Tensor Products of C*-Algebras: the Minimal and Maximal Norms

Intuition Beginner

When you have two separate physical systems and want to study them as one, you glue their algebras of operators together. The raw glue is easy to describe: take products of an operator from the first system with an operator from the second, and allow finite sums of such products. This bare gluing gives an algebra, but it is missing one thing an operator algebra must have — a size, a length assigned to each element that behaves correctly under the star operation.

Here is the surprise. For a single operator algebra the length is forced: there is exactly one sensible way to measure size. For the glued algebra of two systems, the length is not forced. There can be several different sensible measurements, all agreeing on the building blocks but disagreeing on the sums. Two of them are special. One is the smallest sensible length, got by placing both systems honestly side by side as operators on a combined space of arrows. The other is the largest sensible length, got by allowing every possible honest placement at once and taking the biggest size any of them reports.

Most familiar systems have these two lengths coincide, so the gluing is unambiguous. But there are systems built from free groups where the smallest and largest genuinely differ. That gap is the seed of an entire subject.

Visual Beginner

The glued algebra sits between a smallest length and a largest length; when they agree the gluing is unique, when they differ a whole gap opens.

The dictionary reads: gluing the two algebras gives a raw algebra with no length yet; every sensible length sits on the bar between the smallest and the largest; and the entire question of the chapter is whether, for a given pair of systems, the two endpoints are one point or two.

Worked example Beginner

Glue the two-by-two diagonal numbers to themselves. Take system A to be pairs of numbers $(x_1,x_2)$ with slotwise multiplication, which you can picture as diagonal two-by-two matrices. Glue A to a copy of itself. A building block of the glued algebra is a product of one diagonal pair with another diagonal pair, and a general element is a finite sum of such products.

Count what you actually get. A diagonal pair has two slots, so the first copy contributes two independent directions and so does the second. Their products give two times two, which is four independent directions. So the glued algebra is the four-slot algebra: tuples $(y_{11},y_{12},y_{21},y_{22})$ with slotwise operations. The natural length of an element is the largest size among its four slots.

Now ask whether the smallest and largest sensible lengths agree here. Place the first copy as diagonal matrices on a two-arrow space and the second copy on another two-arrow space; side by side they act on a four-arrow space exactly as the four-slot algebra, and the length is the largest slot. Any other honest placement gives the same answer, because each slot is just a number being multiplied. So the smallest and the largest length coincide.

What this tells us: when the pieces are commutative the gluing has a unique length, and the four-slot answer is forced. The gap between smallest and largest only opens for genuinely noncommutative, large pieces such as the free-group algebras.

Check your understanding Beginner

Formal definition Intermediate+

Let $A$ and $B$ be C*-algebras. The algebraic tensor product $A\odot B$ is the $\ast$-algebra generated by elementary tensors $a\otimes b$ with the bilinear relations, the product $(a\otimes b)(a^{\prime }\otimes b^{\prime })=a{a}^{\prime }\otimes b{b}^{\prime }$, and the involution $(a\otimes b{)}^{\ast }=a^{\ast }\otimes b^{\ast }$, extended linearly; a general element is a finite sum ${\sum }_i{a}_i\otimes b_i$. The symbol $\odot$ denotes this purely algebraic gluing, which carries no norm yet ^{[Takesaki 1979 Ch. IV]}.

A C-norm* on $A\odot B$ is a norm $\gamma$ that is submultiplicative, satisfies the C-identity* $\gamma (x^{\ast }x)=\gamma (x{)}^2$, and whose completion is therefore a C*-algebra; it is required to restrict correctly, $\gamma (a\otimes b)=\Vert a\Vert \Vert b\Vert$ on elementary tensors (a cross-norm). The completion of $A\odot B$ in a C*-norm $\gamma$ is written $A{\otimes }_{\gamma }B$.

The minimal (spatial) tensor norm $\Vert \cdot {\Vert }_{\min }$ is defined through faithful representations. Choose faithful $\ast$-representations $A\subseteq B(H)$ and $B\subseteq B(K)$ 39.01.03; then $A\odot B$ acts on the Hilbert-space tensor product $H\otimes K$ by $(a\otimes b)(\xi \otimes \eta )=a\xi \otimes b\eta$, and $$ \Big| \sum_i a_i \otimes b_i \Big|{\min} = \Big| \sum_i a_i \otimes b_i \Big|{B(H \otimes K)}. $$ The defining theorem is that this norm is independent of the chosen faithful representations, so $\Vert \cdot {\Vert }_{\min }$ is well-defined; the completion is $A{\otimes }_{\min }B\subseteq B(H\otimes K)$.

The maximal tensor norm $\Vert \cdot {\Vert }_{\max }$ is the supremum over all $\ast$-representations: $$ \Big| \sum_i a_i \otimes b_i \Big|{\max} = \sup\pi \Big| \pi\Big( \sum_i a_i \otimes b_i \Big) \Big|, $$ the supremum taken over all $\ast$-representations $\pi :A\odot B\to B(H_{\pi })$. Such representations correspond exactly to pairs of commuting representations $({\pi }_A,{\pi }_B)$ of $A$ and $B$ on a common Hilbert space, via $\pi (a\otimes b)={\pi }_A(a){\pi }_B(b)$. The supremum is finite (bounded by $\Vert {\sum }_i{a}_i\otimes b_i{\Vert }_{\max }\le {\sum }_i\Vert a_i\Vert \Vert b_i\Vert$), and $A{\otimes }_{\max }B$ is its completion; by construction $\Vert \cdot {\Vert }_{\max }$ is the largest C*-norm and enjoys the universal property that every $\ast$-homomorphism out of $A\odot B$ into a C*-algebra extends continuously.

Counterexamples to common slips

• The algebraic tensor product is not already complete. $A\odot B$ is only a $\ast$-algebra; even for $A=B=K({\ell }^2)$ (compact operators) it is not norm-closed, and different completions can give genuinely different C*-algebras.
• A cross-norm is more than a norm with $\gamma (a\otimes b)=\Vert a\Vert \Vert b\Vert$. The C*-identity $\gamma (x^{\ast }x)=\gamma (x{)}^2$ on sums is the binding constraint; many submultiplicative cross-norms fail it and do not complete to a C*-algebra.
• Minimal does not mean "smallest dimension" or "fewest operators". Minimal refers to the smallest C*-norm; the spatial completion can be larger as a set than other completions are, because a smaller norm has more Cauchy sequences and thus more limit points.

Key theorem with proof Intermediate+

Theorem (the spatial norm is the smallest C-norm; $\min \le \gamma \le \max$).* For all C-algebras $A,B$ and every C*-norm $\gamma$ on $A\odot B$,* $$ |x|{\min} \le \gamma(x) \le |x|{\max} \qquad (x \in A \odot B). $$ In particular $\Vert \cdot {\Vert }_{\min }$ is independent of the faithful representations used to define it, and is the least C-norm; $\Vert \cdot {\Vert }_{\max }$ is the greatest.* ^{[Takesaki 1964; Brown-Ozawa Ch. 3]}

Proof. The upper bound is immediate: completing $A\odot B$ in any C*-norm $\gamma$ gives a C*-algebra $A{\otimes }_{\gamma }B$, hence a $\ast$-representation ${\pi }_{\gamma }:A\odot B\to A{\otimes }_{\gamma }B\hookrightarrow B(H)$ by Gelfand-Naimark 39.01.03, and $\gamma (x)=\Vert {\pi }_{\gamma }(x)\Vert$ is one of the representations in the supremum defining $\Vert \cdot {\Vert }_{\max }$. Therefore $\gamma (x)\le \Vert x{\Vert }_{\max }$ for every C*-norm $\gamma$.

The lower bound is the substance. Fix a C*-norm $\gamma$ and form the C*-algebra $C=A{\otimes }_{\gamma }B$. Let $\omega$ be a state on $C$ and run the GNS construction of $\omega$ 39.01.03 to obtain ${\pi }_{\omega }:C\to B(H_{\omega })$ with cyclic vector ${\xi }_{\omega }$. The two subrepresentations ${\pi }_A(a)={\pi }_{\omega }(a\otimes 1)$ and ${\pi }_B(b)={\pi }_{\omega }(1\otimes b)$ of $A$ and $B$ commute and are realised on the same Hilbert space. The comparison with the spatial norm rests on Takesaki's inequality: for every state $\omega$ on $A{\otimes }_{\gamma }B$ and every $x\in A\odot B$, $$ \big| \omega(x^* x) \big| \le |x|_{\min}^2 . $$

This inequality uses that on $A\odot B$ the slice maps are completely positive. For a state $\phi$ on $A$ the slice $L_{\phi }:A\odot B\to B$, $a\otimes b\mapsto \phi (a)b$, is completely positive by Stinespring 39.05.01 and contractive for any $\gamma \ge \Vert \cdot {\Vert }_{\min }$; symmetrically for slices on the $B$-leg. Consequently a product state $\phi \otimes \psi$ extends to a state on every completion $A{\otimes }_{\gamma }B$, and the family of such product states is separating and norming for the spatial norm.

Concretely, take faithful representations $A\subseteq B(H)$, $B\subseteq B(K)$ realising $\Vert \cdot {\Vert }_{\min }$. For $x={\sum }_i{a}_i\otimes b_i$ the vector states ${\phi }_{\xi }\otimes {\psi }_{\eta }$ from unit vectors $\xi \otimes \eta$ compute $\Vert x{\Vert }_{\min }={\sup }_{\xi ,\eta }\Vert x(\xi \otimes \eta )\Vert$, and each extends to a state on $A{\otimes }_{\gamma }B$ through the contractive slices. Since $\gamma (x{)}^2=\gamma (x^{\ast }x)={\sup }_{\omega }|\omega (x^{\ast }x)|$ over states $\omega$ on $A{\otimes }_{\gamma }B$, and the product vector states already attain $\Vert x{\Vert }_{\min }^2$, the bound gives $\gamma (x)\ge \Vert x{\Vert }_{\min }$.

Independence of $\Vert \cdot {\Vert }_{\min }$ from the representations follows: any two faithful representations give C*-norms ${\gamma }_1,{\gamma }_2$ each equal to a spatial norm, and the inequality just proved, applied in both directions, forces ${\gamma }_1={\gamma }_2$ to be the unique least norm. $\square$

Bridge. The min/max comparison builds toward the definition of nuclearity and exactness, and it appears again in 39.05.04 where a C*-algebra $A$ is called nuclear precisely when $\Vert \cdot {\Vert }_{\min }=\Vert \cdot {\Vert }_{\max }$ on $A\odot B$ for every $B$. The foundational reason the spatial norm is least is exactly that product vector states, built from the GNS construction 39.01.03, are norming and extend to every completion, so no C*-norm can fall below what these states already see; this is exactly the slice-map / completely-positive machinery of Stinespring 39.05.01 applied one tensor leg at a time, and the bridge is that complete positivity is what makes a single-leg evaluation $a\otimes b\mapsto \phi (a)b$ a contraction. The maximal norm is dual to the minimal one in the categorical sense: $\max$ is the universal (largest, push-out) norm receiving every representation, while $\min$ is the spatial (smallest, concrete) norm, and the central insight is that the two coincide for one factor exactly when that factor is approximable by finite-dimensional pieces. Putting these together, the gap $\max -\min$ measures the failure of a single algebra to be nuclear, and that gap is the organising quantity of the entire chapter.

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 no C*-norm on $A\odot B$: neither the spatial norm $\Vert \cdot {\Vert }_{\min }$ from a pair of faithful representations on $H\otimes K$ with its representation-independence theorem, nor the universal norm $\Vert \cdot {\Vert }_{\max }$ as a supremum over $\ast$-representations / commuting pairs, nor the sandwich $\Vert \cdot {\Vert }_{\min }\le \gamma \le \Vert \cdot {\Vert }_{\max }$, nor the completions $A{\otimes }_{\min }B$ and $A{\otimes }_{\max }B$. The intended statement reads schematically:

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

variable {A B : Type*} [CStarAlgebra A] [CStarAlgebra B]

/-- The spatial norm is the least C*-norm: every C*-norm γ on A ⊙ B
satisfies min ≤ γ ≤ max. -/
theorem cstar_tensor_min_le_max
    (γ : (TensorProduct ℂ A B) → ℝ)
    (hγ : IsCStarNorm γ) :          -- submultiplicative + C*-identity cross-norm
    ∀ x, minTensorNorm A B x ≤ γ x ∧ γ x ≤ maxTensorNorm A B x :=
  sorry  -- product vector states are norming (min ≤ γ); universality (γ ≤ max)

Advanced results Master

The min/max framework organises the entire nuclearity-exactness theory and supplies the first examples of non-uniqueness.

The functoriality and slice-map descriptions. The minimal norm is injective: if $A_0\subseteq A$ and $B_0\subseteq B$ are C*-subalgebras then $A_0{\otimes }_{\min }{B}_0\subseteq A{\otimes }_{\min }B$ isometrically. The maximal norm is projective: it is the unique norm making $A{\otimes }_{\max }B$ the universal C*-algebra receiving all commuting pairs, and it respects quotients, $(A/I){\otimes }_{\max }B=(A{\otimes }_{\max }B)/(I{\otimes }_{\max }B)$. Neither property holds for the other norm in general: $\min$ need not respect quotients (the failure is precisely the definition of exactness), and $\max$ need not respect subalgebras. The slice maps $L_{\phi }:A{\otimes }_{\gamma }B\to B$ for $\phi \in A^{\ast }$, determined by $L_{\phi }(a\otimes b)=\phi (a)b$, are bounded for $\gamma =\min$ and characterise the spatial norm through the formula $\Vert x{\Vert }_{\min }=\sup \{\Vert (\phi \otimes \psi )(x)\Vert :\cdots \}$; their continuity is the technical core of the comparison theorem.

Takesaki's example: $\min \ne \max$. The first pair with distinct tensor norms is $C^{\ast }(F_2)\otimes C^{\ast }(F_2)$, where $C^{\ast }(F_2)$ is the full group C*-algebra of the free group on two generators. Takesaki proved in 1964 that $\Vert \cdot {\Vert }_{\min }\ne \Vert \cdot {\Vert }_{\max }$ on $C^{\ast }(F_2)\odot C^{\ast }(F_2)$: the obstruction is that $C^{\ast }(F_2)$ is not nuclear, equivalently the canonical pair of commuting copies of $C^{\ast }(F_2)$ inside $B(H)$ generates a larger norm than the spatial one. The element witnessing the gap is built from the generators $u_1,u_2$ of the first copy and $v_1,v_2$ of the second: the sum ${\sum }_i{u}_i\otimes \overline{v_i}$ has strictly larger maximal norm than minimal norm. This single example shows the tensor product of C*-algebras is genuinely not a single functor — one must name the norm.

Commuting CP maps and the Connes-Kirchberg framing. The maximal norm is computed by commuting completely positive maps: $\Vert x{\Vert }_{\max }=\sup \{\Vert (\Phi \cdot \Psi )(x)\Vert \}$ over pairs of CP maps with commuting ranges, via Stinespring 39.05.01. This reframes the min/max gap as a question about whether commuting representations can be approximated by tensor-split ones. The deepest instance is the Connes embedding problem and Kirchberg's reformulation: Kirchberg proved $C^{\ast }(F_{\infty }){\otimes }_{\min }{C}^{\ast }(F_{\infty })=C^{\ast }(F_{\infty }){\otimes }_{\max }{C}^{\ast }(F_{\infty })$ is equivalent to the statement that every separable II${}_1$ factor embeds into an ultrapower of the hyperfinite factor. The 2020 refutation of Connes embedding (via ${\mathrm{MIP}}^{\ast }=\mathrm{RE}$) therefore shows this specific min/max equality fails, a striking transfer of a complexity-theoretic result into operator-algebraic tensor norms.

Nuclearity as $\min =\max$ universally. A C*-algebra $A$ is nuclear when $A{\otimes }_{\min }B=A{\otimes }_{\max }B$ for every C*-algebra $B$. Equivalently (the CP-approximation form developed in 39.05.04), the identity on $A$ factors approximately through matrix algebras by UCP maps $A\to M_n\to A$. Commutative C*-algebras, type I C*-algebras, the CAR algebra, and all finite-dimensional algebras are nuclear; $C^{\ast }(F_2)$, $C_r^{\ast }(F_2)$, and the free-group examples are not. Exactness is the weaker, subalgebra-stable property in which only the minimal norm is required to behave well under quotients.

Synthesis. The min/max sandwich is the foundational reason the tensor product of C*-algebras is not a single operation: every sensible norm lives between the spatial $\min$ and the universal $\max$, and this is exactly the statement that $\min$ is the smallest concrete norm while $\max$ is the largest universal one, dual to each other in the categorical sense of pull-back versus push-out. The central insight is that the gap $\max -\min$ vanishes for one factor precisely when that factor is finite-dimensionally approximable, so nuclearity is the equation $\min =\max$ read for all $B$, and this is exactly the completely positive approximation property that Stinespring 39.05.01 makes available. Putting these together, Takesaki's $C^{\ast }(F_2)\otimes C^{\ast }(F_2)$ example, the commuting-CP-map computation of $\max$, and the Connes-Kirchberg transfer all generalise one phenomenon — the failure of commuting representations to split as tensor products — and the bridge to the rest of the chapter is that nuclearity 39.05.04 and exactness are the two ways this gap can be made to close, the first universally over $B$ and the second stably under subalgebras.

Full proof set Master

Proposition (the spatial norm is independent of the faithful representations). Let $A\subseteq B(H_1)$, $A\subseteq B(H_2)$ be two faithful representations and likewise $B\subseteq B(K_1)$, $B\subseteq B(K_2)$. Then the two spatial norms on $A\odot B$ obtained from $H_1\otimes K_1$ and $H_2\otimes K_2$ agree. Proof: it suffices to vary one leg at a time. Fix $B\subseteq B(K)$ and two faithful representations ${\sigma }_1,{\sigma }_2$ of $A$; both are unitarily equivalent to amplifications of a common faithful representation (reduce to the universal representation of $A$), and amplification $a\mapsto a\otimes 1$ does not change the operator norm of ${\sum }_i\sigma (a_i)\otimes b_i$ because $\Vert T\otimes 1\Vert =\Vert T\Vert$. The same for the $B$-leg. Hence the spatial norm is canonical.

Proposition ($\gamma \le \max$ for every C-norm).* Completing $A\odot B$ in a C*-norm $\gamma$ yields a C*-algebra; its inclusion into $B(H)$ via Gelfand-Naimark 39.01.03 restricts to a $\ast$-representation ${\pi }_{\gamma }$ of $A\odot B$ with $\Vert {\pi }_{\gamma }(x)\Vert =\gamma (x)$. Since ${\pi }_{\gamma }$ is one representation in the defining supremum, $\gamma (x)\le \Vert x{\Vert }_{\max }$.

Proposition ($\min \le \gamma$ via norming product states). Let $\gamma$ be a C*-norm, $C=A{\otimes }_{\gamma }B$. For vector states ${\phi }_{\xi }$ on $A\subseteq B(H)$ and ${\psi }_{\eta }$ on $B\subseteq B(K)$ (realising the spatial norm), the product ${\phi }_{\xi }\otimes {\psi }_{\eta }$ defines a state on $A\odot B$ that extends to a state $\omega$ on $C$ because each slice $L_{{\phi }_{\xi }}:A\odot B\to B$, $a\otimes b\mapsto {\phi }_{\xi }(a)b$, is completely positive (it is $a\otimes b\mapsto ⟨(a\otimes b)(\xi \otimes \cdot )\xi ,\cdot ⟩$-type compression, CP by Stinespring 39.05.01) and contractive for any $\gamma \ge \min$. For positive $x^{\ast }x$, $\gamma (x{)}^2=\gamma (x^{\ast }x)={\sup }_{\omega }\omega (x^{\ast }x)\ge {\sup }_{\xi ,\eta }({\phi }_{\xi }\otimes {\psi }_{\eta })(x^{\ast }x)=\Vert x{\Vert }_{\min }^2$. Hence $\gamma (x)\ge \Vert x{\Vert }_{\min }$.

Proposition (universal property of $\max$). Every $\ast$-homomorphism $\rho :A\odot B\to C$ into a C*-algebra is $\Vert \cdot {\Vert }_{\max }$-contractive and extends to $A{\otimes }_{\max }B$. Proof: $\iota \circ \rho$ is a $\ast$-representation for a faithful $\iota :C\hookrightarrow B(H)$, hence dominated by the supremum, so $\Vert \rho (x){\Vert }_C\le \Vert x{\Vert }_{\max }$; extend by continuity. Consequently $A{\otimes }_{\max }B$ is the enveloping C*-algebra of $A\odot B$.

Proposition (finite-dimensional and commutative factors are nuclear). If $A$ is finite-dimensional then $A\odot B$ is already complete in any C*-norm: reducing to $A=M_n(\mathbb{C})$, one has $M_n(\mathbb{C})\odot B=M_n(B)$ with the unique C*-norm of the C*-algebra $M_n(B)$, so $\min =\max$. If $A=C(X)$ is commutative then $A\odot B\subseteq C(X,B)$ and $A{\otimes }_{\min }B=C(X){\otimes }_{\min }B=C(X,B)=A{\otimes }_{\max }B$, the algebra of $B$-valued continuous functions, again forcing $\min =\max$. Thus commutative and finite-dimensional C*-algebras are nuclear.

Proposition (Takesaki: $C^{\ast }(F_2)\otimes C^{\ast }(F_2)$ has $\min \ne \max$). Let $u_1,u_2$ be the canonical unitary generators of one copy of $C^{\ast }(F_2)$ and $w_1,w_2$ of a second copy, and consider the element $x=u_1\otimes w_1+u_2\otimes w_2\in C^{\ast }(F_2)\odot C^{\ast }(F_2)$. The maximal and minimal norms of $x$ differ, so the two completions are distinct C*-algebras and $C^{\ast }(F_2)$ is not nuclear ^{[Takesaki 1964]}.

For the maximal norm, take the pair of commuting representations sending both $u_i$ and $w_i$ to a single fixed unitary $s_i$ on a common space, with $s_1,s_2$ chosen freely. This is one admissible commuting pair, and on it $x$ acts as $s_1{s}_1^{\prime }+s_2{s}_2^{\prime }$; choosing the second factor to align phases shows the supremum over commuting pairs reaches the full value $\Vert x{\Vert }_{\max }=2$, the largest possible for a sum of two products of unitaries.

For the minimal norm, the spatial completion uses the left-regular representation $\lambda$ of $F_2$, where the generators act as free Haar unitaries. By Kesten's theorem the self-adjoint operator ${\sum }_{i=1}^n(\lambda (g_i)+\lambda (g_i{)}^{\ast })$ over $n$ free generators has norm $2\sqrt{2n-1}$, strictly below the value $2n$ it would have for commuting unitaries; the same free-probability spreading (Haagerup's inequality) forces $\Vert x{\Vert }_{\min }<2$ for the two-generator element $x$, while the commuting-pair computation above gives $\Vert x{\Vert }_{\max }=2$. The strict inequality $\Vert x{\Vert }_{\min }<\Vert x{\Vert }_{\max }$ is Takesaki's 1964 theorem.

The structural reason is uniform: the maximal norm sees commuting representations of $C^{\ast }(F_2)$ that need not factor through any spatial tensor product, while the minimal norm is computed inside the single free-group von Neumann algebra and feels the spectral spreading forced by non-amenability of $F_2$. Non-amenability is exactly what separates the two norms; for an amenable group the regular representation already realises every commuting pair and the gap closes.

Connections Master

• Completely positive maps and the Stinespring dilation theorem 39.05.01 — the slice maps $a\otimes b\mapsto \phi (a)b$ that compute the spatial norm and the commuting-CP-map description of the maximal norm both run on Stinespring dilations; complete positivity of single-leg evaluations is exactly why product states are norming, so the CP machinery of that unit is the engine of the min/max comparison theorem here.

• States, the GNS construction, and Gelfand-Naimark 39.01.03 — the maximal norm is a supremum over $\ast$-representations produced by GNS from states on $A{\otimes }_{\gamma }B$, and the norming product vector states that pin the minimal norm are GNS vector states on the two faithful representations; the entire sandwich $\min \le \gamma \le \max$ is read off the state spaces of the two factors.

• Nuclear C-algebras and the CPAP 39.05.04* — nuclearity is the equation $A{\otimes }_{\min }B=A{\otimes }_{\max }B$ for all $B$, equivalently the completely positive approximation property; this unit supplies the two norms whose coincidence defines that chapter's central object, making it the immediate sequel.

• Amenable groups, Folner sets, and invariant means 39.05.06 — a discrete group is amenable iff its group C*-algebra $C^{\ast }(G)$ is nuclear iff $C^{\ast }(G)=C_r^{\ast }(G)$; the non-nuclearity of $C^{\ast }(F_2)$ that drives Takesaki's $\min \ne \max$ example is exactly the non-amenability of the free group, tying the tensor-norm gap to the geometric group theory of that unit.

• Density matrix, pure and mixed states 12.17.01 — the composite-quantum-system picture, where the joint algebra of two subsystems is a tensor product and entangled states are non-product states, is the finite-dimensional shadow of this theory; there $\min =\max$ because matrix algebras are nuclear, which is why finite quantum systems never see the tensor-norm ambiguity that infinite ones do.

Historical & philosophical context Master

The systematic theory of C*-algebra tensor products begins with Masamichi Takesaki, who in 1964 proved that the spatial (minimal) cross-norm is the smallest C*-norm on the algebraic tensor product and exhibited the first pair of C*-algebras for which the minimal and maximal norms differ, namely two copies of the full group C*-algebra of the free group $F_2$ ^{[Takesaki 1964]}. This settled that the tensor product of C*-algebras is not a single canonically normed object, in sharp contrast to the unique norm of a single C*-algebra, and it isolated the property — later named nuclearity by Lance — of an algebra for which the ambiguity disappears against every partner. The general framework was consolidated in Takesaki's 1979 monograph and, for the operator-space refinements, by Gilbert Effros, Zhong-Jin Ruan, and Gilles Pisier ^{[Pisier 2020]}.

The physical and logical reach of the min/max gap became clear later. Eberhard Kirchberg connected the equality $C^{\ast }(F_{\infty }){\otimes }_{\min }{C}^{\ast }(F_{\infty })=C^{\ast }(F_{\infty }){\otimes }_{\max }{C}^{\ast }(F_{\infty })$ to Alain Connes's 1976 embedding problem for II${}_1$ factors, recasting a question about von Neumann algebras as one about a single tensor-norm coincidence. The 2020 proof that the complexity class ${\mathrm{MIP}}^{\ast }$ equals $\mathrm{RE}$, by Ji, Natarajan, Vidick, Wright, and Yuen, refuted the Connes embedding conjecture and therefore established that this minimal-equals-maximal equality fails, an unexpected route by which a result in quantum-complexity theory determined the value of a C*-algebraic tensor norm. The Brown-Ozawa development organises the subject around finite-dimensional approximation, presenting nuclearity and exactness as the two regimes in which the min/max gap is controlled ^{[Brown-Ozawa 2008]}.

Bibliography Master

• Takesaki, M., "On the cross-norm of the direct product of C*-algebras", Tohoku Mathematical Journal 16 (1964), 111-122.
• Takesaki, M., Theory of Operator Algebras I, Springer, 1979. Ch. IV §1-4.
• Lance, E. C., "On nuclear C*-algebras", Journal of Functional Analysis 12 (1973), 157-176.
• Wassermann, S., Exact C-Algebras and Related Topics*, Lecture Notes Series 19, Seoul National University, 1994.
• Brown, N. P. and Ozawa, N., C-Algebras and Finite-Dimensional Approximations*, Graduate Studies in Mathematics 88, American Mathematical Society, 2008. Ch. 3.
• Pisier, G., Tensor Products of C-Algebras and Operator Spaces: The Connes-Kirchberg Problem*, London Mathematical Society Student Texts 96, Cambridge University Press, 2020.
• Ji, Z., Natarajan, A., Vidick, T., Wright, J., and Yuen, H., "MIP* = RE", Communications of the ACM 64 (2021), 131-138.

---

Operator-algebras spine, second structural unit of the nuclearity-exactness chapter. Produced as the C-tensor-product anchor: the algebraic tensor product $A\odot B$ and the C*-norms it can carry, the minimal (spatial) and maximal tensor norms and the sandwich $\min \le \gamma \le \max$, Takesaki's $C^{\ast }(F_2)\otimes C^{\ast }(F_2)$ example of $\min \ne \max$, the commuting-CP-map and Connes-Kirchberg framing, and nuclearity as $\min =\max$ for all $B$ previewing 39.05.04.*