Entanglement Membrane Picture

• The Entanglement Membrane Picture is a variational framework representing entanglement dynamics in chaotic many-body systems via codimension-one membranes that encode local entropy production.
• It organizes phenomena such as entanglement growth, operator spreading, and butterfly physics by reducing the problem to a minimization over spacetime surfaces with an orientation-dependent tension function E(v).
• Despite its success in random circuits, Floquet models, large-N systems, and holography, the approach requires modifications for free fermions, 2D CFTs, and impurity problems, highlighting its model dependence.

The entanglement membrane picture is a coarse-grained spacetime description of entanglement dynamics in chaotic many-body quantum systems in which the entropy of a region is obtained from a variational problem over a codimension-one membrane, or in $1+1$ dimensions a directed line, with a local tension $E(v)$ or $\mathcal E(v)$ set by its orientation $v$ in spacetime. It was first formulated as a leading-order description of state and operator entanglement in short-range chaotic systems, then extended from random-circuit settings to deterministic Floquet models, Brownian large-$N$ models, and holography. Across these settings, the membrane organizes entanglement growth, operator spreading, and butterfly physics, while more recent work has clarified that its standard local form is neither exact nor universal: free-fermion systems, pure $2$d CFT, and finite-interval impurity problems require either modified spacetime objects or additional degrees of freedom, and in some cases expose qualitative failures of the usual membrane framework (Jonay et al., 2018, Zhou et al., 2019, Mezei et al., 3 Dec 2025, Fraenkel et al., 2024, Jiang et al., 2024).

1. Variational formulation and basic quantities

In one spatial dimension, the membrane picture can be written either as a local entropy-production law or as a spacetime minimization principle. For a pure state with entanglement entropy $S(x,t)$ across a cut at position $x$, the local growth law is

$$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$

while the equivalent variational form is

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$

Here $E(v)$0 is the equilibrium entropy density, $E(v)$1 is the line tension, and the minimizing curve runs from $E(v)$2 to some $E(v)$3. In higher dimensions, the same structure becomes a minimization over codimension-one membranes in the spacetime slab, with a local surface-tension density determined by the membrane orientation (Jonay et al., 2018).

The local and variational formulations are related by Legendre transform. In $E(v)$4 dimensions,

$E(v)$5

This identifies the membrane not as an independent ansatz, but as a geometric form of coarse-grained entanglement hydrodynamics. The entanglement velocity is $E(v)$6, while the butterfly velocity is singled out by

$E(v)$7

Convexity, $E(v)$8, is a central consistency condition in the coarse-grained theory (Jonay et al., 2018).

The same formalism appears in later large-$E(v)$9 and holographic work with notation $\mathcal E(v)$0 for the $\mathcal E(v)$1-th Rényi membrane tension. In that language, a quench formula takes the form

$\mathcal E(v)$2

and for higher-dimensional holographic systems one writes

$\mathcal E(v)$3

These formulations differ in derivation and regime, but they share the same central object: a local tension function governing entanglement growth at large scales (Mezei et al., 3 Dec 2025, Jiang et al., 2024).

2. Replica structure, permutation variables, and spacetime statistical mechanics

The membrane picture emerged most cleanly in random unitary circuits, where replicated tensor networks reduce to effective statistical mechanics of local pairing or permutation variables. Rényi entropies and OTOCs require multiple forward and backward copies,

$\mathcal E(v)$4

and after Haar averaging the effective degrees of freedom are local permutations $\mathcal E(v)$5. In this setting the membrane is a domain wall in spacetime between different pairing sectors, and the entanglement problem becomes a partition function for those domain walls (Zhou et al., 2019).

A decisive extension beyond averaged random circuits is the exact projector decomposition introduced for deterministic chaotic systems. For a two-site gate, the replicated identity is resolved as

$\mathcal E(v)$6

with $\mathcal E(v)$7 projecting onto pairing states and $\mathcal E(v)$8 onto the orthogonal complement. The effective local spin variable is therefore

$\mathcal E(v)$9

The paired projectors obey

$v$0

so the microscopic dynamics survives only through the $v$1 sector. In the resulting spacetime partition sum, the membrane is no longer a strictly thin random-circuit wall: it becomes a “thick” wall dressed by finite $v$2-clusters. The central coarse-graining assumption is that this thickness remains $v$3, allowing an emergent line tension $v$4 to be defined from fixed-endpoint partition sums (Zhou et al., 2019).

An independent microscopic derivation appears in Brownian chaotic models. There, disorder averaging maps the Lorentzian $v$5-th Rényi problem to Euclidean evolution under a replicated superhamiltonian $v$6, and the low-energy sector is argued to consist of a universal gapped band of plane-wave, locally dressed domain walls between permutation-labeled ferromagnetic ground states. Their dispersion relation determines the membrane tension through large-deviation/saddle-point relations. In this approach, the membrane is not just an effective geometric surface; it is the spacetime large-deviation avatar of specific low-lying replicated excitations (Vardhan et al., 2024).

3. Line tension, operator entanglement, and butterfly physics

A major strength of the membrane framework is that the same tension function governs both state entanglement and operator spreading. For the time-evolution operator $v$7, viewed as a doubled state, the operator entanglement satisfies

$v$8

while for larger separations

$v$9

This makes $N$0 directly accessible through operator entanglement of the unitary and clarifies the distinguished role of $N$1: beyond the butterfly cone, the leading dynamical contribution to $N$2 disappears (Jonay et al., 2018).

The same line tension controls the large-deviation form of OTOCs. In the deterministic Floquet formulation,

$N$3

for $N$4. Thus $N$5 simultaneously encodes entanglement growth and operator-front decay outside the light cone (Zhou et al., 2019).

Concrete solvable examples show how model dependence enters the tension. For a large-$N$6 random circuit with random spacetime structure,

$N$7

For a regular random circuit at finite $N$8, the second-Rényi tension is

$N$9

with

$2$0

These examples illustrate a recurring theme of the literature: the membrane structure is robust, but the detailed shape of $2$1 is model-dependent (Jonay et al., 2018).

Higher Rényi indices can introduce additional structure. In Brownian symmetry-free models, the second Rényi entropy is controlled by a dressed single-domain-wall band, but for the third Rényi entropy the proposed tension $2$2 can require phase transitions between bound-state and unbound-domain-wall branches in order to satisfy the consistency condition that an equilibrium state should not continue to gain entropy. This makes the membrane tension itself piecewise and indicates that higher-Rényi membrane theories may contain nontrivial internal phase structure (Vardhan et al., 2024).

4. Deterministic Floquet systems and solvable chaotic large-$2$3 realizations

In deterministic Floquet systems, the membrane picture survives coarse-graining despite the absence of circuit averaging. The fixed-endpoint partition function is written in terms of irreducible membrane steps $2$4, leading to the exact renewal equation

$2$5

Generating functions then yield the line tension $2$6. For generic chaotic Floquet models the resulting $2$7 is convex and the membrane exhibits diffusive wandering, whereas dual-unitary circuits form a special wavelike universality class with flat line tension

$2$8

so $2$9 and $S(x,t)$0. In that case the continuum membrane dynamics is governed by a wave equation rather than diffusion (Zhou et al., 2019).

The Brownian SYK chain provides a solvable interacting large-$S(x,t)$1 derivation in which the membrane is a traveling-wave saddle of collective bilinear fields on the replicated Schwinger–Keldysh contour. For $S(x,t)$2, the relevant collective variables are $S(x,t)$3, $S(x,t)$4, and $S(x,t)$5, constrained by

$S(x,t)$6

and the continuum action takes the canonical form

$S(x,t)$7

In the large-$S(x,t)$8, large-subsystem limit, the membrane is a localized traveling wave,

$S(x,t)$9

interpolating between asymptotic contour-gluing patterns. The membrane free energy is the on-shell action density of this soliton (Mezei et al., 3 Dec 2025).

This derivation makes the relation to scrambling explicit. For $x$0, the membrane is a bound state of two fronts and has finite width. For $x$1, a single localized membrane ceases to exist and instead splits into two fronts, each moving at $x$2, separated by a macroscopic region of width

$x$3

In that regime the localized membrane description breaks down, while operator entanglement still defines an effective slope

$x$4

For $x$5, the computed $x$6 and $x$7 are convex and satisfy

$x$8

This is one of the clearest microscopic realizations of the membrane as an actual saddle-point object rather than a postulated variational surface (Mezei et al., 3 Dec 2025).

5. Modified membrane phenomenology: free fermions and $x$9d CFT

Not all systems with spacetime entanglement descriptions realize the standard sharp membrane with nonzero local tension. In noisy free fermion chains without conservation laws, disorder-averaged Rényi entropies map not to permutation spins but to an $$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$0 ferromagnetic Heisenberg chain in imaginary time,

$$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$1

For $$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$2 this becomes a spin-$$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$3 Heisenberg chain, and the relevant spacetime object is a smooth coherent-state saddle described by two independent fields $$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$4 and $$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$5, obeying nonlinear diffusion-type equations. The domain wall broadens diffusively rather than remaining localized, and the purity cost scales as

$$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$6

with

$$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$7

This produces $$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$8 entanglement growth, not ballistic $$\frac{\partial S}{\partial t}=s_{\mathrm{eq}}\,\Gamma\!\left(\frac{\partial S}{\partial x}\right),$$9-growth. Weak interactions explicitly break the continuous replica symmetry, pin the wall to finite width

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$0

and induce a crossover time

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$1

after which

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$2

The membrane idea therefore survives only in a generalized sense in the Gaussian universality class: the controlling spacetime object is a smooth diffusive wall, and the sharp tensionful membrane emerges only after interactions discretize the vacuum manifold (Swann et al., 2023).

A different modification occurs in pure $$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$3d CFT. Holographically, BTZ gives

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$4

so the ordinary membrane tension is completely degenerate. This suffices for several leading entanglement-entropy results, but it fails for reflected entropy. The proposed remedy is a generalized membrane carrying an additional worldline degree of freedom

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$5

with effective Lagrangian

$$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$6

The extra $$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$7-profile captures radial information that a degenerate $$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$8 discards and is necessary to reproduce the $$S(x,t)=\min_y\left[t\,s_{\mathrm{eq}}\,E\!\left(\frac{x-y}{t}\right)+S(y,0)\right].$$9d CFT reflected-entropy plateau $E(v)$00 rather than $E(v)$01 (Jiang et al., 2024).

Relevant deformations of the $E(v)$02d CFT restore an ordinary nondegenerate membrane. In the scalar-hairy planar BTZ setup, the deformation generates a finite $E(v)$03 and yields

$E(v)$04

with a conventional convex tension function obeying the usual butterfly-point conditions. In this sense, the extra $E(v)$05 degree of freedom is effectively gapped out under RG flow, and the generalized $E(v)$06d membrane interpolates back to the standard higher-dimensional form (Jiang et al., 2024).

6. Defects, impurities, and limitations of the standard local membrane

A recent sharp test of the membrane paradigm is provided by dual-unitary quantum circuits with a fixed non-dual-unitary impurity gate. The circuit is a one-dimensional brickwork model on $E(v)$07 sites with possibly different local dimensions on the two sides of the impurity,

$E(v)$08

All bulk gates are dual unitary, while the impurity at bond $E(v)$09 is only unitary and defines a vertical defect line in spacetime. The entanglement region is a contiguous interval

$E(v)$10

of length $E(v)$11, lying a distance $E(v)$12 from the impurity (Fraenkel et al., 2024).

For a semi-infinite subsystem, the membrane picture works exactly. The Rényi entropy is

$E(v)$13

where $E(v)$14 is the operator entanglement of the impurity gate. In the membrane language, the bulk has constant line tension $E(v)$15, while the defect line has lower tension $E(v)$16. The optimal membrane runs along the light cone to the impurity and then vertically down the impurity worldline, giving

$E(v)$17

In this geometry the membrane, quasiparticle, and exact calculations coincide (Fraenkel et al., 2024).

For a finite interval near the impurity, however, the exact result can be qualitatively non-monotonic. In the integrable bulk realization,

$E(v)$18

Because typically $E(v)$19, the slope is negative in the window $E(v)$20 and positive again for $E(v)$21. The quasiparticle interpretation is that reflected entangled pairs can temporarily place both members inside the interval, reducing the bipartite entropy, before later restoring maximal contribution when one partner exits (Fraenkel et al., 2024).

The standard membrane picture, by contrast, predicts monotone growth until saturation for a connected interval if one assumes a local positive line tension and the usual extremization logic. The discrepancy is not confined to integrable circuits. In random dual-unitary circuits with entangling power $E(v)$22, lower bounds on the second-Rényi purity show that non-monotonicity can persist in a chaotic regime. Thus the impurity problem reveals a structural limitation of the standard local membrane theory: reflected-pair memory and coupling between the two interval boundaries are not naturally encoded by a local positive-tension surface functional (Fraenkel et al., 2024).

7. Applications to Page curves, information recovery, and present scope

The membrane picture has been used as a coarse-grained theory of black-hole-information dynamics in many-body toy models. In an evaporating black-hole model built from a chaotic system of length $E(v)$23 with moving boundary $E(v)$24, the entropy of the black-hole or radiation region is determined by competition between two candidate membranes. Their costs are

$E(v)$25

with

$E(v)$26

The Page time is the crossing point,

$E(v)$27

so the Page curve arises from a transition between competing membrane saddles. In the same framework, Hayden–Preskill recovery appears as a rapid transfer of mutual information around Page time, again driven by changes in the dominant membrane configuration rather than by microscopic gate-by-gate analysis (Blake et al., 2023).

In these black-hole toy models the membrane is not claimed to replace gravitational replica path integrals or quantum extremal surfaces. Rather, it supplies a many-body variational mechanism whose saddle transitions mirror Page-curve and island-like phenomena. The same paper emphasizes that many results depend only on coarse quantities such as $E(v)$28, $E(v)$29, $E(v)$30, and total black-hole entropy $E(v)$31, while intermediate-time behavior remains model-dependent through the detailed form of $E(v)$32 (Blake et al., 2023).

Taken together, the literature supports a precise but limited statement. In generic chaotic systems without additional obstructions, the membrane gives a powerful leading-order description of entanglement growth, operator entanglement, and butterfly-cone structure. Deterministic Floquet systems, Brownian replicated Hamiltonians, and solvable large-$E(v)$33 models all provide concrete microscopic realizations of this statement. At the same time, the framework is incomplete in several important regimes: free noisy fermions require a diffusing smooth wall rather than a sharp membrane; pure $E(v)$34d CFT requires an extra degree of freedom $E(v)$35 for reflected entropy; and impurity problems can produce finite-time, finite-interval behavior that a local membrane with positive tension misses qualitatively. A plausible implication is that the long-time, large-scale membrane theory is best regarded not as a universal exact law, but as one effective description inside a broader landscape of spacetime entanglement theories that interpolate among membrane, quasiparticle, and more elaborate replicated saddle structures (Swann et al., 2023, Jiang et al., 2024, Fraenkel et al., 2024).