Why a gapped quantum sheet should have only boundary-sized entanglement

← All digests

Family 265 · Mathematical physics
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-25

1. The question

A sheet of quantum spins has an enormous state space. If there are $n$ sites and each has $q$ states, a general pure state requires $q^n$ complex amplitudes. Yet ground states of local physical systems often appear much simpler. How much information can a region share with the rest of the sheet?

An area law says that this shared quantum information grows with the boundary of the region, rather than with its volume. In two spatial dimensions, “area” means boundary length. Family 265 claims such an entropy bound for unique ground states of finite-range square-lattice Hamiltonians using only a positive gap above the full system’s ground state.

A companion manuscript then claims a compact tensor-network approximation on an open square grid. These are related but separate statements: controlling each cut’s entropy does not automatically produce one globally consistent approximation across all cuts.

2. Background and essential definitions

At each lattice site place a copy of $\mathbb C^q$. The full Hilbert space is their tensor product. A local Hamiltonian has the form

$$H=\sum_X h_X,$$

where each Hermitian term $h_X$ acts only on a small set $X$. The interaction range $R$ bounds the diameter of these supports, and $\|h_X\|\leq J$ bounds their strength. The manuscript indexes at most one term per support, so fixed range and strength also bound the total interaction strength touching each site.

Let $\Omega$ be the unique normalized ground vector, up to its physically irrelevant phase, and let $E_0$ be its energy. A spectral gap $\Delta>0$ means

$$H-E_0 I\geq\Delta\bigl(I-|\Omega\rangle\langle\Omega|\bigr).$$

Every state orthogonal to the ground vector costs at least $\Delta$ additional energy. For a family of growing systems, the claimed uniform conclusion requires a lower bound on $\Delta$ independent of system size.

For a region $A$, discard the other sites by taking a partial trace. This gives its reduced density matrix

$$\rho_A=\operatorname{Tr}_{A^c}|\Omega\rangle\langle\Omega|.$$

Its von Neumann entropy is $S(A)=-\operatorname{Tr}(\rho_A\log\rho_A)$, with $0\log0=0$. For a pure global state, this measures entanglement between the two sides. The elementary bound is $S(A)\leq |A|\log q$, proportional to the number of sites.

A projected entangled-pair state, or PEPS, replaces the full amplitude table by local tensors. Each site tensor has a physical index and virtual indices joined along lattice edges. Summing over paired virtual indices gives the many-body vector. The largest virtual index size is the bond dimension $D$. Small $D$ limits the tensor sizes, although finding or contracting the tensors may still be difficult.

3. Development history and previous partial results

Early links between entropy and boundary area arose in quantum-field studies motivated by black-hole entropy, including work of Bombelli and collaborators and Srednicki. These provided important models and physical motivation. The finite-dimensional lattice-spin question asks for a theorem under explicit locality and gap assumptions.

Hastings’s one-dimensional area law established a size-independent entropy bound for unique gapped ground states of local spin chains. It also connected the gap to matrix-product-state approximation. Later work by Arad, Landau, Vazirani and collaborators sharpened the method using approximate ground-state projectors: operators that preserve the ground state while suppressing excitations without introducing too much entanglement.

Exponential decay of correlations is another consequence of a spectral gap, but it does not by itself transparently bound the entire reduced-state spectrum. Entropy sums contributions from all eigenvalues of $\rho_A$, including many individually small ones. Brandão and Horodecki’s one-dimensional result uses decay of correlations between arbitrary separated regions, a stronger information condition than a statement just about pairs of single-site observables.

Higher-dimensional area laws were known in restricted settings, such as suitable Gaussian oscillator systems. Other results used low-energy counting, heat-capacity assumptions, local gaps, or frustration-free interactions. A Hamiltonian is frustration-free when its ground state minimizes each local term simultaneously; a general ground state need not do this.

In two dimensions, Anshu, Arad and Gosset obtained subvolume estimates and subsequent area-law results under local-gap and frustration-free hypotheses, with specified cuts. M1’s historical discussion distinguishes those hypotheses and geometric settings from the global-gap, arbitrary-region theorem it claims here.

PEPS were introduced as higher-dimensional counterparts of matrix product states. Earlier approximation results addressed local observables, stronger entropy hypotheses, extra low-energy spectral assumptions, or tensor networks on an enlarged lattice. Those variants remain distinct from a small global vector error on the original grid. M2 expressly develops more than the entropy bound to reach that conclusion.

4. What the new manuscripts claim

An entropy law for every region

M1 considers any finite induced subgraph $\Lambda$ of the square lattice. “Induced” means that every nearest-neighbor edge between included sites is present. The domain may have holes or disconnected components. Interaction range is measured in the graph distance within that domain, with infinite distance between different components.

For fixed $q,R,J,\Delta$, it claims a finite constant $C$ such that every admissible unique gapped ground state satisfies

$$S(A)\leq C\,|\partial_\Lambda A| \qquad(A\subseteq\Lambda).$$

Here $\partial_\Lambda A$ is the set of nearest-neighbor edges with one endpoint in $A$ and the other in $\Lambda\setminus A$. The constant is independent of domain size, domain shape, the chosen region and the individual Hamiltonian.

The gap assumption is imposed only on the full Hamiltonian. There is no requirement that every restriction to a smaller region remain gapped. Nor are commutativity, translation invariance or frustration freeness assumed. If the boundary is empty, the stated inequality gives zero entropy; the induced-graph interaction convention and uniqueness are relevant to that degenerate case.

A global PEPS approximation

M2 specializes to an open $L\times L$ square grid, $L\geq2$, with bounded on-site and nearest-neighbor interactions. For fixed $q\geq2$, $J$ and $\Delta$, it claims constants $C,c>0$ and a nonzero PEPS $\Phi$ on that same grid with

$$D\leq C L^c,$$

and

$$\min_{\theta\in\mathbb R} \left\|\frac{\Phi}{\|\Phi\|}-e^{i\theta}\Omega\right\| \leq L^{-1}.$$

The norm is the Hilbert-space norm of the entire state. The minimization accounts for global phase. The tensors can have arbitrary complex entries and need not be isometries. Constants are uniform over Hamiltonians with the fixed local parameters.

This is an existence theorem. It does not assert that the tensors can be found efficiently, that the resulting PEPS can be contracted efficiently, or that the approximant has an injective structure or a gapped parent Hamiltonian. It also does not silently replace the open square by a torus or claim the same representation theorem for every irregular domain covered by M1.

5. Why the advance matters

Provisional importance: 5/5. A general two-dimensional gap-to-area theorem would resolve a central structural question about quantum many-body ground states. The asserted removal of local-gap and frustration-free assumptions is particularly significant, because physical interactions often compete and restricted Hamiltonians can behave differently from the full system.

The PEPS result adds a global representation statement with polynomial bond dimension. It gives a mathematical explanation for compact tensor-network descriptions in the claimed class, while keeping the computational task separate. A concise representation can exist even when discovering its coefficients is hard.

The two manuscripts also show why the distinction matters conceptually. An entropy estimate concerns a bipartition at a time. A state representation must make all those local descriptions compatible, control their accumulated errors, and route the required indices through the actual grid. That extra work is a substantive part of the family.

6. The main idea

M1 compares information gained with excitation energy spent. Its gap says that a state with very little excess energy must lie mostly in the ground-state direction. The proof seeks constructions that would create too much independent information at too little energy if the desired area bound failed.

The first step obtains a weaker entropy estimate, below volume growth, on suitable boxes. Local spectral filters then replace the Hamiltonian by positive constraints that individually annihilate $\Omega$, retain a global gap, and have rapidly decaying spatial tails. This is an analytical construction; the original Hamiltonian need not have been frustration-free.

A useful information quantity is mutual information:

$$I(B:C)=S(B)+S(C)-S(B\cup C)$$

for disjoint regions. The proof moves partitions through a collar separating a target region from the remote system. It encodes entropy comparisons using operators on many symmetric copies of the state. The same interpolating family controls both information and energy, allowing the global gap to constrain the information cost.

The resulting collar estimate is strengthened using localized products of the positive constraints. Finally, the planar geometry permits two ordered families of separated regions after removing only a boundary-sized collection of sites. Two entropy-chain-rule comparisons cancel the interior contributions. The remaining exceptional sites and summed information errors have boundary size, yielding the claimed area law.

M2 imports the area law and stronger collar estimates from M1. It does not infer PEPS approximation from entropy alone. Its construction imagines distributing physical registers among auxiliary parties, beginning with one party holding the whole vector and ending with one party per site. This starting point is legitimate for an existence theorem and also makes clear why it is not an algorithm for discovering the ground state.

At successive geometric scales, small information across buffers allows registers to change owners with controlled error. Small patches near junctions need separate encodings, because ordinary separation disappears there. The protocol is arranged so each party participates in only boundedly many interactions over its lifetime.

A compression theorem then replaces large auxiliary pair resources by polynomial-size sums in a density-operator calculation. A suitable product-basis column recovers an approximate vector. Finally, the connections are routed through the original square grid with bounded congestion, so polynomial virtual dimensions remain polynomial. The error budget is chosen to control the normalized global vector, rather than only individual reduced states.

7. Scope, limitations, and open questions

Uniqueness of the ground state, finite local dimension, bounded local interactions and a positive full-system gap are explicit hypotheses. The displayed theorem does not address arbitrary degenerate ground spaces, gapless systems, unbounded oscillator sites or all higher spatial dimensions.

The area law covers finite induced square-lattice domains with intrinsic graph-range interactions. The PEPS theorem has the narrower open-square nearest-neighbor setting and the stated inverse-linear accuracy. Alternative boundary conditions, stronger approximation guarantees and additional tensor structure require their own arguments.

Neither compact entanglement nor polynomial bond dimension alone settles efficient simulation. Searching for the tensors, representing their entries with controlled numerical precision, and contracting the network are separate problems. The manuscript explicitly leaves those algorithmic claims outside its theorem.

8. Sources and further reading

Hastings and the Anshu–Arad–Gosset papers linked in section 3 provide primary background for the one-dimensional theorem and restricted two-dimensional advances. M1’s historical discussion separates correlation decay, entropy, local gaps and global gaps. M2 compares global PEPS approximation with local-observable and enlarged-lattice formulations, and its Appendix A records the stronger companion estimates used beyond the entropy inequality itself.

Source manuscripts

• M1. OpenAI, A two-dimensional area law from a global spectral gap (September 24, 2026).
• M2. OpenAI, Polynomial PEPS approximation of gapped square-grid ground states (September 24, 2026).

Sources consulted on 25 October 2026 at repository revision 1581d4c. Download the manuscript citations.