Entanglement-Wedge Reconstruction

Entanglement-wedge reconstruction follows when the boundary region preserves all relative-entropy distinguishability carried by the corresponding bulk wedge algebra. In a finite-dimensional regulated code subspace, equality in data processing implies a recovery channel and hence boundary representatives of bulk operators. Gravity supplies the relative-entropy equality perturbatively through generalized entropy; it also supplies the approximation, center, and code-domain qualifications.

Required background. Quantum Information supplies relative entropy, data processing, Petz recovery, and operator-algebra codes; Gauge Theory and String-Theory Preparation supplies holographic encoding and gauge constraints.

Helpful background. Semiclassical Gravity supplies QES wedges and the area term; Quantum Fields in Curved Spacetime supplies regulated bulk algebras; Black-Hole Information supplies islands and code subspaces.

Boundary and wedge algebras

Fix a boundary region $A$, its complement $\bar A$, and a semiclassical code subspace $\mathcal H_{\mathrm{code}}$. Let $a$ be the generalized-entropy wedge bounded by $A$ and the selected QES. After finite-dimensional regulation, denote the wedge operator algebra by $\mathcal M_a$ and the boundary algebra by $\mathcal B(\mathcal H_A)$.

For code states $\rho$ and $\sigma$, write their reduced states as $\rho_A,\sigma_A$ and their restrictions to $\mathcal M_a$ as $\rho_a,\sigma_a$. Quantum relative entropy is

$$D(\rho\Vert\sigma) =\operatorname{Tr}\rho(\log\rho-\log\sigma),$$

with the usual support condition. The relevant equality is

$$D(\rho_A\Vert\sigma_A) =D(\rho_a\Vert\sigma_a)$$

for every pair in the declared state family, exactly or with a stated error. The area or charge center must be treated consistently on both sides.

JLMS relative entropy

Fix a faithful reference state $\sigma$ in a perturbative holographic code. Pulling the boundary modular Hamiltonian back by the encoding $V$, its relation to the bulk operator has the schematic form

$$V^\dagger\bigl(K_{\sigma_A}^{\mathrm{CFT}}\otimes I_{\bar A}\bigr)V = \mathcal L_A +K_{\sigma_a}^{\mathrm{bulk}} +O(1/N),$$

where $K_\sigma=-\log\sigma$ and $\mathcal L_A$ is the central gravitational entropy operator, with leading term $\widehat{\operatorname{Area}}(X_A)/(4G)$ and the local terms appropriate to the approximation. Every operator in this equation acts on the same code Hilbert space. Combining it with $S(\rho_A)=\langle\mathcal L_A\rangle_\rho+S(\rho_a)$ cancels the gravitational term in

$$D(\rho_A\Vert\sigma_A) =\operatorname{Tr}(\rho_AK_{\sigma_A})-S(\rho_A),$$

leaving the bulk relative entropy to the retained order. The derivation and its state-family assumptions are in Jafferis, Lewkowycz, Maldacena, and Suh 2016, §§2–4. A center’s classical sector probabilities remain part of the algebraic relative entropy; canceling the area term does not discard them.

This is stronger than equality of one entropy. Relative entropy compares every pair of code states and therefore tests distinguishability. It also obeys data processing: a channel cannot increase it. Equality means that no distinguishability relevant to the wedge algebra was lost when restricting to $A$.

The QES and wedge can change between states. A common reconstruction theorem requires a state family for which the algebra and wedge assignment are controlled, or a block/center formulation that records the change. Quietly swapping wedges between the two sides of the equality makes the statement ill posed.

From equality to reconstruction

Theorem. Assume a finite-dimensional regulator, a fixed code subspace and encoding channel, fixed algebras for $a$ and $A$, and relative-entropy equality for all relevant code states. Then the encoding channel to $A$ is sufficient for $\mathcal M_a$: there exists a recovery channel whose Heisenberg adjoint assigns every $O_a\in\mathcal M_a$ a boundary operator $O_A$ satisfying

$$O_A V=VO_a$$

on the code. With the corresponding complement equality, complementary recovery reconstructs the commutant in $\bar A$. The reconstruction theorem and proof are given in Dong, Harlow, and Wall 2016, complete Letter and proof.

The proof uses equality in monotonicity. In the factor-like regulated case, let $\mathcal N$ denote the induced channel that sends the logical wedge state to its encoded state on $A$, with complementary logical data fixed or included as a center. Choose a faithful reference state $\sigma$ on the logical input algebra. Saturation of data processing makes the Petz map

$$\mathcal R_{\sigma}(X) = \sigma^{1/2} \mathcal N^\dagger\!\left( \mathcal N(\sigma)^{-1/2}X \mathcal N(\sigma)^{-1/2} \right) \sigma^{1/2}$$

recover every code state in the family. Passing to the adjoint gives the boundary operator representation. For operator algebras with centers, the construction is performed blockwise with the central data retained.

For a general operator-algebra code, $\rho_a$ need not by itself determine $\rho_A$ unless the complementary and central data have been handled this way. The full Dong–Harlow–Wall argument supplies that algebraic structure; writing a channel directly from an arbitrary reduced wedge state to an arbitrary boundary reduction without it would assume part of the result.

The displayed Petz formula assumes inverses on the relevant supports. It is trace preserving on the support of $\mathcal N(\sigma)$ and can be extended to a channel on the full output space by assigning the orthogonal subspace a fixed input state. Zero modes therefore require support restriction or a generalized inverse; ignoring them can turn a well-defined recovery statement into a singular expression.

Approximate equality and reconstruction error

Semiclassical JLMS equality has $1/N$ or $G$ corrections. When the regulated setup supplies the induced channel above and $\mathcal N(\rho_a)=\rho_A$, its data-processing loss

$$0\le D(\rho_a\Vert\sigma_a) -D(\mathcal N(\rho_a)\Vert\mathcal N(\sigma_a)) \le\delta$$

uniformly on the state family, quantitative recovery theorems can bound fidelity or channel error. If one only knows a perturbative JLMS mismatch $\lvert D(\rho_a\Vert\sigma_a)-D(\rho_A\Vert\sigma_A)\rvert\le\delta$ without such a channel construction, no sign should be assigned to that raw mismatch. The conversion depends on the norm, reference state, dimensions, and whether the bound is uniform. Universal recovery maps are developed in Cotler et al. 2019, Theorem 1 and §§2–3.

An expectation-value match for a few simple operators is weaker than a small diamond-norm reconstruction error. Similarly, $O(1/N)$ for each fixed correlator need not remain small for an exponentially large code or a circuit whose complexity grows with $N$.

Checks and counterexamples

Identical-state check. Setting $\rho=\sigma$ gives zero relative entropy on both sides. This checks normalization but not reconstruction; all state pairs are needed.

Monotonicity check. Once an induced channel from the logical wedge data to $A$ has been defined, it cannot increase relative entropy. A sign opposite to data processing then signals mismatched algebras, states, or channel direction. Without that induced channel, the raw comparison is a JLMS relation rather than a free-standing monotonicity inequality.

Counterexample from an enlarged code. If two orthogonal logical states have identical reductions on $A$ but different restrictions to $\mathcal M_a$, their boundary relative entropy is zero while their bulk distinguishability is nonzero. Equality and reconstruction both fail. This occurs when a code is enlarged beyond the erasure threshold.

The theorem establishes information-theoretic existence, not efficient decoding. Gravitational dressing, generic AdS/CFT foundations, and nonperturbative completion remain outside its proof.

Exercises

Encode a qubit as $\lvert0\rangle\mapsto\lvert00\rangle$, $\lvert1\rangle\mapsto\lvert11\rangle$, and trace out the second qubit. Compare the logical states $\lvert+\rangle$ and $\lvert-\rangle$. What happens to relative entropy and logical $X$ reconstruction?

Solution

The encoded states are $(\lvert00\rangle\pm\lvert11\rangle)/\sqrt2$, which are orthogonal, so their input relative entropy is infinite in the ideal pure-state support convention. After tracing the second qubit, both give $I_2/2$, so the output relative entropy is zero. The channel has erased the phase distinguishing $+$ from $-$; equality in data processing fails maximally, and no operator on the first qubit can represent logical $X$ on the full code. Full-rank regularizations make the entropies finite but preserve the strict loss.

References

• Cotler et al. 2019, Jordan Cotler, Patrick Hayden, Geoffrey Penington, Grant Salton, Brian Swingle, and Michael Walter, “Entanglement Wedge Reconstruction via Universal Recovery Channels,” Physical Review X 9 (2019) 031011.
• Dong, Harlow, and Wall 2016, Xi Dong, Daniel Harlow, and Aron C. Wall, “Reconstruction of Bulk Operators within the Entanglement Wedge in Gauge-Gravity Duality,” Physical Review Letters 117 (2016) 021601.
• Jafferis et al. 2016, Daniel L. Jafferis, Aitor Lewkowycz, Juan Maldacena, and S. Josephine Suh, “Relative Entropy Equals Bulk Relative Entropy,” Journal of High Energy Physics 06 (2016) 004.