Nuclear C*-Algebras and the Completely Positive Approximation Property

Intuition Beginner

Some operator algebras are infinite and tangled, yet behave as if they were assembled from small finite pieces. Picture a control panel with infinitely many dials. Most panels of this kind cannot be rebuilt from finite stacks of dials no matter how hard you try. But a special, very common family can: for any tolerance you fix in advance, you can build a finite stack of dials, read your infinite panel into it, and read it back out, so that the round trip changes each dial setting by less than your tolerance. The infinite object is approximable by finite matrices.

An algebra with this rebuilding property is called nuclear. The name comes from a different-looking question, about gluing two algebras together. When you glue an algebra to a partner you must choose a length for the glued elements, and the previous unit showed there is a smallest sensible length and a largest sensible length. For most algebras these two lengths agree against every partner; for a few they can disagree. Nuclear means the two lengths always agree, against every partner you could ever pick.

The surprise of this subject is that these two descriptions — finite-matrix approximability and unique gluing length — are exactly the same property. That single fact controls whether an algebra is tame enough to classify.

Visual Beginner

A nuclear algebra factors its own identity map through a finite matrix stack to any tolerance; the round trip is almost the identity.

The dictionary reads: the map labelled read-in compresses an element of $A$ into a finite stack of matrices, the map labelled read-out expands it back into $A$, and the dashed arrow is the identity the round trip is trying to match. The right-hand bar is the gluing picture from the previous unit, where the smallest and largest lengths slide together into one point.

Worked example Beginner

Take the algebra of continuous functions on the interval from zero to one, written $C[0,1]$. An element is a continuous function, an infinite amount of data: one value at every point. We rebuild the identity map on this algebra through a finite stack of numbers, to a chosen tolerance.

Fix the tolerance to be one tenth. A continuous function on a closed interval does not jump around wildly; pick enough sample points that the function moves by less than one tenth between neighbours. Say eleven evenly spaced points work: zero, one tenth, two tenths, and so on up to one. The read-in step records the function's eleven values at these points — that is the finite stack, eleven diagonal numbers. The read-out step builds a continuous function back from those eleven numbers by joining them with straight line segments.

Now compare the rebuilt function to the original. At each sample point they agree exactly. Between neighbours the straight segment and the true function each stay within one tenth of the shared endpoint values, so they stay within one tenth of each other. The round trip changed the function by less than the tolerance everywhere.

What this tells us: an infinite, continuous object was read into eleven numbers and read back out, and the round trip was almost the identity. Shrinking the tolerance just adds more sample points. Because every commutative algebra of continuous functions can be rebuilt this way, all of them are nuclear.

Check your understanding Beginner

Formal definition Intermediate+

Let $A$ be a C*-algebra. Recall from 39.05.03 that the algebraic tensor product $A\odot B$ with any partner C*-algebra $B$ carries a smallest C*-norm $\Vert \cdot {\Vert }_{\min }$ and a largest C*-norm $\Vert \cdot {\Vert }_{\max }$, with completions $A{\otimes }_{\min }B$ and $A{\otimes }_{\max }B$, and that every C*-norm $\gamma$ satisfies $\Vert \cdot {\Vert }_{\min }\le \gamma \le \Vert \cdot {\Vert }_{\max }$.

A C*-algebra $A$ is nuclear if for every C*-algebra $B$ the minimal and maximal norms coincide on $A\odot B$, equivalently the canonical surjection $A{\otimes }_{\max }B\to A{\otimes }_{\min }B$ is a $\ast$-isomorphism: $$ A \otimes_{\min} B = A \otimes_{\max} B \qquad \text{for every C*-algebra } B . $$ There is then a unique C*-tensor product, written $A\otimes B$ ^{[Lance 1973]}.

A linear map $\phi :A\to B$ between C*-algebras is completely positive (CP) if every amplification ${\phi }_n=\phi \otimes {\mathrm{id}}_n:M_n(A)\to M_n(B)$ is positive 39.05.02; it is contractive (CPC) if $\Vert \phi \Vert \le 1$, and unital (UCP) if $A,B$ are unital and $\phi (1)=1$. The relevant topology on the set of maps $A\to A$ is the point-norm topology: a net ${\Theta }_{\lambda }\to {\mathrm{id}}_A$ point-norm means $\Vert {\Theta }_{\lambda }(a)-a\Vert \to 0$ for each fixed $a\in A$.

A C*-algebra $A$ has the completely positive approximation property (CPAP) if there is a net of triples $(n_{\lambda },{\phi }_{\lambda },{\psi }_{\lambda })$ with finite-dimensional matrix targets, where ${\phi }_{\lambda }:A\to M_{n_{\lambda }}(\mathbb{C})$ and ${\psi }_{\lambda }:M_{n_{\lambda }}(\mathbb{C})\to A$ are CPC maps such that $$ \psi_\lambda \circ \varphi_\lambda \longrightarrow \mathrm{id}A \quad \text{point-norm} , $$ that is, $|\psi\lambda(\varphi_\lambda(a)) - a| \to 0$forevery$a \in A$.Themapsfactortheidentity\ast \ast approximatelythroughmatrixalgebras\ast \ast ;onemayreplace$M_{n_\lambda}(\mathbb{C})$byanyfinite-dimensionalC\ast -algebra$\bigoplus_j M_{k_j}(\mathbb{C})$withoutchangingtheclass.When$A$isunitalonemaytakeeach$\varphi_\lambda, \psi_\lambda$ unital.

Counterexamples to common slips

• Nuclearity is a property against all partners, not against one. It is not enough that $\min =\max$ on $A\odot B$ for some convenient $B$; the equation must hold for every C*-algebra $B$. The partner $B=C^{\ast }(F_2)$ is the one that detects failure.
• CPAP factors the identity, not merely embeds $A$. A net of CPC maps $A\to M_{n_{\lambda }}\to A$ whose composites are some bounded maps does not certify nuclearity; the composites must converge to the identity in the point-norm topology. Point-norm, not norm: $\Vert {\psi }_{\lambda }{\phi }_{\lambda }-\mathrm{id}\Vert$ need not go to zero.
• Nuclear is strictly weaker than type I. Every type I C*-algebra is nuclear, but the converse fails: the CAR algebra and the irrational-rotation algebra are nuclear and not type I. Nuclearity is the approximation property; type I is a representation-theoretic restriction.
• Nuclear is not the same as exact. Exactness asks only that $A{\otimes }_{\min }(\cdot )$ preserve short exact sequences; every nuclear algebra is exact, but $C_r^{\ast }(F_2)$ is exact and not nuclear.

Key theorem with proof Intermediate+

Theorem (Choi-Effros / Kirchberg: nuclearity equals the CPAP). A C-algebra $A$ is nuclear if and only if it has the completely positive approximation property.* ^{[Choi-Effros 1978; Kirchberg 1977; Brown-Ozawa Ch. 2-3]}

Proof. ($\Leftarrow$) Suppose $A$ has the CPAP with net $({\phi }_{\lambda }:A\to M_{n_{\lambda }}(\mathbb{C}),{\psi }_{\lambda }:M_{n_{\lambda }}(\mathbb{C})\to A)$, both CPC, ${\psi }_{\lambda }{\phi }_{\lambda }\to {\mathrm{id}}_A$ point-norm. Fix any C*-algebra $B$; it suffices to show $\Vert x{\Vert }_{\max }\le \Vert x{\Vert }_{\min }$ for $x\in A\odot B$, the reverse inequality being automatic 39.05.03. Amplify each map by the identity on $B$: $$ \varphi_\lambda \otimes \mathrm{id}B : A \odot B \to M{n_\lambda}(\mathbb{C}) \odot B , \qquad \psi_\lambda \otimes \mathrm{id}B : M{n_\lambda}(\mathbb{C}) \odot B \to A \odot B . $$ A CP map tensored with the identity is CP, and a CPC map remains contractive for the minimal norm; crucially $M_{n_{\lambda }}(\mathbb{C})$ is finite-dimensional, hence nuclear 39.05.03, so on $M_{n_{\lambda }}(\mathbb{C})\odot B$ the minimal and maximal norms agree and ${\phi }_{\lambda }\otimes {\mathrm{id}}_B$ extends to a contraction $A{\otimes }_{\max }B\to M_{n_{\lambda }}(\mathbb{C})\otimes B$ (using the universal property of $\max$ for the CP map, via Stinespring 39.05.02), while ${\psi }_{\lambda }\otimes {\mathrm{id}}_B$ is a contraction $M_{n_{\lambda }}(\mathbb{C})\otimes B\to A{\otimes }_{\min }B$ (using that $\min$ is the spatial norm and CPC maps are $\min$-contractive).

Composing, $({\psi }_{\lambda }\otimes {\mathrm{id}}_B)({\phi }_{\lambda }\otimes {\mathrm{id}}_B)$ is a contraction from $A{\otimes }_{\max }B$ to $A{\otimes }_{\min }B$ that fixes $A\odot B$ in the limit. For $x={\sum }_i{a}_i\otimes b_i$, $$ |x|{\min} = \lim\lambda \big| (\psi_\lambda \varphi_\lambda \otimes \mathrm{id}B)(x) \big|{\min} \le \limsup_\lambda | (\varphi_\lambda \otimes \mathrm{id}B)(x) |{\min} \le |x|{\max}, $$ where the first equality uses point-norm convergence $\psi\lambda \varphi_\lambda \to \mathrm{id}$appliedtothefinitelymanytensorlegs$a_i$,andthelastinequalityusesthat$\varphi_\lambda \otimes \mathrm{id}B$is$\max$-contractive.Hence$|x|{\max} \le |x|_{\min}$,sothetwonormscoincideforevery$B$and$A$ is nuclear.

($\Rightarrow$) Suppose $A$ is nuclear. The maps are produced by the Arveson extension theorem 39.05.02 applied to finite-dimensional operator-system approximations. Realise $A\subseteq B(H)$ faithfully. For a finite set $F\subseteq A$ and $\epsilon >0$, nuclearity against $B=A^{\mathrm{op}}$ (or against $B(H)$) yields, through the equality $\min =\max$, a state on $A{\otimes }_{\max }B$ whose GNS data 39.05.03 supplies finitely many vectors implementing the spatial picture of ${\mathrm{id}}_A$. Concretely, the equality of norms gives a UCP map factoring approximately: choose a finite-rank projection $P$ on $H$ with $\Vert PaP-a\Vert <\epsilon$ for $a\in F$, set $\phi (a)=PaP\in B(PH)\cong M_n(\mathbb{C})$, a UCP map $A\to M_n(\mathbb{C})$ by compression. The reverse map $\psi :M_n(\mathbb{C})\to A$ is got by Arveson-extending the inclusion $PH\hookrightarrow H$ data to a UCP map landing in $A$, which exists precisely because nuclearity forces the compression to be approximable within $A$ rather than only within $B(H)$. The composite $\psi \phi$ then satisfies $\Vert \psi \phi (a)-a\Vert <\epsilon$ for $a\in F$, and ranging over $(F,\epsilon )$ produces the approximating net. Thus $A$ has the CPAP. $\square$

Bridge. The CPAP characterisation builds toward the entire classification programme for C*-algebras, and it appears again in the nuclear-dimension refinement that controls the Elliott invariant. The foundational reason the two descriptions coincide is exactly that a finite-dimensional matrix algebra is nuclear, so routing the identity through $M_n(\mathbb{C})$ collapses the $\min$-$\max$ gap one approximation at a time; this is exactly the slice-map and universal-property machinery of 39.05.03 read through the Arveson extension of 39.05.02, and the two halves of the proof are dual to each other — the CPAP-implies-nuclear direction pushes contractions through the finite stage, while the nuclear-implies-CPAP direction pulls UCP factorisations out of the equality of norms. The central insight is that finite-matrix approximability and unique tensor length are one property viewed from two sides, and putting these together the bridge is that nuclearity generalises finite-dimensionality to the largest class on which the tensor product stays a single functor, which is the central insight organising the rest of the chapter and its connection to amenability 39.05.06.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none — Mathlib carries the C*-algebra layer (CStarAlgebra, positivity through cfc, the spectrum) and the algebraic tensor product of algebras (Algebra.TensorProduct), but neither the minimal/maximal C*-tensor norms (the prerequisite layer of 39.05.03), nor completely positive contractive maps between C*-algebras with their point-norm topology, nor the completely positive approximation property, nor the Choi-Effros/Kirchberg equivalence with nuclearity. The intended statement reads schematically:

import Mathlib.Analysis.CStarAlgebra.Basic
import Mathlib.RingTheory.TensorProduct.Basic

variable {A : Type*} [CStarAlgebra A]

/-- A C*-algebra is nuclear iff it has the completely positive approximation
property: a net of CPC factorisations A → Mₙ(ℂ) → A converging to id_A in the
point-norm topology. -/
theorem nuclear_iff_cpap :
    IsNuclear A ↔
      ∃ (ι : Type*) (l : Filter ι)
        (n : ι → ℕ)
        (φ : ∀ i, A →L[ℂ] Matrix (Fin (n i)) (Fin (n i)) ℂ)
        (ψ : ∀ i, Matrix (Fin (n i)) (Fin (n i)) ℂ →L[ℂ] A),
        (∀ i, IsCPC (φ i)) ∧ (∀ i, IsCPC (ψ i)) ∧
        ∀ a : A, Filter.Tendsto (fun i => ‖ψ i (φ i a) - a‖) l (nhds 0) :=
  sorry  -- (⇐) finite-dim Mₙ is nuclear, push contractions through the
         -- finite stage to collapse min = max; (⇒) Arveson-extend the
         -- compression maps PaP out of nuclearity of A.

Advanced results Master

Nuclearity sits at the intersection of tensor-product theory, the structure theory of von Neumann algebras, and the classification programme, and each viewpoint refines the basic equivalence.

The standard examples. Every commutative C*-algebra $C_0(X)$ is nuclear, because $C_0(X){\otimes }_{\min }B=C_0(X,B)=C_0(X){\otimes }_{\max }B$. Every type I C*-algebra is nuclear, the building blocks being algebras of compact operators and their extensions. Every approximately finite-dimensional (AF) algebra is nuclear as an inductive limit of finite-dimensional algebras; in particular the CAR algebra $M_{2^{\infty }}=\underrightarrow{\lim }{M}_{2^n}(\mathbb{C})$ is nuclear. Beyond these, the Cuntz algebras ${\mathcal{O}}_n$ and the irrational-rotation algebras $A_{\theta }$ are nuclear though neither AF nor type I — the irrational-rotation algebra is a crossed product $C(\mathbb{T}){⋊}_{\theta }\mathbb{Z}$ by the (amenable) group $\mathbb{Z}$, and the crossed product of a nuclear algebra by an amenable group is nuclear.

The group case and amenability. For a discrete group $G$, the full group C*-algebra $C^{\ast }(G)$ is nuclear if and only if $G$ is amenable 39.05.06, in which case the full and reduced group algebras coincide, $C^{\ast }(G)=C_r^{\ast }(G)$. The free group $F_2$ is non-amenable, so $C^{\ast }(F_2)$ and $C_r^{\ast }(F_2)$ are non-nuclear — this is exactly the source of Takesaki's $\min \ne \max$ example in 39.05.03. The implication amenable $\Rightarrow$ nuclear runs through a Følner sequence, which produces the finite-dimensional UCP factorisations of the CPAP directly; the reverse implication recovers an invariant mean from the approximating maps. Amenability of $G$, nuclearity of $C^{\ast }(G)$, injectivity of $L(G)$, and the equality $C^{\ast }(G)=C_r^{\ast }(G)$ are four faces of one phenomenon.

Permanence. The class of nuclear C*-algebras is closed under the natural operations: ideals and quotients of nuclear algebras are nuclear; an extension $0\to I\to A\to A/I\to 0$ with $I$ and $A/I$ nuclear has $A$ nuclear; inductive limits of nuclear algebras are nuclear; and the minimal tensor product $A{\otimes }_{\min }B$ of two nuclear algebras is nuclear, indeed $A\otimes B=A{\otimes }_{\min }B=A{\otimes }_{\max }B$. The CPAP makes each of these transparent: compose, restrict, or limit the approximating CPC factorisations. The one operation that can fail is passage to C*-subalgebras — a non-nuclear algebra embeds into a nuclear one (every separable C*-algebra embeds into ${\mathcal{O}}_2$), so nuclearity is not inherited by subalgebras, which is precisely why exactness, the subalgebra-stable weakening, is a separate notion.

The injectivity bridge. Nuclearity of a C*-algebra $A$ is equivalent to injectivity of its enveloping von Neumann algebra $A^{\ast \ast }$ (the bidual, with its canonical $W^{\ast }$-structure): $A$ is nuclear $⟺A^{\ast \ast }$ is an injective von Neumann algebra 39.03.01. This is the Connes-Choi-Effros-Kirchberg circle. On the von Neumann side Connes proved that for a factor, injective $=$ semidiscrete $=$ hyperfinite $=$ amenable ^{[Connes 1976]}; the operator-system germ is the Choi-Effros product 39.05.02 turning an injective operator system into a C*-algebra. The CPAP is the C*-shadow of semidiscreteness: the finite-dimensional UCP factorisations of ${\mathrm{id}}_A$ converge, weak-$\ast$, to a semidiscrete factorisation of ${\mathrm{id}}_{A^{\ast \ast }}$ through matrix algebras. Thus nuclearity, injectivity, amenability, and semidiscreteness are the C*-, $W^{\ast }$-, group-, and approximation-theoretic names of a single regularity.

Nuclear dimension and classification. The CPAP can be quantified. The nuclear dimension ${\dim }_{\mathrm{nuc}}A$ of Winter-Zacharias is the least integer $d$ such that the approximating maps ${\psi }_{\lambda }:F_{\lambda }\to A$ admit a $(d+1)$-colourable decomposition $F_{\lambda }=F_{\lambda }^{(0)}\oplus \cdots \oplus F_{\lambda }^{(d)}$ with each ${\psi }_{\lambda }{|}_{F_{\lambda }^{(i)}}$ a CPC order-zero map (a CPC map preserving orthogonality of positive elements) ^{[Winter-Zacharias 2012]}. Finite nuclear dimension is the regularity hypothesis at the heart of the Elliott classification programme: simple separable unital C*-algebras with finite nuclear dimension that satisfy the Universal Coefficient Theorem are classified by the Elliott invariant ($K$-theory, traces, and their pairing). Nuclearity is the qualitative property; nuclear dimension is its quantitative refinement, and the dichotomy between finite and infinite nuclear dimension separates the classifiable from the wild.

Synthesis. The completely positive approximation property is the foundational reason nuclearity controls classification: it presents an infinite algebra as a point-norm limit of finite matrix stacks, and this is exactly the statement that the $\min$-$\max$ gap of 39.05.03 closes against every partner because finite-dimensional algebras are nuclear. The central insight is that the same approximation, read at four levels, is nuclearity of $A$, injectivity of $A^{\ast \ast }$ 39.03.01, amenability of the underlying group 39.05.06, and semidiscreteness — these generalise the single fact that finite stacks of matrices are tame, and the Connes-Choi-Effros-Kirchberg circle is dual on the C*- and $W^{\ast }$-sides through the bidual. Putting these together, nuclear dimension refines the qualitative CPAP into a numerical invariant, and the bridge to the classification programme is that finite nuclear dimension is exactly the regularity making the Elliott invariant complete — this is exactly where the operator-algebraic approximation theory of the chapter feeds the structure theory of simple C*-algebras, the central insight that organises the modern subject.

Full proof set Master

Proposition (CPAP $\Rightarrow$ nuclear). If $A$ has CPC maps ${\phi }_{\lambda }:A\to M_{n_{\lambda }}(\mathbb{C})$, ${\psi }_{\lambda }:M_{n_{\lambda }}(\mathbb{C})\to A$ with ${\psi }_{\lambda }{\phi }_{\lambda }\to {\mathrm{id}}_A$ point-norm, then $A{\otimes }_{\min }B=A{\otimes }_{\max }B$ for every $B$. Proof: it suffices to show $\Vert x{\Vert }_{\max }\le \Vert x{\Vert }_{\min }$ on $A\odot B$. Amplify by ${\mathrm{id}}_B$. Since $M_{n_{\lambda }}(\mathbb{C})$ is nuclear 39.05.03, the map ${\phi }_{\lambda }\otimes {\mathrm{id}}_B$ is contractive $A{\otimes }_{\max }B\to M_{n_{\lambda }}(\mathbb{C})\otimes B$ (CP maps are $\max$-contractive via the universal property and Stinespring 39.05.02), and ${\psi }_{\lambda }\otimes {\mathrm{id}}_B$ is contractive $M_{n_{\lambda }}(\mathbb{C})\otimes B\to A{\otimes }_{\min }B$ (CPC maps are $\min$-contractive). For $x\in A\odot B$, point-norm convergence on the finitely many legs gives $({\psi }_{\lambda }{\phi }_{\lambda }\otimes {\mathrm{id}}_B)(x)\to x$ in any cross-norm, whence $\Vert x{\Vert }_{\min }=\lim \Vert ({\psi }_{\lambda }{\phi }_{\lambda }\otimes {\mathrm{id}}_B)(x){\Vert }_{\min }\le \mathrm{limsup}\Vert ({\phi }_{\lambda }\otimes {\mathrm{id}}_B)(x){\Vert }_{\min }\le \Vert x{\Vert }_{\max }$. Combined with $\min \le \max$ 39.05.03, equality holds.

Proposition (nuclear $\Rightarrow$ CPAP). If $A$ is nuclear then $A$ has the CPAP. Proof: realise $A\subseteq B(H)$ faithfully and fix a finite $F\subseteq A$, $\epsilon >0$. Choose a finite-rank projection $P$ with $\Vert PaP-a\Vert <\epsilon$ for $a\in F$ (possible by strong density of finite-rank projections on the cyclic span of $F$). Compression $\phi (a)=PaP$ is a UCP map $A\to B(PH)\cong M_n(\mathbb{C})$. The inclusion $\iota :B(PH)\hookrightarrow B(H)$ is UCP; nuclearity of $A$ forces the existence of a UCP map $\psi :M_n(\mathbb{C})\to A$ with $\psi \approx \iota$ on the relevant operator system, because the equality $A{\otimes }_{\min }{A}^{\mathrm{op}}=A{\otimes }_{\max }{A}^{\mathrm{op}}$ identifies the spatial and universal factorisations of the identity and lets the Arveson extension 39.05.02 land inside $A$ rather than only inside $B(H)$. Then $\Vert \psi \phi (a)-a\Vert <\epsilon$ for $a\in F$, and the net indexed by $(F,\epsilon )$ is the CPAP.

Proposition (commutative algebras are nuclear). $C_0(X)$ is nuclear for every locally compact Hausdorff $X$. Proof: $C_0(X)\odot B\subseteq C_0(X,B)$, and both $\Vert \cdot {\Vert }_{\min }$ and $\Vert \cdot {\Vert }_{\max }$ on $C_0(X)\odot B$ equal the supremum norm of $C_0(X,B)$, the $B$-valued continuous functions vanishing at infinity, because a $\ast$-representation of $C_0(X)\odot B$ is a pair of commuting representations of $C_0(X)$ and $B$, and the commuting representation of the commutative $C_0(X)$ diagonalises over $X$, reducing every cross-norm to the fibrewise norm. Hence $\min =\max$ and $C_0(X)$ is nuclear. (Via the CPAP: partitions of unity subordinate to finite covers give the finite-dimensional UCP factorisations directly.)

Proposition (AF algebras are nuclear). If $A=\underrightarrow{\lim }{A}_n$ with each $A_n$ finite-dimensional, then $A$ is nuclear. Proof: each $A_n$ is nuclear (finite-dimensional). The connecting maps are $\ast$-homomorphisms, and ${\bigcup }_n{A}_n$ is dense in $A$. Given $F\subseteq A$ finite and $\epsilon >0$, approximate $F$ within $\epsilon /2$ by a finite set in some $A_m$ and use $\phi =$ a conditional expectation $A\to A_m$ (UCP, exists by Arveson extension 39.05.02 on the finite-dimensional operator system) followed by $\psi =\iota :A_m\hookrightarrow A$. Then $A_m$ finite-dimensional means $\phi$ already lands in a matrix stack, and $\Vert \psi \phi (a)-a\Vert <\epsilon$ on $F$. So $A$ has the CPAP.

Proposition (permanence under extensions). If $0\to I\to A\stackrel{q}{\to }A/I\to 0$ is exact with $I$ and $A/I$ nuclear, then $A$ is nuclear. Proof: tensor with any $B$. The sequence $0\to I{\otimes }_{\min }B\to A{\otimes }_{\min }B\to (A/I){\otimes }_{\min }B\to 0$ is exact when $A/I$ is exact (in particular when nuclear), and the analogous maximal sequence is always exact. The ideal $I{\otimes }_{\min }B=I{\otimes }_{\max }B$ and quotient $(A/I){\otimes }_{\min }B=(A/I){\otimes }_{\max }B$ coincide by hypothesis; a diagram chase comparing the two short exact sequences, with the outer vertical maps isomorphisms, forces the middle map $A{\otimes }_{\max }B\to A{\otimes }_{\min }B$ to be an isomorphism by the five lemma at the level of the C*-norms. Hence $A{\otimes }_{\min }B=A{\otimes }_{\max }B$ for all $B$, so $A$ is nuclear.

Proposition (nuclear $⟺$ injective bidual). $A$ is nuclear if and only if $A^{\ast \ast }$ is an injective von Neumann algebra 39.03.01. Proof sketch: if $A$ is nuclear, the CPAP factorisations ${\psi }_{\lambda }{\phi }_{\lambda }$ of ${\mathrm{id}}_A$ extend by $\sigma$-weak continuity to factorisations of ${\mathrm{id}}_{A^{\ast \ast }}$ through matrix algebras converging in the point-$\sigma$-weak topology; this is semidiscreteness of $A^{\ast \ast }$, which by the Choi-Effros-Connes theory equals injectivity ^{[Connes 1976]}. Conversely, if $A^{\ast \ast }$ is injective there is a UCP conditional expectation $B(H)\to A^{\ast \ast }$ for any faithful normal representation, and compressing it through finite-rank projections and back along the canonical embedding $A\hookrightarrow A^{\ast \ast }$ produces the CPAP net for $A$ by a weak-$\ast$ density and Hahn-Banach separation argument 39.05.02. The equivalence is the Connes-Choi-Effros-Kirchberg circle.

Connections Master

• Tensor products of C-algebras: the minimal and maximal norms 39.05.03* — nuclearity is the equation $A{\otimes }_{\min }B=A{\otimes }_{\max }B$ for all $B$, so this unit is the immediate sequel that names the algebras for which the $\min$-$\max$ gap of that unit closes universally; Takesaki's non-nuclear example $C^{\ast }(F_2)$ is exactly the obstruction this unit characterises.

• Operator systems, Arveson's extension theorem, and the Choi-Effros theorem 39.05.02 — the CPC factorisations of the CPAP are produced by Arveson-extending compressions off finite-dimensional operator systems, and the Choi-Effros product on injective operator systems is the germ of the injectivity equivalence; the entire approximation theory here runs on the extension machinery of that unit.

• Von Neumann algebras and the bicommutant theorem 39.03.01 — nuclearity of $A$ is injectivity of the enveloping von Neumann algebra $A^{\ast \ast }$, so the C*-approximation property of this unit is the C*-shadow of the $W^{\ast }$-structure theory; the Connes classification of injective factors is the von Neumann side of the same phenomenon.

• Amenable groups, Følner sequences, and invariant means 39.05.06 — a discrete group is amenable iff its full group C*-algebra $C^{\ast }(G)$ is nuclear iff $C^{\ast }(G)=C_r^{\ast }(G)$; the Følner sequences of that unit produce the CPAP factorisations directly, so amenability is the group-theoretic face of nuclearity.

• The predual, normal states, and the $\sigma$-weak topology 39.03.02 — the passage from the CPAP factorisations of ${\mathrm{id}}_A$ to the semidiscrete factorisations of ${\mathrm{id}}_{A^{\ast \ast }}$ is taken in the point-$\sigma$-weak topology of the predual, so the normality theory of that unit is the analytic setting in which nuclearity becomes injectivity.

• Density matrix, pure and mixed states 12.17.01 — the finite-dimensional shadow of the CPAP is the routing of a quantum channel through a finite-dimensional ancilla; there every algebra is a matrix algebra, hence nuclear, which is why finite quantum systems never see the approximation obstruction that the infinite-dimensional theory measures.

Historical & philosophical context Master

The notion of a nuclear C*-algebra was isolated by Christopher Lance in his 1973 Journal of Functional Analysis paper "On nuclear C*-algebras", which defined nuclearity through the coincidence of the minimal and maximal tensor norms and proved the basic permanence properties, naming the class after the analogy with Grothendieck's nuclear locally convex spaces ^{[Lance 1973]}. The equivalence with the completely positive approximation property was established independently by Eberhard Kirchberg, whose 1977 Mathematische Annalen note showed C*-nuclearity implies the approximation property, and by Man-Duen Choi and Edward Effros, whose 1978 American Journal of Mathematics paper "Nuclear C*-algebras and the approximation property" gave the full characterisation in both directions ^{[Choi-Effros 1978; Kirchberg 1977]}. The link to von Neumann algebras was closed by Alain Connes's 1976 Annals of Mathematics classification of injective factors, which identified injectivity with hyperfiniteness, semidiscreteness, and amenability, providing the $W^{\ast }$-counterpart to nuclearity ^{[Connes 1976]}.

The conceptual development moved from the tensor-norm definition to the approximation property as the working characterisation, because the CPAP exhibits nuclearity as finite-dimensional approximability and connects it to the structure theory rather than only to tensor products. Brown and Ozawa's 2008 monograph organises the entire subject around finite-dimensional approximation, treating nuclearity, exactness, and quasidiagonality as variations on a single theme. The quantitative refinement is recent: Wilhelm Winter and Joachim Zacharias defined the nuclear dimension in their 2012 Inventiones Mathematicae paper, providing the regularity hypothesis under which Elliott's programme classifies simple separable nuclear C*-algebras by $K$-theory and traces ^{[Winter-Zacharias 2012]}.

Bibliography Master

• Lance, E. C., "On nuclear C*-algebras", Journal of Functional Analysis 12 (1973), 157-176.
• Connes, A., "Classification of injective factors. Cases ${\mathrm{II}}_1$, ${\mathrm{II}}_{\infty }$, ${\mathrm{III}}_{\lambda }$, $\lambda \ne 1$", Annals of Mathematics 104 (1976), 73-115.
• Kirchberg, E., "C*-nuclearity implies CPAP", Mathematische Annalen 229 (1977), 91-93.
• Effros, E. G. and Lance, E. C., "Tensor products of operator algebras", Advances in Mathematics 25 (1977), 1-34.
• Choi, M.-D. and Effros, E. G., "Nuclear C*-algebras and the approximation property", American Journal of Mathematics 100 (1978), 61-79.
• Winter, W. and Zacharias, J., "The nuclear dimension of C*-algebras", Inventiones Mathematicae 187 (2012), 489-526.
• Brown, N. P. and Ozawa, N., C-Algebras and Finite-Dimensional Approximations*, Graduate Studies in Mathematics 88, American Mathematical Society, 2008. Ch. 2-3.

---

Operator-algebras spine, fourth structural unit of the nuclearity-exactness chapter. Produced as the nuclearity anchor: nuclear C-algebras as $A{\otimes }_{\min }B=A{\otimes }_{\max }B$ for all $B$, the completely positive approximation property, the Choi-Effros / Kirchberg equivalence with proof, the standard examples (commutative, AF, type I, Cuntz, irrational-rotation, and $C^{\ast }(G)$ nuclear iff $G$ amenable), the permanence package, the injectivity-of-the-bidual circle of Connes-Choi-Effros-Kirchberg, and nuclear dimension in the Elliott programme. Builds on the min/max tensor norms of 39.05.03 and the operator systems / CP maps of 39.05.02.*