---
title: Weak Dixmier Approximation Theorem
url: https://www.emergentmind.com/topics/weak-dixmier-approximation-theorem
type: topic
---

# Weak Dixmier Approximation Theorem

The weak Dixmier approximation theorem is a family of averaging principles in operator algebras asserting that an element can be driven, by convex combinations of unitary conjugates or related unital completely positive averages, toward central data. In the von Neumann algebra setting, the classical form says that the convex hull of the unitary orbit of an element meets the center; in finite factors, the target is the scalar trace. Relative versions replace the center by a relative commutant for an inclusion \(N\subset M\). In \(C^*\)-algebra theory, closely related results are typically formulated through Dixmier sets, weak centrality, local averaging criteria, and uniform finite averaging estimates rather than under a single universal title [1907.00615][1304.3523][1611.08263].

## 1. Classical formulation and the meaning of “weak”

For a von Neumann algebra \(M\), the classical Dixmier phenomenon concerns the unitary orbit
\[
\{uxu^*:u\in \mathcal U(M)\}
\]
of an element \(x\in M\). In the weak form recalled in the relative literature, one asks that
\[
\overline{\operatorname{co}^{\,w^*}\bigl(\{uxu^*:u\in\mathcal U(M)\}\bigr)}\cap Z(M)\neq \varnothing,
\]
where \(Z(M)=M'\cap M\) is the center. This is the absolute case \(N=M\) of the relative formulation. In a finite factor \((\mathcal M,\tau)\), the conclusion specializes to the scalar trace:
\[
\tau(X)I \in \overline{\operatorname{conv}\{U^*XU : U\in \mathcal U(\mathcal M)\}}.
\]
The closure depends on the ambient formulation: in finite von Neumann algebras and factors one often uses norm closure for self-adjoint elements, while weak-operator or ultraweak closure also appears in general formulations [1907.00615][2303.10602].

The adjective “weak” does not have a completely uniform role across the literature. In the relative von Neumann algebra setting, it refers to weak\(^*\) or \(\sigma\)-weak closure of the convex hull. In \(C^*\)-algebra theory, the closest analogues are frequently norm-closed Dixmier sets
\[
D_A(a)=\overline{\operatorname{co}\{uau^*:u\in\mathcal U(A)\}},
\]
together with exact or approximate intersection with the center. This difference in formulation is part of the reason the phrase “weak Dixmier approximation theorem” functions more as a thematic label than as a single canonical theorem across all operator-algebraic settings [2106.00098][1611.08263].

## 2. Relative weak Dixmier approximation for inclusions of von Neumann algebras

For an inclusion \(N\subset M\) of von Neumann algebras, Popa’s weak relative Dixmier property requires that for every \(x\in M\),
\[
\overline{\operatorname{co}^{\,w^*}\bigl(\{uxu^*:u\in\mathcal U(N)\}\bigr)}\cap (N'\cap M)\neq \varnothing.
\]
Here the averaging unitaries come only from \(N\), and the limiting object lies in the relative commutant \(N'\cap M\). Marrakchi proved that every inclusion of von Neumann algebras with faithful normal conditional expectation has this property. Equivalently, if \(E_N:M\to N\) is faithful normal, then \(N\subset M\) has the weak relative Dixmier property [1907.00615].

A central structural reformulation uses the weak\(^*\)-closed convex semigroup
\[
D(N,M)\subset \mathrm{UCP}(M)
\]
generated by the inner automorphisms \(\operatorname{Ad}(u)\), \(u\in \mathcal U(N)\). The fixed-point algebra of this semigroup is
\[
M^{D(N,M)}=N'\cap M,
\]
and the inclusion \(N\subset M\) has the weak relative Dixmier property if and only if \(D(N,M)\) contains a conditional expectation onto \(N'\cap M\). This is stronger than a pointwise intersection statement, because a single map \(\Phi\in D(N,M)\) then sends every \(x\in M\) into \(N'\cap M\) [1907.00615].

The proof uses compact convex semigroups of ucp maps and an improvement of Ellis’ lemma. For a compact convex semigroup \(K\), every minimal idempotent \(e\in K\) satisfies
\[
exe=e \qquad \text{for all }x\in K.
\]
Applied to \(D(N,M)\), this rigidity shows that a faithful minimal idempotent is automatically a conditional expectation onto the fixed-point algebra. The difficult case is when \(N\) is a type \(\mathrm{III}_1\) factor. There the key lemma states that for a faithful normal state \(\varphi\) on \(N\),
\[
\varphi\in D(N,N),
\]
with \(\varphi\) viewed as the ucp map \(x\mapsto \varphi(x)1\). The construction uses an irreducible AFD type \(\mathrm{III}_1\) subfactor with expectation together with Connes–Størmer transitivity [1907.00615].

The theorem pinpoints the expectation hypothesis as decisive. Without expectation, the relative property can fail: \(M\subset B(H)\) has the weak relative Dixmier property if and only if \(M\) is AFD, and for a type \(\mathrm{III}_1\) factor \(M\), the inclusion of its continuous core has the weak relative Dixmier property if and only if \(M\) has trivial bicentralizer. It was already elementary when \(M\) is tracial or when there exists a faithful normal state \(\varphi\) with \(N\subset M_\varphi\), but Marrakchi’s theorem shows that the full expected case is always covered [1907.00615].

## 3. Local \(C^*\)-algebraic variants: Dixmier sets, Magajna sets, and quotient criteria

For a unital \(C^*\)-algebra \(A\), a unitary mixing operator is a map
\[
\phi(x)=\sum_{i=1}^n t_i u_i x u_i^*
\]
with \(u_i\in U(A)\), \(t_i\ge 0\), and \(\sum_i t_i=1\). The associated Dixmier set of \(a\in A\) is
\[
D_A(a)=\overline{\{\phi(a):\phi\in \operatorname{Av}(A,U(A))\}},
\]
the norm-closed convex hull of the unitary orbit. The algebra \(A\) has the Dixmier property exactly when
\[
D_A(a)\cap Z(A)\neq \varnothing \qquad (\forall a\in A).
\]
The local theory isolates individual elements: \(a\) is a Dixmier element if \(D_A(a)\cap Z(A)\neq\varnothing\), and \(a\) belongs to the approximate class \(\overline{\operatorname{Dix}(A)}\) if \(\operatorname{dist}(D_A(a),Z(A))=0\) [2106.00098].

A parallel weak-centrality theory replaces unitary mixing by unital completely positive elementary operators
\[
\phi(x)=\sum_{i=1}^n a_i x a_i^*, \qquad \sum_{i=1}^n a_i a_i^*=1.
\]
The corresponding Magajna set
\[
M_A(a)=\overline{\{\phi(a):\phi\in \operatorname{EUCP}(A)\}}
\]
gives the local weak-centrality classes \(\operatorname{Mag}(A)\) and \(\overline{\operatorname{Mag}(A)}\). Since every unitary mixing operator is a unital completely positive elementary operator,
\[
D_A(a)\subseteq M_A(a),
\]
hence
\[
\operatorname{Dix}(A)\subseteq \operatorname{Mag}(A),\qquad \overline{\operatorname{Dix}(A)}\subseteq \overline{\operatorname{Mag}(A)}.
\]
The paper emphasizes that exact central averaging and mere distance zero need not coincide for general elements, although they do coincide for selfadjoint ones [2106.00098].

The geometric object controlling this theory is the multifunction
\[
\Psi_a(N):=\bigcap_{M\in \operatorname{Max}^N(A)} W_{A/M}(a+M), \qquad N\in \operatorname{Glimm}(A),
\]
built from quotient numerical ranges. For Magajna averaging, a central element \(z\in Z(A)\) lies in \(M_A(a)\) exactly when \(z(N)\in \Psi_a(N)\) for all \(N\in \operatorname{Glimm}(A)\). Theorem 3.3 states that for \(a\in A\),
\[
(i)\Rightarrow(ii)\iff(iii),
\]
where \((i)\) is \(a\in \operatorname{Mag}(A)\), \((ii)\) is \(\operatorname{dist}(M_A(a),Z(A))=0\), and \((iii)\) is \(\Psi_a(N)\neq \varnothing\) for all \(N\in \operatorname{Glimm}(A)\); if \(a\) is selfadjoint, then all three are equivalent [2106.00098].

For genuine Dixmier averaging, traces enter. If
\[
Y=\{N\in \operatorname{Glimm}(A): T(A/N)\neq \varnothing\},
\]
then Proposition 4.2 characterizes exact central values \(z\in D_A(a)\cap Z(A)\) by the simultaneous conditions
\[
z(N)=\tau(a+N)\quad \text{for all }N\in Y,\ \tau\in T(A/N),
\]
and
\[
z(N)\in \Psi_a(N)\quad \text{for all }N\in \operatorname{Glimm}(A).
\]
Theorem 4.4 gives the approximate version: \(\operatorname{dist}(D_A(a),Z(A))=0\) if and only if the tracial value function exists on \(Y\), lies in \(\Psi_a(N)\) there, and \(\Psi_a(N)\neq\varnothing\) on \(\operatorname{Glimm}(A)\setminus Y\). Again, if \(a\) is selfadjoint, approximate and exact central averaging coincide. This is one of the sharpest elementwise formulations of weak Dixmier-type approximation in the \(C^*\)-setting [2106.00098].

## 4. Traces, weak centrality, QTS, and uniform averaging

A global characterization of the Dixmier property for unital \(C^*\)-algebras states that \(A\) has the Dixmier property if and only if it is weakly central and satisfies specific tracial conditions on maximal quotients. More precisely, weak centrality means that the map
\[
\Psi:{\rm Max}(A)\to {\rm Max}(Z(A)),\qquad \Psi(M)=M\cap Z(A),
\]
is injective, and the tracial conditions require quotientwise uniqueness and factorization behavior for tracial states. When \(A\) has the singleton Dixmier property, there is a unique center-valued trace \(R\) and
\[
D_A(a)\cap Z(A)=\{R(a)\}\qquad (a\in A).
\]
The same work gives a distance formula:
\[
d(D_A(a),D_A(b))
\]
is the minimum \(r\ge 0\) such that \(|\tau(a-b)|\le r\) for all tracial states and
\[
d(W_{A/M}(q_M(a)),W_{A/M}(q_M(b)))\le r
\]
for all maximal ideals \(M\). This makes the geometry of Dixmier approximation explicitly trace- and quotient-controlled [1611.08263].

The uniform Dixmier property strengthens existence to uniform finite averaging. It requires that for every \(\epsilon>0\) there exists \(n\in\mathbb N\) such that for all \(a\in A\) one can choose unitaries \(u_1,\dots,u_n\) with
\[
d\Big(\sum_{i=1}^n \frac1n u_i a u_i^*,\, Z(A)\Big)\le \epsilon \|a\|.
\]
Equivalently, there exist constants \((m,\gamma)\), with \(m\in\mathbb N\) and \(0<\gamma<1\), such that for every self-adjoint \(a\in A\),
\[
\Big\|\sum_{i=1}^m \frac1m u_i a u_i^* - z\Big\|\le \gamma \|a\|
\]
for some central \(z\). Von Neumann algebras satisfy this uniform property, but not all \(C^*\)-algebras with the Dixmier property do. The literature also supplies broad sufficient conditions, such as finite radius of comparison-by-traces [1611.08263].

Ozawa’s trace-sensitive theorem provides a complementary form of weak Dixmier approximation. For a unital \(C^*\)-algebra \(A\), the following are equivalent: \(A\) has the QTS property, meaning every non-zero quotient of \(A\) has a tracial state; and for every \(\varepsilon>0\) and \(a\in A\) with
\[
\sup_{\tau\in T(A)}|\tau(a)|<\varepsilon,
\]
there are \(u_1,\dots,u_k\in \mathcal U(A)\) such that
\[
\left\|\frac1k\sum_{i=1}^k u_i a u_i^*\right\|<\varepsilon.
\]
In particular, if \(\tau(a)=0\) for all \(\tau\in T(A)\), then
\[
0 \in \overline{\operatorname{co}\{uau^*:u\in \mathcal U(A)\}}.
\]
If \(A\) has a unique tracial state \(\tau\), applying this to \(a-\tau(a)1\) yields the scalar weak Dixmier conclusion
\[
\tau(a)1 \in \overline{\operatorname{co}\{uau^*:u\in \mathcal U(A)\}}.
\]
Under nuclearity, QTS is equivalent to symmetric amenability, so the trace-controlled averaging property is tied to a global homological regularity condition [1304.3523].

## 5. Exact finite averaging in type \({\rm II}\) factors

In type \({\rm II}\) factors, the weak Dixmier approximation phenomenon admits a sharp exact strengthening. Given a type \({\rm II}\) factor \((\mathcal M,\tau)\) and finitely many operators \(X_1,\dots,X_n\in \mathcal M\), there exist mutually orthogonal nonzero projections \(E_1,\dots,E_N\in\mathcal M\) such that
\[
I=\sum_{j=1}^N E_j,\qquad E_jX_iE_j=\tau(X_i)E_j
\]
for all \(j\) and all \(i\). Equivalently, there exists a unitary \(U\in\mathcal M\) such that
\[
\frac1N\sum_{j=0}^{N-1}(U^*)^j X_i U^j=\tau(X_i)I,\qquad i=1,\dots,n.
\]
This is strictly stronger than weak Dixmier approximation: it gives exact equality instead of approximation, uses a single finite cyclic average, and works simultaneously for a finite family [2303.10602].

The proof reduces to the self-adjoint trace-zero case, treats the two-variable case by block decompositions, and then proceeds by induction. Its structural ingredients are the factor property, the diffuse trace, maximal abelian subalgebras, polar decomposition, and repeated reduction to balanced \(2\times2\) and \(4\times4\) block forms. A key splitting lemma states that if \(A=PAQ\), then \(P\) and \(Q\) can be split into equal-trace halves
\[
P=P_1+P_2,\qquad Q=Q_1+Q_2,
\]
with
\[
P_2AQ_1=P_1AQ_2=0.
\]
This allows the off-diagonal structure to be arranged so that exact blockwise trace cancellation becomes possible [2303.10602].

The applications show how far exact finite averaging goes beyond mere approximation. Every trace-zero element in a type \({\rm II}_1\) factor is a single commutator, every self-adjoint trace-zero element is a single self-commutator, every self-adjoint element is a linear combination of \(4\) projections, and every operator in a finite factor decomposes as
\[
X=N+K
\]
with \(N\) normal and \(K\) nilpotent. These consequences depend on exact elimination of diagonal blocks, not on asymptotic convex-hull arguments [2303.10602].

## 6. Operator-valued weights, type \({\rm III}\) subfactors, and the modern relative theory

A further extension replaces conditional expectations by faithful normal semifinite operator-valued weights. If
\[
A \subset M
\]
is an inclusion of von Neumann algebras with faithful normal semifinite operator-valued weight
\[
E_A\colon M\to A,
\]
then every positive element \(x\in M\) with \(E_A(x)<\infty\) satisfies
\[
\overline{\operatorname{co}^{\,\sigma\text{-weak}\{uxu^*:u\in \mathcal U(A)\}}}\cap (A'\cap M)\neq \varnothing.
\]
Writing
\[
\mathfrak n_{E_A}:=\{x\in M\mid \|E_A(x^*x)\|_\infty<\infty\},\qquad \mathfrak m_{E_A}:=(\mathfrak n_{E_A})^*\mathfrak n_{E_A},
\]
the stronger statement is that there exists a single ucp map
\[
\Psi\in DSG(A\subset M)
\]
such that
\[
\Psi(\mathfrak m_{E_A})\subset A'\cap M.
\]
When \(E_A\) is a faithful normal conditional expectation, \(\mathfrak m_{E_A}=M\), so Marrakchi’s theorem is recovered [2508.17592].

The proof separates semifinite and type \({\rm III}\) parts. In the semifinite case, one uses a faithful normal semifinite invariant weight and an \(L^2\)-minimization argument on a weakly closed convex hull. In the type \({\rm III}\) case, the core mechanism is different: the \(E_A\)-finite algebra \(\mathfrak m_{E_A}\) is shown to be stable under averaging, minimal idempotents in the weak Dixmier semigroup are combined with a Choi–Effros product argument, and a rigidity computation forces the image algebra into the relative commutant. A decisive estimate is the operator-valued-weight inequality
\[
E_A(\theta(x))\le \theta(E_A(x))
\]
for \(x\in \mathfrak m_{E_A}^+\) and \(\theta\in DSG(A\subset M)\), which ensures that averaging preserves the integrable domain [2508.17592].

This operator-valued-weight version is especially significant in type \({\rm III}\) theory. Applied inside the basic construction \(\langle M,B\rangle\), it produces nonzero elements in relative commutants that remain finite for the canonical operator-valued weight, and this is the averaging step needed for a type \({\rm III}\) reformulation of Popa’s intertwining criterion without tracial assumptions. The same work applies the theorem to an extension of Ozawa’s relative solidity theorem to the type \({\rm III}\) setting and to a Galois-type correspondence for crossed products by totally disconnected groups [2508.17592].

Across these forms, the weak Dixmier approximation theorem is best understood not as a single isolated statement but as an organizing principle: central or relatively central information is extracted from an element by averaging its unitary orbit, with the relevant closure, target algebra, and admissible averaging operators determined by the ambient category—finite factor, general von Neumann inclusion, unital \(C^*\)-algebra, or operator-valued-weight setting. The modern theory shows that the decisive obstructions are tracial data, quotient numerical ranges, and the presence or absence of conditional expectations or operator-valued weights [1907.00615][2106.00098][1611.08263][1304.3523][2508.17592].

Source: https://www.emergentmind.com/topics/weak-dixmier-approximation-theorem