---
title: Entanglement Purification Gap
url: https://www.emergentmind.com/topics/entanglement-of-purification-gap
type: topic
---

# Entanglement Purification Gap

The entanglement of purification gap concerns the difference between the purification-based correlation measure \(E_P(A:B)\) and other quantities that are easier to compute, more geometric, or more directly tied to lower bounds. For a bipartite mixed state \(\rho_{AB}\), the entanglement of purification is defined by minimizing the entanglement entropy over all purifications, \(E_P(A:B)=\min_{\rho_{AB}=\mathrm{Tr}_{A'B'}(\ket{\psi}\bra{\psi})} S(AA')_{\ket{\psi}}\). In the literature, the phrase “entanglement of purification gap” is not completely standardized: it can mean the practical mismatch between the optimization defining \(E_P\) and the holographic entanglement wedge cross section \(E_W\), the separation between \(E_P\) and \(\frac12 S_R\) or \(\frac12 S_R^{(q)}\), or the explicit residual \(g(A:B)=2E_p(A:B)-I(A:B)\) above the lower bound \(\frac12 I(A:B)\) [1708.09393] [2309.02506] [2510.04596] [2406.09033].

## 1. Definitions and competing meanings of the gap

The common starting point is the standard purification problem. For a mixed state \(\rho_{AB}\), one introduces ancillas and minimizes the entropy across a bipartition of the purified state:
\[
E_P(A:B)= \min_{\rho_{AB} = \mathrm{Tr}_{A'B'}(\ket{\psi} \bra{\psi})} S(AA')_{\ket{\psi}}.
\]
This optimization is natural but difficult, because it ranges over all purifications and all ancilla splittings. The basic lower and upper bounds repeatedly used in the literature are
\[
\frac{1}{2}I(A:B)\le E_P(A:B)\le \min(S_A,S_B),
\]
with \(I(A:B)=S(A)+S(B)-S(AB)\) [1708.09393].

Several distinct “gap” notions appear around this definition.

| Usage | Expression | Role |
|---|---|---|
| Geometric/computational gap | \(E_P(A:B)\) versus \(E_W(A:B)\) | Mismatch between the abstract optimization and a directly computable holographic quantity |
| Reflected-entropy gap | \(E_P(A:B)-\frac12 S_R^{(q)}(A:B)\), or \(\Delta(q)=S(AA')-\frac12 S_R^{(q)}(A:B)\) | Tests whether reflected entropy provides a universal lower bound |
| Mutual-information gap | \(g(A:B)=2E_p(A:B)-I(A:B)\) | Measures how far \(E_P\) lies above its lower bound \(\frac12 I(A:B)\) |

The reflected entropy \(S_R(A:B)\) is defined from a canonical purification in a doubled Hilbert space, while the Rényi-reflected entropy \(S_R^{(q)}(A:B)\) replaces the von Neumann entropy by a Rényi entropy. The entanglement wedge cross section is the holographic quantity
\[
EW(A:B)=\frac{\operatorname{Area}(\Gamma_{A:B})}{4G_N},
\]
where \(\Gamma_{A:B}\) is the minimal surface splitting the entanglement wedge of \(AB\) into parts containing \(A\) and \(B\) [2306.06163].

The absence of a single universally adopted definition is itself a feature of the subject. Some papers explicitly define a gap observable, such as \(g(A:B)\), while others use “gap” more descriptively for the discrepancy between \(E_P\) and a surrogate quantity or for the practical distance between an optimization problem and its geometric realization [2510.04596] [2406.09033].

## 2. Holographic formulation: from purification cost to bulk cross sections

The foundational holographic proposal identifies the entanglement wedge cross section with entanglement of purification. For two disjoint boundary subsystems \(A\) and \(B\), the entanglement wedge cross section is defined as the minimal bulk surface inside the entanglement wedge that separates the \(A\)-side from the \(B\)-side:
\[
E_W(\rho_{AB}) =\min_{\Gamma^{(A)}_{AB}\subset \Gamma^{\min}_{AB}} \left[\frac{A(\Sigma^{\min}_{AB})}{4G_N}\right].
\]
The proposed dictionary is
\[
E_W(\rho_{AB})=E_P(\rho_{AB}),
\]
supported by matched inequalities, including
\[
E_W(\rho_{AB})\leq \min \left[S(\rho_A),S(\rho_B)\right],\qquad
E_W(\rho_{AB})\geq \frac{1}{2}I(A:B),
\]
together with extensiveness and strong superadditivity [1708.09393].

This proposal reframes the gap as a possible discrepancy between a geometric bulk bottleneck and an information-theoretic minimization over purifications. In the holographic picture, the claim is that the gap is absent at leading order in large \(N\): the bulk geometry already implements the optimal purification cost. The original tensor-network argument takes the boundary of the entanglement wedge as an effective purified state, with the choice of cut through that boundary supplying the purification ancillas [1708.09393].

A later generalization introduced overlapping-region versions \(E_W^G\) and \(E_p^G\), and proved parallel inequalities for both. In particular,
\[
I(A:B|C)\le 2E_W^G(AC:BC),\qquad
I(A:B|C)\le 2E_p^G(AC:BC),
\]
along with bounds on tripartite information. The same work also proposed that nonminimal entanglement wedge cuts correspond to suboptimal purifications, so that the difference between a minimal and nonminimal cut may be interpreted as the cost of using a suboptimal purification rather than the optimal one [1710.07643].

The bit-thread formulation sharpened this interpretation. With a divergenceless, norm-bounded flow \(v_{AB}\) constrained to the entanglement wedge,
\[
E_P(A:B)=\max_{v_{AB}:\,\hat n\cdot v_{AB}|_{m_{AB}}=0}\int_A v_{AB},
\]
so \(E_P(A:B)\) becomes the maximal flux from \(A\) to \(B\). In a pure tripartite state \(|\Psi\rangle_{ABC'}\), the total entropy of \(A\) decomposes as
\[
S(A)=E_P(A:B)+\Delta(C'>A),
\]
where the residual term is the quantum advantage of dense coding. In this formulation, one form of the “gap” is the part of \(S(A)\) that cannot be routed specifically to \(B\) inside the entanglement wedge [1904.06871].

## 3. Reflected entropy, lower bounds, and explicit counterexamples

A major strand of the literature studies whether reflected entropy universally lower-bounds entanglement of purification. For integer \(q\ge 2\), one has
\[
E_P(A:B)\ge \frac{1}{2}S_R^{(q)}(A:B),
\]
and, in particular,
\[
E_P(A:B)\ge \frac{1}{2}S_R^{(2)}(A:B).
\]
This lower bound is proved using twist operators and a Cauchy–Schwarz inequality in the replica formalism [2306.06163].

In random tensor networks at large bond dimension, this lower bound can close the gap completely. The key mechanism is that in a large class of networks,
\[
\frac{1}{2}S_R^{(2)}(A:B)=EW(A:B),
\]
while a standard geometric purification argument gives \(E_P(A:B)\le EW(A:B)\). Combining the two inequalities yields
\[
E_P(A:B)=EW(A:B)
\]
up to corrections vanishing as \(D\to\infty\). In these models the gap between the purification optimization and the geometric cross section vanishes at leading order, providing controlled evidence for the holographic conjecture [2306.06163].

The corresponding \(q=1\) statement fails in general. The conjectured inequality
\[
E_P(A:B)\stackrel{?}{\ge} \frac{1}{2}S_R(A:B)
\]
was tested numerically and refuted by explicit counterexamples. The relevant witness is
\[
\Delta(q)\coloneqq S(AA')-\frac{1}{2}S_R^{(q)}(A:B).
\]
At \(q=1\), the conjecture would require \(\Delta(1)\ge 0\), but states with \(\Delta(1)<0\) were found by numerical optimization using PyTorch, the `tensornetwork` package, and the ADAM optimizer [2309.02506].

Two explicit examples were reported. For on-site dimensions \([3,3,2,2]\),
\[
S(AA') \approx 0.81941,\qquad
S_R(A:B) \approx 1.64454,\qquad
S(AA')-\frac12 S_R(A:B)\approx -0.00286.
\]
For on-site dimensions \([4,4,2,2]\),
\[
S(AA') \approx 0.89796,\qquad
S_R(A:B) \approx 1.80783,\qquad
S(AA')-\frac12 S_R(A:B)\approx -0.00596.
\]
Because \(S(AA')\) upper-bounds \(E_P(A:B)\), these immediately imply \(E_P(A:B)<\frac12 S_R(A:B)\) for the reduced states on \(AB\) [2309.02506].

The significance of these results is sharply bifurcated. For generic quantum states, \(\frac12 S_R\) is not a universal lower bound on \(E_P\). For restricted classes—especially CFT states with semiclassical gravity duals, holographic states, and MERA-like tensor network states—the bound may still hold. The violating examples had positive maximal tripartite mutual information, which is inconsistent with the holographic entropy cone, and no violations were found in the tested 8- and 16-qubit MERA states [2309.02506].

## 4. CFT constructions: approximate and alternative purifications

Conformal field theory provides two distinct responses to the purification gap problem. One approach retains the standard purification framework but restricts purifications to states generated from the complement of \(AB\). Using the Reeh–Schlieder theorem, purification states of \(\rho_{AB}\) in vacuum CFT can be approximated by acting with operators localized in \(\overline{AB}\), and the purification constraint together with locality and the separating property implies that the relevant operators are unitary. Combined with surface/state correspondence, this leads to a proof of the holographic formula
\[
E_P(\rho_{AB})=E_W(\rho_{AB})=\frac{\mathrm{Area}(\Sigma_{AB})}{4G}
\]
in the holographic setting [1901.00330].

The same work examined conformal-basis projective measurements on \(\overline{AB}\). Although the projective measurement is not unitary, the post-measurement entropy asymptotically reproduces the holographic entanglement of purification up to a universal constant offset. In the large-\(l\) symmetric limit,
\[
S_A=\frac{c}{3}\log\frac{l}{s}+\cdots,\qquad
E_{AB}=\frac{c}{3}\log\frac{2l}{s},
\]
so
\[
E_{AB}-S_A=\frac{c}{3}\log 2
\]
up to boundary contributions. The same constant appears in the asymmetric limit, again modulo boundary terms. This identifies a small but systematic gap between holographic EoP and a measurement-generated approximation to the optimal purification [1901.00330].

A different CFT construction avoids ancilla addition altogether. Instead of purifying by adjoining \(\bar A,\bar B\), one subtracts the “undetectable regions” from the Euclidean path-integral manifold. For two disjoint subsystems
\[
A=(a_2,b_1),\qquad B=(-\infty,a_1)\cup(b_2,\infty),\qquad a_1<a_2<b_1<b_2,
\]
removing two discs produces a doubly connected region with conformally invariant boundary conditions. Mapping this region to an annulus of width
\[
W=\log\frac{r_2}{r_1},
\]
the replica calculation gives
\[
S_{\mathrm{vN}}(A:B)=\frac{c}{6}W.
\]
For the two-interval geometry this becomes
\[
S_{\mathrm{vN}}(A:B) = \frac{c}{6}\log\left[1+\frac{2}{z}+2\sqrt{\frac{1}{z}\left(\frac{1}{z}+1\right)}\right],
\]
with cross ratio
\[
z=\frac{(a_2-a_1)(b_2-b_1)}{(b_1-a_2)(b_2-a_1)}.
\]
The key statement is
\[
E_W(A:B)=S_{\mathrm{vN}}(A:B),
\]
and similarly for the complementary asymmetric construction \((C,D)\) [2406.09033].

This subtraction-based picture does not redefine entanglement of purification itself. Rather, it narrows the conceptual and computational gap between the optimization defining \(E_P\) and the geometric quantity \(E_W\) by giving a canonical CFT entropy equal to the entanglement wedge cross section. The construction also exhibits a phase transition. The two entropies satisfy
\[
\sinh\left[\frac{3}{c}S_{\mathrm{vN}}(A:B)\right]\sinh\left[\frac{3}{c}S_{\mathrm{vN}}(C:D)\right]=1,
\]
with critical value
\[
S_{\mathrm{vN}}^{*}=\frac{c}{3}\mathrm{arcsinh}\,1=\frac{c}{6}\log\left(3+2\sqrt{2}\right),
\]
matching the known EWCS phase-transition point [2406.09033].

## 5. Numerical and many-body manifestations of purification gaps

Outside holography, the gap often appears as the difference between a canonical or constrained purification and the true minimum. An MPS-based disentangling method was introduced to iteratively minimize the second Rényi entropy of a purified one-dimensional state. The optimization acts only on ancilla degrees of freedom and is designed to approach
\[
E_p[\rho_{LR}] = \min_{|\psi\rangle} S_{LL'}[\psi].
\]
Applied to thermofield double purifications, the method shows that the standard TFD purification is not minimally entangled: in the large-\(\beta\) limit, the optimized entanglement becomes approximately half of that in the standard TFD purification [1711.01288].

This is a concrete computational version of the purification gap. The difference is not between \(E_P\) and a different correlation measure, but between the entanglement carried by a natural purification and the minimal value over all purifications. The same work also applied the method to real-time dynamics after a local operator quench and found that optimized local disentanglers strongly reduce entanglement growth relative to no disentangling, and typically more than backward time evolution on the ancilla [1711.01288].

In many-body lattice models, the optimized purification itself can behave nontrivially. Numerical studies of a free scalar field on a lattice and the transverse-field Ising chain found that \(E_P\) can be non-monotonic in the separation \(d\) between \(A\) and \(B\) when the total system size is small, while for larger systems it becomes monotonic and shows a plateau-like behavior. In the scalar theory, at \(N=60\), \(w=|A|=|B|=4\), and \(m=10^{-4}\), the EoP shows a plateau from \(d=0\) to \(d=1\). In the critical Ising chain, for small \(N\), \(E_P\) can even increase with distance [1902.02369].

The same study also found that the optimal purification can spontaneously break the \(A\leftrightarrow B\) reflection symmetry of the mixed state. In the scalar case, the asymmetry parameter
\[
\mathcal A = \|K_{A,\tilde B}-K_{B,\tilde A}^{R}\|_2
\]
is essentially zero for \(d\neq 1\) but becomes nonzero at \(d=1\), with the effect growing with subsystem size \(w\). In the Ising model, the same \(Z_2\) symmetry breaking occurs only in the ferromagnetic phase \(h<1\) [1902.02369].

That paper does not define a universal “entanglement of purification gap” as a dedicated observable, but it does exhibit quantified separations that function as gap-like quantities. In Werner states it explicitly notes a regime with
\[
S_A-E_P>0,
\]
and in the scalar conformal scaling regime it finds that \(E_P\) and mutual information have similar scaling but differ by an approximately constant offset \(C\approx 0.25\) over the studied range. The authors interpret these differences in terms of the interplay between short-range quantum entanglement and longer-range classical correlations [1902.02369].

## 6. Explicit gap observables and multipartite generalization

A more formal notion of entanglement-of-purification gap appears when the lower bound by half the mutual information is taken as the reference. In this setting,
\[
g(A:B)\equiv 2E_p(A:B)-I(A:B),
\]
which is nonnegative and vanishes exactly when \(E_P\) saturates its lower bound. For a tripartite pure state \(ABC\), where \(C\) purifies \(\rho_{AB}\), the recalled result is
\[
\ket{\psi_{ABC}}\text{ is 2-producible}\quad \Longleftrightarrow \quad g(A:B)=0.
\]
The same quantity also admits the conditional-mutual-information expressions
\[
g(A:B)=I(\bar A:B\bar B|A)+I(\bar B:A|B)
      =I(\bar B:A\bar A|B)+I(\bar A:B|A),
\]
which give it a recovery-theoretic interpretation [2510.04596].

This bipartite gap ceases to be sufficient beyond three parties. Using 4-partite random stabilizer states with size ratios
\[
N_A:N_B:N_C:N_D=1:3:3:3,
\]
it was shown that vanishing pairwise gaps \(g(A_i:A_j)=0\) for all pairs does not force 2-producibility. In the quoted example, the pairwise EoP gaps and pairwise negativities become asymptotically negligible, yet subsystem \(A\) is still entangled with \(BCD\) with probability tending to 1 [2510.04596].

To address this, a generalized multipartite entanglement of purification was introduced:
\[
E_p(A_1,\dots,A_n) =\frac{1}{2}\min \sum_{i=1}^n S(A_iR_i)
=\frac{1}{2}\min D\!\left(\rho_{\bigotimes_{i=1}^{n+1}A_i}\,\Big\|\,\bigotimes_{i=1}^n \rho_{A_iR_i}\right),
\]
where the minimization is over all purifications and partitions of the purifier into \(R_1,\dots,R_n\). The corresponding generalized gap is
\[
g(A_1,\dots,A_n)
=\frac{1}{2}\min_{(R_1:\cdots:R_n)} \left[\sum_{i=1}^n \big(S(A_iR_i)-S(A_i)\big)+S(\bar A_{n+1})\right],
\]
or equivalently
\[
g(A_1,\dots,A_n)
=\frac{1}{2}\min_{(R_1:\cdots:R_n)}\sum_{i=1}^n I\!\left(R_i:\bar A_i\bar R_i\,\middle|\,A_i\right).
\]
Its operational meaning is the optimal total quantum communication cost of sequentially redistributing purifier fragments to the parties [2510.04596].

The central structural theorem is that a multipartite pure state is 2-producible if and only if the generalized EoP gaps vanish on all relevant subsystem sets. The same work further proved lower bounds linking the generalized gap to local recovery and distance from the set of 2-producible states:
\[
g(A_1,\dots,A_n)\ge -2\log F\!\left(\rho_A,\mathcal R^{\mathrm{LO}}_{\bar A_{n+1}\to A}(\rho_{\bar A_{n+1}})\right),
\]
and
\[
\max_{|\alpha|<n} g(\alpha)\le \min_{\sigma\sim_{\mathrm{LU}}\rho}\min_{\mu\in \mathrm{BPS}_n} D(\sigma\Vert\mu).
\]
For states with a generalized Schmidt decomposition,
\[
\ket{\psi}=\sum_l \sqrt{p_l}\bigotimes_{i=1}^{n+1}\ket{\psi_{A_i}^l},
\]
the formulas simplify to
\[
E_p(A_1,\dots,A_n)=\frac{n}{2}H(p_l),\qquad
g(A_1,\dots,A_n)=\frac{1}{2}H(p_l).
\]
In this line of work, the entanglement of purification gap is no longer merely a mismatch between proxy quantities; it becomes an explicit criterion for irreducible multipartite entanglement [2510.04596].

Source: https://www.emergentmind.com/topics/entanglement-of-purification-gap