirelson 1993]}; Narutaka Ozawa's surveys established the chain of equivalences. The resolution came from theoretical computer science: Zhengfeng Ji, Anand Natarajan, Thomas Vidick, John Wright, and Henry Yuen proved ${\mathrm{MIP}}^{\ast }=\mathrm{RE}$ in 2020, exhibiting nonlocal games separating the two quantum models and so refuting Tsirelson's problem and the Connes embedding problem ^{[JNVWY 2020]}. Soficity and hyperlinearity, isolated by Gromov and named by Weiss, Elek, and Szabó, were the group- and factor-level versions of the embedding question ^{[Elek-Szabó 2005]}.

Bibliography Master

@article{Connes1976cep,
  author  = {Connes, Alain},
  title   = {Classification of injective factors. Cases $II_1$, $II_\infty$, $III_\lambda$, $\lambda \neq 1$},
  journal = {Annals of Mathematics},
  volume  = {104},
  year    = {1976},
  pages   = {73--115}
}

@article{Haagerup1979,
  author  = {Haagerup, Uffe},
  title   = {An example of a non-nuclear C*-algebra which has the metric approximation property},
  journal = {Inventiones Mathematicae},
  volume  = {50},
  year    = {1979},
  pages   = {279--293}
}

@article{CowlingHaagerup1989,
  author  = {Cowling, Michael and Haagerup, Uffe},
  title   = {Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one},
  journal = {Inventiones Mathematicae},
  volume  = {96},
  year    = {1989},
  pages   = {507--549}
}

@article{Kirchberg1993,
  author  = {Kirchberg, Eberhard},
  title   = {On non-semisplit extensions, tensor products and exactness of group C*-algebras},
  journal = {Inventiones Mathematicae},
  volume  = {112},
  year    = {1993},
  pages   = {449--489}
}

@article{ElekSzabo2005,
  author  = {Elek, G{\'a}bor and Szab{\'o}, Endre},
  title   = {Hyperlinearity, essentially free actions and $L^2$-invariants. The sofic property},
  journal = {Mathematische Annalen},
  volume  = {332},
  year    = {2005},
  pages   = {421--441}
}

@article{Tsirelson1993,
  author  = {Tsirelson, Boris},
  title   = {Some results and problems on quantum Bell-type inequalities},
  journal = {Hadronic Journal Supplement},
  volume  = {8},
  year    = {1993},
  pages   = {329--345}
}

@article{JNVWY2020,
  author  = {Ji, Zhengfeng and Natarajan, Anand and Vidick, Thomas and Wright, John and Yuen, Henry},
  title   = {MIP* = RE},
  journal = {arXiv:2001.04383; Communications of the ACM},
  volume  = {64},
  year    = {2020/2021},
  pages   = {131--138}
}

@book{BrownOzawa2008cep,
  author    = {Brown, Nathanial P. and Ozawa, Narutaka},
  title     = {C*-Algebras and Finite-Dimensional Approximations},
  series    = {Graduate Studies in Mathematics},
  volume    = {88},
  publisher = {American Mathematical Society},
  year      = {2008}
}

---

Operator-algebras spine, capstone unit of the nuclearity-exactness chapter. Produced as the approximation-hierarchy summit: the Haagerup property (a-T-menability) through proper CND functions and $c_0$ positive-definite sequences, weak amenability and the Cowling-Haagerup constant ${\Lambda }_{\mathrm{cb}}(G)$, the CBAP for $C^_r(G)$and$L(G)$, sofic and hyperlinear groups, the Connes embedding problem with Kirchberg's QWEP / unique-norm reformulation and Tsirelson's problem, and the 2020 MIP*=RE refutation. Builds on amenability and the Følner condition of 39.05.06 and the completely positive approximation property of 39.05.04; co-produced 39.05.05 (exactness) and 39.05.07 (group C*-algebras) appear in the Connections.*