gpt 6 astra ultra attempt on Phys-3

Entanglement Area Law for 2D Gapped Quantum Systems

Definitions and mathematical statement

Let $H_\Lambda$ be a finite-range bounded Hamiltonian on a finite square-lattice region, with fixed finite site dimension, unique normalized ground state $|\psi_\Lambda\rangle$, and spectral gap at least $\Delta>0$ uniformly in $\Lambda$. For $A\subseteq\Lambda$, put \[ \rho_A=\operatorname{Tr}_{\Lambda\setminus A}|\psi_\Lambda\rangle\langle\psi_\Lambda|, \qquad S(\rho_A)=-\operatorname{Tr}(\rho_A\log\rho_A). \] The conjecture is $S(\rho_A)\le C|\partial A|$, where $|\partial A|$ counts lattice edges crossing to the complement. The constant may depend on the fixed interaction parameters and gap but not on the region or system size.

Formulation and concept references: [S1].

Short English statement

Must entanglement in every uniformly gapped two-dimensional local ground state grow at most in proportion to a region's boundary?

Research attempt

Prepared by GPT-6 Astra Ultra on 27 September 2026, with primary-source review, explicit example checks, and examination of the stated conditional reductions. This is an unresolved research attempt.

Scope and checked literature (27 September 2026).

The target concerns a unique ground state, a fixed local site dimension, bounded finite-range interactions, and a gap uniform in system size. No frustration-free assumption, local gap for every sub-Hamiltonian, or preparation by a bounded-depth circuit is present in the target. Arad--Firanko--Jain [S1] assumes a sufficiently good approximate ground-state projector in its general theorem. The known two-dimensional applications require additional hypotheses. The primary construction of Anshu--Arad--Gosset [E1] concerns locally gapped frustration-free systems and specified cuts. A lower bound for topological entropy [S2] does not establish the upper bound requested here.

The attempt derives a precise entropy criterion from Schmidt tails, proves the area law for a restricted preparation class, and tests what a spectral gap alone gives through a simple projector approximation. The obstruction is quantitative: a small error in a low-rank approximation can still carry entropy proportional to the volume. All logarithms and entropies below use the natural logarithm.

A complete restricted case: finite-depth local preparation.

Suppose the ground state is obtained from a product state by $D$ layers of disjoint nearest-neighbor two-site gates, with $D$ independent of the system size. Let $b=|\partial A|$ and let the site dimension be $d$. A gate wholly in $A$ or its complement does not change $S(\rho_A)$. A gate crossing the cut changes that entropy by at most $2\log d$.

For the last claim, write $A=A_0\cup\{a\}$, where $a$ is the site of the gate in $A$. The reduced state on $A_0$ is unchanged by a unitary on the two gate sites. Before and after the gate, subadditivity and the Araki--Lieb inequality give \[ S(A_0)-\log d\le S(A_0a)\le S(A_0)+\log d. \] Their difference is at most $2\log d$. There are at most $b$ crossing gates per layer, so starting with entropy zero proves \begin{equation}\label{wsref-0-61-72-65-61-3a-63-69-72-63-75-69-74} S(\rho_A)\le 2Db\log d. \end{equation} This applies to every subset $A$ for such a preparation. The proof does not show that a general uniformly gapped ground state has one.

The area scaling is attained in a simple gapped family. Choose disjoint nearest-neighbor edges and put a maximally entangled pair of dimension $d$ on each, with any uncovered sites in fixed pure states. The sum of the projectors onto the orthogonal complements of these desired local states has a unique ground state and gap one. A cut crossing $m$ dimers has exactly $m\log d$ entropy. The interaction range, norm, and site dimension are uniform, and $m\le b$. Thus the boundary dependence cannot in general be replaced by a system-size-independent constant.

What one accurate low-rank approximation actually proves.

Let $\lambda_1\ge\lambda_2\ge\cdots$ be the eigenvalues of $\rho_A$, including zeros up to Schmidt dimension $D_A\le d^{\min(|A|,|\Lambda\setminus A|)}$. Put $\delta_R=\sum_{j>R}\lambda_j$. The largest squared overlap of $|\psi\rangle$ with a normalized Schmidt-rank-at-most-$R$ state is \[ 1-\delta_R=\sum_{j\le R}\lambda_j. \] One proof uses the singular-value decomposition of the coefficient matrix of $\psi$: orthogonal projection onto any rank-$R$ Schmidt subspace captures at most the sum of its $R$ largest squared singular values, and truncating the Schmidt decomposition attains the bound.

Grouping the first $R$ eigenvalues and the tail gives the exact entropy decomposition into a binary entropy and two conditional entropies. Each conditional distribution is maximized when uniform, so, for $1\le R<D_A$, \begin{equation}\label{wsref-0-61-72-65-61-3a-6f-6e-65-74-61-69-6c} S(\rho_A)\le h_2(\delta_R) +(1-\delta_R)\log R+\delta_R\log(D_A-R). \end{equation} Here $h_2(t)=-t\log t-(1-t)\log(1-t)$, with the usual endpoint convention. This uses the actual tail mass; it does not replace it inside every term by an upper bound without checking monotonicity. In particular, a fixed small tail bound and $\log R=O(b)$ leave a possible term of order $\delta_R|A|$.

The obstruction occurs in the explicit normalized state \[ |\Psi\rangle=\sqrt{1-\delta}|0,0\rangle +\sqrt{\frac{\delta}{D-1}} \sum_{j=1}^{D-1}|j,j\rangle. \] It has squared overlap $1-\delta$ with a product state and entropy $h_2(\delta)+\delta\log(D-1)$. Encode each side in $m$ sites, so $D=d^m$. Taking $\delta=m^{-1/4}$ makes the fidelity tend to one while the entropy grows as $m^{3/4}\log d$. For square regions with boundary of order $\sqrt m$, this exceeds an area bound. This family is a test of the proposed inference from fidelity; it is not asserted to be the unique gapped ground state of a local Hamiltonian.

A sufficient criterion using all approximation scales.

The following condition does control the entire entropy. Suppose there are constants $a,b_0,c,C>0$, independent of the cut and system size, such that, for every integer $j\ge0$, \begin{equation}\label{wsref-0-61-72-65-61-3a-74-61-69-6c-73} \sum_{r>R_j}\lambda_r\le Ce^{-cj},\qquad R_j=\left\lceil\exp(a|\partial A|+b_0j)\right\rceil. \end{equation} Ranks beyond $D_A$ are harmless. For a random index $I$ with law $\Pr(I=r)=\lambda_r$, let $J$ be the first bucket containing it: the zeroth bucket is $r\le R_0$, and bucket $j\ge1$ is $R_{j-1}<r\le R_j$. Then \[ \Pr(J>j)\le Ce^{-cj},\qquad \mathbb EJ\le\frac{C}{1-e^{-c}}. \] The entropy chain rule and the size of a bucket give \[ S(\rho_A)=H(I)\le H(J)+\mathbb E\log R_J \le H(J)+a|\partial A|+b_0\mathbb EJ+\log2. \] For any $\theta>0$, compare the law of $J$ with the geometric distribution $q_j=(1-e^{-\theta})e^{-\theta j}$. Nonnegativity of relative entropy gives \[ H(J)\le-\log(1-e^{-\theta})+\theta\mathbb EJ. \] Consequently \begin{equation}\label{wsref-0-61-72-65-61-3a-65-6e-74-72-6f-70-79} S(\rho_A)\le a|\partial A|+\log2-\log(1-e^{-\theta}) +\frac{(b_0+\theta)C}{1-e^{-c}}. \end{equation} For cuts with $|\partial A|\geq1$ this is an area law after enlarging the constant. The empty and full-region cuts have entropy zero. All constants in (area:tails) must be uniform; allowing $C$ to grow exponentially with the volume would defeat the argument.

This proof exhibits what additional approximation information is needed beyond one high fidelity: the rank cost grows exponentially in the accuracy index, while the tail mass decays exponentially in that same index. The entropy contribution from arbitrarily small eigenvalues is then summable with a bound independent of volume.

A projector reduction and the cost hidden in the starting overlap.

Suppose a Hermitian operator $K$ satisfies $K\psi=\psi$, has norm at most $\eta<1$ on $\psi^\perp$, and has operator Schmidt rank at most $r$ across the chosen cut. For a product vector $\phi$ with $\alpha=\langle\psi,\phi\rangle\ne0$, write $\phi=\alpha\psi+\phi_\perp$. Then \[ K^m\phi=\alpha\psi+K^m\phi_\perp, \qquad\|K^m\phi_\perp\|\le\eta^m. \] The vector on the left has Schmidt rank at most $r^m$, since applying each sum of $r$ product operators multiplies that rank by at most $r$. After normalization its squared overlap with $\psi$ is at least $|\alpha|^2/(|\alpha|^2+\eta^{2m})$. Hence \begin{equation}\label{wsref-0-61-72-65-61-3a-61-67-73-70} \delta_{r^m}\le\frac{\eta^{2m}}{|\alpha|^2}. \end{equation} This is a valid conditional route to (area:tails), but the prefactor cannot be discarded. The best squared product overlap is $\lambda_1$, and the elementary lower bound is only $\lambda_1\ge D_A^{-1}$. Removing its possible volume-dependent cost by choosing a larger initial $m$ can itself require \[ \log R_0\quad\text{of order}\quad \frac{\log r}{2\log(1/\eta)}\log(1/|\alpha|^2). \] An area-law proof needs a better overlap argument or a projector construction that compensates for this cost. Assuming the desired Schmidt tail as the starting overlap estimate would be circular.

The elementary gap filter is insufficient quantitatively.

Let $E_0$ be the ground energy and choose $W\ge\|H-E_0\|$. Then \[ K_0=I-(H-E_0)/W \] fixes the ground state and contracts its orthogonal complement by at most $1-\Delta/W$. A sum of uniformly bounded local terms permits $W=O(|\Lambda|)$, which gives only a contraction guarantee approaching one with volume. Across a nearest-neighbor cut, the operator Schmidt rank of $K_0$ is at most $2+O(|\partial A|)d^4$: terms wholly on either side contribute two product operators, and each crossing term has bounded operator Schmidt rank. Fixed finite range changes the geometric constants for ordinary rectangular lattice boxes.

Thus the straightforward estimate for $\log r/\log(1/\eta)$ can grow like $|\Lambda|\log(2+|\partial A|)/\Delta$. Raising this same filter to powers improves contraction and increases the rank cost together; it does not remove the ratio. This proves no impossibility theorem for better filters. It identifies why this immediate use of a gap does not supply the constants in (area:tails).

Verification, gap, and outcome.

The companion calculations check the entropy of the explicit tail state, the dimer count, and the geometric-tail entropy bound on finite distributions. The arguments above cover every finite Schmidt dimension and use the actual eigenvalue distribution, not a continuity bound with its dimension factor omitted.

The first missing step is to derive a uniform family of Schmidt-tail bounds such as (area:tails) from the stated two-dimensional Hamiltonian assumptions. Concrete continuation would improve the projector rank-versus-contraction tradeoff, control the starting overlap without assuming an area law, or establish a different dimension-independent entropy estimate. Existing extra hypotheses must remain visible when applying their constructions.

UNRESOLVED. Finite-depth preparations satisfy the claimed scaling, and the all-scale tail criterion is proved. The general uniform-gap hypothesis has not been shown to imply that criterion. The high-entropy test state disproves an approximation shortcut, not the area-law conjecture. No novelty or formal verification is claimed.

Sources

• S1 I. Arad, R. Firanko, R. Jain, Area Laws and Tensor Networks for Maximally Mixed Ground States, Communications in Mathematical Physics 407, 54 $2026$. https://link.springer.com/article/10.1007/s00220-026-05554-z. Formulation source.
• S2 Universal lower bound on topological entanglement entropy. https://arxiv.org/abs/2302.00689. Further reference; not a proof endorsement.
• E1 Anurag Anshu, Itai Arad, and David Gosset, An area law for 2D frustration-free spin systems, arXiv:2103.02492. https://arxiv.org/abs/2103.02492. Primary source for a result with additional local-gap and frustration-free hypotheses, not the unrestricted statement; consulted 27 September 2026.