---
title: Popa's Intertwining Criterion
url: https://www.emergentmind.com/topics/popa-s-intertwining-criterion
type: topic
---

# Popa's Intertwining Criterion

Popa’s intertwining criterion, usually written \(A \preceq_M B\), is a criterion for deciding when a corner of a von Neumann subalgebra \(A\) embeds into a corner of another subalgebra \(B\) inside an ambient von Neumann algebra \(M\). In the finite tracial setting it is equivalent to the existence of an intertwining partial isometry, to the presence of a finite-right-dimensional \(A\)-\(B\) bimodule inside \(L^2(M)\), and to the failure of a certain asymptotic vanishing condition for conditional expectations; later work extended the criterion to \(\sigma\)-finite, nontracial, and type III settings by replacing tracial \(L^2\)-methods with modular theory, continuous cores, basic constructions, operator-valued weights, and weak relative Dixmier techniques [1902.01049], [2508.17592].

## 1. Classical tracial formulation

Let \(M\) be a separable II\(_1\) factor with faithful normal trace \(\tau\), and let \(A,B \subset M\) be von Neumann subalgebras. In the standard formulation, \(A \preceq_M B\) means that \(A\) embeds into \(B\) inside \(M\) in the sense of intertwining-by-bimodules. One standard form of the criterion states that the following are equivalent [2003.10004].

First, there exist projections \(p \in A\) and \(q \in B\), a \(*\)-homomorphism
\[
\theta : pAp \to qBq,
\]
and a nonzero partial isometry \(v \in qMp\) such that
\[
\theta(x)v = vx \qquad \text{for all } x \in pAp.
\]

Second, there exists a nonzero \(A\)-\(B\) sub-bimodule \(\mathcal H \subset L^2(M)\) whose right \(B\)-dimension is finite:
\[
\dim_B(\mathcal H) < \infty.
\]

Third, there is no sequence of unitaries \((u_n) \subset \mathcal U(A)\) such that
\[
\|E_B(xu_n y)\|_2 \xrightarrow[n\to\infty]{} 0
\qquad \text{for all } x,y \in M,
\]
where \(E_B : M \to B\) is the trace-preserving conditional expectation.

This formulation isolates the criterion’s central analytic content. The bimodule picture expresses a representation-theoretic relation between \(A\) and \(B\); the partial-isometry picture gives a concrete intertwiner inside \(M\); and the unitary-sequence formulation turns the embedding problem into a vanishing/nonvanishing test for conditional expectations.

Several common misunderstandings are excluded by the classical formulation itself. The relation \(A \preceq_M B\) does not mean \(A \subset B\); it means that a corner of \(A\) can be conjugated into a corner of \(B\) by a partial isometry. It is weaker than unitary conjugacy, which would require a unitary \(u \in M\) with \(uAu^* \subset B\), and stronger than information about intersections or relative commutants alone, because it detects an actual \(A\)-\(B\) bimodule link inside \(L^2(M)\) [2003.10004].

## 2. \(\sigma\)-finite and type III generalizations

In the nontracial setting, the ambient algebra is a \(\sigma\)-finite von Neumann algebra \(M\), and the relevant inclusions \(A,B \subset M\) are assumed to be with expectation. Isono formulates \(A \preceq_M B\) by requiring projections \(e \in A\), \(f \in B\), a partial isometry \(v \in eMf\), and a unital normal \(*\)-homomorphism
\[
\theta : eAe \to fBf
\]
such that \(\theta(eAe) \subset fBf\) is with expectation and
\[
av = v\,\theta(a) \qquad \text{for all } a \in eAe
\]
[1902.01049].

A more flexible formulation uses an auxiliary separable Hilbert space \(H\), a fixed minimal projection \(e_{1,1} \in B(H)\), a projection \(f \in B \,\overline\otimes\, B(H)\), a unital normal \(*\)-homomorphism
\[
\pi : A \to f(B \,\overline\otimes\, B(H))f
\]
such that \(\pi(A)\) is with expectation, and a partial isometry
\[
w \in (1_A \otimes e_{1,1})(M \,\overline\otimes\, B(H))f
\]
satisfying
\[
w\,\pi(a) = (a \otimes e_{1,1})\,w = (a \otimes 1)\,w
\qquad \forall a \in A.
\]
This preserves the original intertwining structure while accommodating nontracial basic-construction methods [2508.17592].

The general analytic criterion replaces tracial \(L^2\)-norms by strong or \(\sigma\)-strong convergence. In Isono’s Popa–Houdayer–Isono criterion, if \(A\) is finite, then \(A \preceq_M B\) is equivalent to the nonexistence of a net of unitaries \((u_i)_i \subset \mathcal U(A)\) such that
\[
E_B(b^*u_i a) \to 0 \quad \text{strongly for all } a,b \in M1_B.
\]
The same framework uses the basic construction \(\langle M,\widetilde B\rangle\), the Jones projection \(e_B\), and the canonical operator-valued weight \(\widehat E_{\widetilde B}\) to reformulate intertwining as the existence of a nonzero positive element
\[
d \in A' \cap 1_A\langle M,\widetilde B\rangle 1_A
\]
with
\[
d = dJ1_BJ, \qquad \widehat E_{\widetilde B}(d) \in M
\]
[1902.01049].

The 2025 reformulation removes tracial assumptions altogether for large classes of inclusions. Under either the hypothesis that \(A\) has no direct summand that is semifinite and properly infinite, or that \(B\) is properly infinite, the following are equivalent: \(A \preceq_M B\); the absence of a net \((u_i)\subset \mathcal U(A)\) with
\[
E_B(b^*u_i a)\to 0 \quad \sigma\text{-strongly for all } a,b\in M1_B;
\]
the existence of a positive element \(d_0 \in \langle M,\widetilde B\rangle^+\) with
\[
\widehat E_{\widetilde B}(d_0)<\infty,\qquad d_0=d_0J1_BJ,\qquad 0\notin \mathcal K(d_0,A),
\]
where
\[
\mathcal K(d_0,A)=\overline{\mathrm{conv}^{\rm weak}\{u d_0 u^* \mid u\in\mathcal U(A)\}};
\]
and the nonexistence of certain unital completely positive maps in the weak Dixmier semigroup \(DSG(A\subset 1_A\langle M,\widetilde B\rangle 1_A)\) that vanish on the basic-construction coefficients \(ae_{\widetilde B}b^*\) [2508.17592].

## 3. Modular actions and continuous cores

A major refinement of the criterion relates intertwining to modular theory. Let \(\varphi\) and \(\psi\) be faithful normal states preserved by conditional expectations onto \(A\) and \(B\), respectively. Isono defines
\[
(A,\sigma^\varphi)\preceq_M (B,\sigma^\psi)
\]
by requiring a witness \((e,f,\theta,v)\) for \(A\preceq_M B\), with \(e \in A_\varphi\), together with a generalized cocycle \((u_t)_{t\in\mathbb R}\) for the modular action \(\sigma^\psi\) with support projection \(f\), such that if \(\omega_t=[D\varphi:D\psi]_t\), then
\[
v u_t = \omega_t\,\sigma_t^\varphi(v) \qquad \text{for all } t\in\mathbb R,
\]
and
\[
u_t\,\sigma_t^\psi(\theta(a))\,u_t^*
= \theta(\sigma_t^\varphi(a))
\qquad \text{for all } a\in eAe,\ t\in\mathbb R
\]
[1902.01049].

The first part of Isono’s Theorem A states that \(A \preceq_M B\) is equivalent to the existence of a faithful normal state \(\varphi\), preserved by some conditional expectation onto \(A\), such that
\[
(A,\sigma^\varphi) \preceq_M (B,\sigma^\psi).
\]
Thus an intertwining element can be chosen so that it also aligns appropriate modular flows.

The second part of Theorem A passes to continuous cores. For a faithful normal semifinite weight \(\varphi\), the continuous core is
\[
\mathcal C_\varphi(M) := M \rtimes_{\sigma^\varphi} \mathbb R.
\]
After tensoring with a type III\(_1\) factor \(N\) equipped with a faithful normal state \(\omega\), Isono proves that the following are equivalent: modular-flow intertwining, conditional-expectation intertwining, and the intertwining of continuous cores,
\[
\mathcal C_\omega(AN) \preceq_{\mathcal C_\omega(MN)} \mathcal C_\omega(BN),
\]
inside the core of \(M\overline\otimes N\) [1902.01049].

This core characterization is structurally significant because \(\mathcal C_\varphi(M)\) is semifinite, so one recovers access to the original analytic machinery of Popa’s criterion in a nontracial environment. Isono emphasizes that tensoring with a type III\(_1\) factor is necessary: without the tensor product, one can have \(A \preceq_M B\) while the corresponding core inclusion fails [1902.01049].

## 4. Weak relative Dixmier property and structural applications

The operator-valued-weight approach yields a nontracial replacement for tracial \(L^2\)-averaging. Let \(A\subset M\) be an inclusion with a faithful normal semifinite operator-valued weight
\[
E_A : M \to A.
\]
For \(x\in M\), define the \(\sigma\)-weakly closed convex hull of its \(A\)-unitary orbit by
\[
\mathcal K(x,A):=
\overline{\mathrm{conv}^{\rm weak}\{uxu^*\mid u\in\mathcal U(A)\}}.
\]
Isono proves that for every positive element \(x\in M\) with \(E_A(x)<\infty\), the set \(\mathcal K(x,A)\) intersects \(A'\cap M\). This is the weak relative Dixmier property for integrable positive elements, and it extends Marrakchi’s conditional-expectation нәтиҗيجة to the operator-valued-weight setting [2508.17592].

The same paper uses this property to derive a general nontracial reformulation of Popa’s criterion. The implication from the existence of \(d_0\) with \(0\notin \mathcal K(d_0,A)\) to \(A\preceq_M B\) proceeds by extracting a nonzero central element from the convex hull, thereby producing the positive intertwiner needed in the basic construction. In this sense, weak relative Dixmier averaging replaces tracial Hilbert-space averaging [2508.17592].

These extensions feed directly into rigidity theory. Isono’s 2019 paper uses the criterion and its modular/core refinements to analyze unitary conjugacy of subfactors, crossed-product decompositions, W\(^*\)-superrigidity, and stable strong solidity. In particular, mutual intertwining can force unitary conjugacy under suitable irreducibility and factoriality hypotheses; the paper’s Theorem C gives a W\(^*\)-superrigidity result for Bernoulli shifts of ICC countable discrete groups in the class \(\mathcal C\) acting on amenable type III\(_1\) factors; and Theorem G and Corollary H characterize stable strong solidity for amalgamated free products and free product factors in terms of intertwining alternatives and the behavior of stable normalizers [1902.01049].

The 2025 paper develops a complementary set of type III applications. For a crossed product \(M=B\rtimes_\alpha\Gamma\) by a discrete bi-exact group \(\Gamma\) acting on an amenable \(\sigma\)-finite von Neumann algebra \(B\), Theorem E states that \(M\) is solid relative to \(B\): if \(A\subset pMp\) is with expectation and \(A\not\preceq_M B\), then \(A'\cap pMp\) is amenable. The same techniques yield a Galois-type correspondence for crossed products by totally disconnected groups: under the stated outerness and minimality assumptions, every intermediate subfactor between \(B\) and \(B\rtimes G\) is of the form \(B\rtimes H\) for some closed subgroup \(H\le G\) [2508.17592].

## 5. Spectral gap, factorial commutants, and model-theoretic reformulations

Although Goldbring’s paper on Popa’s factorial commutant embedding problem does not restate the intertwining criterion explicitly, it is organized around rigidity inputs that are closely related to Popa-style spectral-gap/intertwining phenomena. The motivating question asks whether every \(R^{\mathcal U}\)-embeddable II\(_1\) factor \(N\) admits an embedding
\[
T:N\hookrightarrow R^{\mathcal U}
\]
such that the relative commutant
\[
T(N)'\cap R^{\mathcal U}
\]
is a factor [2003.10004].

Goldbring replaces \(R\) by a locally universal McDuff II\(_1\) factor \(M\) and studies property (T) factors. The central rigidity notion is \(w\)-spectral gap: a subfactor \(N\subset M\) has \(w\)-spectral gap in \(M\) if
\[
N' \cap M^{\mathcal U} = (N'\cap M)^{\mathcal U}.
\]
The paper recalls that property (T) factors have \(w\)-spectral gap in any extension. A key structural fact states that if \(N\) is a \(w\)-spectral gap subfactor of an existentially closed factor \(M\), then
\[
(N'\cap M)'\cap M = N.
\]
From this, Goldbring deduces that \(N'\cap M\) is a factor, and hence
\[
N'\cap M^{\mathcal U}=(N'\cap M)^{\mathcal U}
\]
is also a factor [2003.10004].

The main theorem then asserts that if \(M\) is an infinitely generic II\(_1\) factor, then for any property (T) II\(_1\) factor \(N\), there exists an embedding
\[
T:N\hookrightarrow M^{\mathcal U}
\]
such that
\[
T(N)'\cap M^{\mathcal U}
\]
is a factor. Since infinitely generic factors are existentially closed, locally universal, and McDuff, this gives a fixed McDuff locally universal ambient factor whose ultrapower admits factorial relative commutants for all property (T) factors. The paper describes this as a “poor man’s resolution” of Popa’s factorial commutant embedding problem for property (T) factors, because the target is \(M^{\mathcal U}\) rather than \(R^{\mathcal U}\) [2003.10004].

This connection is conceptually close to intertwining theory. Spectral gap prevents the appearance of asymptotically central sequences, and Popa’s criterion detects the failure of asymptotic disjointness through conditional expectations. The papers do not identify these frameworks, but they place them on the same deformation/rigidity axis.

## 6. Interpretation, scope, and recurrent points of confusion

The criterion’s most persistent conceptual feature is that it converts an embedding problem into an obstruction to averaging away a subalgebra. In the tracial II\(_1\) case, one tests whether unitaries in \(A\) can force all coefficients \(E_B(xu_n y)\) to vanish in \(L^2\)-norm; failure of this vanishing is exactly the existence of an intertwiner. In the \(\sigma\)-finite and type III settings, the same philosophy survives, but traces are replaced by strong-topology convergence, basic constructions, operator-valued weights, and completely positive averaging semigroups [2003.10004], [2508.17592].

A first misconception is to identify \(A\preceq_M B\) with inclusion. The criterion says only that a corner of \(A\) embeds into a corner of \(B\), generally after compression and conjugation by a partial isometry. A second misconception is to treat it as merely a statement about relative commutants. Relative commutants measure commutation, whereas intertwining detects finite-dimensional \(A\)-\(B\) bimodule structure and therefore an actual representation-theoretic relation inside the ambient algebra [2003.10004].

A third recurrent point concerns the nontracial setting. The absence of a faithful normal trace does not eliminate intertwining theory; rather, it changes its analytic realization. Isono’s modular formulation shows that an intertwiner can be made compatible with modular flows, and his later operator-valued-weight formulation shows that weak relative Dixmier averaging of integrable positive elements can play the role that \(L^2\)-averaging played in the classical proof [1902.01049], [2508.17592].

Taken together, these developments locate Popa’s intertwining criterion at the center of modern deformation/rigidity theory. In II\(_1\) factors it is the standard device for converting asymptotic orthogonality into algebraic embedding; in type III theory it becomes a modular and core-compatible tool; and in model-theoretic work on factorial commutant embeddings it reappears indirectly through spectral-gap rigidity and ultrapower commutant control [1902.01049], [2003.10004].

Source: https://www.emergentmind.com/topics/popa-s-intertwining-criterion