Undecidability of the Spectral Gap (full version)

Abstract: We show that the spectral gap problem is undecidable. Specifically, we construct families of translationally-invariant, nearest-neighbour Hamiltonians on a 2D square lattice of d-level quantum systems (d constant), for which determining whether the system is gapped or gapless is an undecidable problem. This is true even with the promise that each Hamiltonian is either gapped or gapless in the strongest sense: it is promised to either have continuous spectrum above the ground state in the thermodynamic limit, or its spectral gap is lower-bounded by a constant in the thermodynamic limit. Moreover, this constant can be taken equal to the local interaction strength of the Hamiltonian.

• The paper shows that the spectral gap problem for 2D quantum spin Hamiltonians is undecidable by reducing it to the Halting Problem.
• The authors construct a family of translationally invariant Hamiltonians using quantum Turing machines and aperiodic tilings to encode computational history states.
• The result implies that quantum many-body phase diagrams can exhibit uncomputably complex behavior, highlighting fundamental limits in predicting macroscopic properties from microscopic descriptions.

This paper (1502.04573) presents a groundbreaking result in mathematical physics and theoretical computer science: the undecidability of the spectral gap for a family of 2D quantum spin lattice Hamiltonians. This means there is no algorithm that can take such a Hamiltonian as input and determine whether it has a gap above the ground state or not. The result holds even under strong conditions, such as translational invariance, nearest-neighbour interactions, a fixed local Hilbert space dimension, and algebraic interaction strengths, and even for Hamiltonians that are arbitrarily small quantum perturbations of classical systems.

The spectral gap is the difference between the ground state energy (lowest energy eigenvalue) and the first excited state energy. It is a crucial property in condensed matter physics, influencing phenomena like phase transitions, correlation lengths, and the stability of ground state properties against perturbations. A gapped system has a non-zero minimum energy required to excite it from the ground state, leading to stable ground states and short-range correlations. A ^^^^1^^^^ system has excitations arbitrarily close to the ground state energy, often associated with quantum phase transitions and long-range correlations. The paper defines "gapped" and "gapless" in a strong sense to avoid ambiguity (\cref{def:gapped,def:gapless}).

The core of the main result (\cref{thm:promise}) is the explicit construction of a family of translationally-invariant, nearest-neighbour Hamiltonians $H^{\Lambda(L)}(n)$ on a 2D square lattice for a fixed local Hilbert space dimension $d$. These Hamiltonians are indexed by a natural number $n$. The construction is such that determining whether $H^{\Lambda(L)}(n)$ is gapped or gapless in the thermodynamic limit ($L \to \infty$) is equivalent to determining whether a given Universal Turing Machine (UTM) halts on input $n$. Since the Halting Problem is undecidable, the spectral gap problem for this family of Hamiltonians is also undecidable.

The constructed Hamiltonians exhibit specific properties crucial for the proof:

• They are defined by three Hermitian matrices: $h_1(n)$ (on-site), $h_{\text{row}(n)$ (horizontal), and $h_{\text{col}(n)$ (vertical).
• The local Hilbert space dimension $d$ is fixed for the entire family, depending only on the details of the construction, not on $n$ or $L$.
• The matrix entries are algebraic numbers, meaning they are computable.
• The interactions $h_{\text{row}(n), h_{\text{col}(n)$ are arbitrarily small quantum perturbations of a classical Hamiltonian (defined by matrices $A$ and $D$, which are diagonal with integer entries). The quantum perturbation strength is controlled by a rational parameter $\beta$ that can be taken as small as desired.
• The dependence on $n$ appears in the on-site term $h_1(n)$ via a prefactor $\alpha(n)$ and in the row interaction $h_{\text{row}(n)$ via phase factors involving $e^{i\pi\varphi}$ and $e^{i\pi2^{-\abs{\varphi}}$, where $\varphi$ is a rational number encoding $n$.

The implications of this result for physics are profound:

• It suggests that phase diagrams of certain quantum many-body systems can be uncomputably complex.
• It shows that predicting the thermodynamic limit behavior (gapped or gapless) from finite-size properties can be impossible for some systems, as the gap could close and reopen at arbitrarily large system sizes in an uncomputable manner.
• It implies that general gapped systems are not robust to arbitrary local perturbations, in contrast to known results for frustration-free or free-fermion systems.
• It opens the possibility that answering fundamental conjectures about the spectral gap for specific physical models (like the Haldane conjecture for 1D spin chains or the existence of spin liquid phases in 2D) might be axiomatically undecidable from standard mathematical axioms.

The proof of \cref{thm:promise} is constructive and involves several intricate steps, combining techniques from quantum information theory, quantum computation, and aperiodic tilings.

1. Quantum Phase Estimation Turing Machine (Section 3): The first step is to build a family of Quantum Turing Machines (QTMs) indexed by $n$. A key challenge is feeding an arbitrary input $n$ to a UTM encoded in a fixed-size QTM (fixed alphabet and number of internal states). This is achieved by constructing a QTM $P_n$ based on the quantum phase estimation algorithm. When given a unary representation of a sufficiently large number $N$ as input, $P_n$ deterministically computes the binary representation of $n$ on its tape and halts. The transition rule amplitudes of $P_n$ depend on $n$ via phases involving $e^{i\pi\varphi}$. The construction uses various reversible TM subroutines (copying, shifting, comparing, incrementing, looping, unary-to-binary conversion) and ensures the QTMs are well-formed, normal form, and unidirectional, properties crucial for later encoding into a Hamiltonian. The inverse Quantum Fourier Transform required for phase estimation introduces an exponential runtime $O(\poly(N) 2^N)$, but this is acceptable as the QTM is used only to generate input and halts, not for general computation efficiency.
2. Encoding QTMs in Local Hamiltonians (Section 4): Building on the work of Kitaev [quant-ph/9707021, quant-ph/0204112] and Gottesman & Irani [quant-ph/(0905.2419)], the evolution of the family of QTMs $P_n$ is encoded into the ground state of a 1D translationally-invariant nearest-neighbour Hamiltonian. The local Hilbert space of the 1D chain is structured into multiple "tracks" (\cref{fig:track-cartoon}): a clock oscillator (Track 1), a counter TM (Tracks 2 & 3), the QTM (Tracks 4 & 5), and a time-wasting tape (Track 6). The clock oscillator provides a periodic signal. The counter TM increments a number based on this signal, tracking the computation time. The QTM evolves one step for each full cycle of the clock oscillator. The Hamiltonian's zero-energy ground state (within a specific subspace of "bracketed" configurations) is a superposition of computational history states (\cref{def:history-state}), where each term corresponds to a configuration of the QTM at a specific time step. Penalty terms are used to constrain the valid configurations, and transition terms encode the evolution rules. A crucial technical point is ensuring that the effective quantum transitions on the QTM state are unitary, even when the QTM halts or runs out of tape space; this is handled by switching to a "time-wasting" computation on Track 6 that uses up the remaining clock cycles without affecting the QTM's final output tape configuration. The "Clairvoyance Lemma" (\cref{clairvoyance}) provides a general framework to prove that only the history state has zero energy, provided the Hamiltonian is in "standard form" (\cref{def:standard-form_H}) and the rules satisfy certain properties related to well-formedness and reachability (\cref{evolve_to_illegal}). The result is a 1D Hamiltonian whose ground state energy depends on whether the encoded QTM halts on an input determined by the length of the chain. However, this energy difference vanishes exponentially in the chain length.
3. Quasi-periodic Tilings (Section 5): To amplify the exponentially small energy difference to something substantial in the thermodynamic limit, the construction moves to a 2D lattice and exploits the geometric properties of Robinson's aperiodic tiling [Berger, Robinson]. The Hamiltonian is constructed using a "tiling layer" and a "quantum layer" (\cref{tiling+quantum}). The tiling layer enforces configurations corresponding to a valid Robinson tiling of the 2D lattice. A modified set of Robinson tiles is used to ensure the periodic structure of "borders" (lines formed by specific tiles) is robust. The quantum layer encodes the 1D QTM Hamiltonian from the previous step. The coupling between layers is designed such that the 1D QTM Hamiltonian effectively runs only on specific horizontal segments within the tiling pattern, specifically on the "borders" of size $4^n$ (\cref{fig:tiling,fig:3-square}). Crucially, the paper proves new "rigidity" results for the Robinson tiling (\cref{Robinson_rigidity,segment_bound}) showing that these border segments appear with a certain density in large regions, even in the presence of defects.
4. Putting it Together (Section 6): The 1D Hamiltonian whose ground state energy encodes the Halting Problem (from Section 4) is incorporated into the quantum layer of the 2D tiling Hamiltonian (from Section 5). This creates a Hamiltonian $H_u$ whose ground state energy density depends on whether the UTM halts on input $n$ (\cref{thm:gs_density}). Specifically, if the UTM halts, the 1D Hamiltonians on the tiling segments corresponding to sufficient computation time have positive ground state energy, leading to a positive ground state energy density for the 2D Hamiltonian $H_u$. If the UTM does not halt, these 1D Hamiltonians have zero ground state energy, leading to a zero ground state energy density for $H_u$. This proves the undecidability of the ground state energy density problem (\cref{thm:gs_density}).

   Finally, the paper constructs the main Hamiltonian $H$ by combining $H_u$ with two other Hamiltonians: a trivial gapped Hamiltonian $H_0$ (with a unique 0-energy ground state and a gap of 1, implemented by a projector term) and a critical gapless Hamiltonian $H_d$ (with 0 ground state energy density but spectrum becoming dense near zero in the thermodynamic limit). The construction carefully combines the local Hilbert spaces and interactions such that the overall spectrum is related to the sum of the spectra of $\tilde{H}_u, \tilde{H}_d$, and $\tilde{H}_0$ (\cref{eq:promise_H}).

   • If the UTM halts, $\tilde{H}_u$ has positive minimum energy $\lambda_0(\tilde{H}_u) \geq 1$ for sufficiently large $L$ (\cref{promise_Hu}). Since $\tilde{H}_d \geq 0$, the lowest energy state comes from $\tilde{H}_0$, which has a gap of 1. Thus $H$ is gapped with a gap of at least 1 (\cref{fig:gs-energy-to-spectral-gap}).
   • If the UTM does not halt, $\lambda_0(\tilde{H}_u) \leq 0$ for all $L$. The spectrum of $H$ near zero is dominated by the sum of the spectra of $\tilde{H}_u$ and $\tilde{H}_d$. Since $\tilde{H}_d$ is gapless and $\tilde{H}_u$ can have energies at or below zero, the combined spectrum is gapless.

This completes the proof of \cref{thm:promise} for open boundary conditions. The extension to periodic boundary conditions is achieved by modifying the tiling layer using special "boundary tiles" (\cref{fig:boundary_tiles,eq:boundary_tile_weights}) that force the tiling on a torus to effectively contain a linear defect line, mapping the periodic boundary condition problem to an open boundary condition problem on a slightly smaller lattice (\cref{periodic-to-open-bc}).

From an implementation perspective in terms of building such systems, the paper provides the detailed mathematical description of the local Hilbert space dimension ($d$ is large but fixed, determined by the size of the QTM alphabet and state space, and the number of tracks) and the algebraic matrix entries for the local interaction terms ($h_1, h_{\text{row}}, h_{\text{col}}$). While these Hamiltonians are explicitly constructed, their complexity (large local dimension, intricate interactions encoding complex logic) makes physical realization challenging. The key takeaway is not how to build these specific systems, but the fundamental computational limitations they reveal about properties of quantum matter, suggesting that for certain systems, determining basic macroscopic properties from microscopic descriptions is an impossible task in principle. Numerical simulation of such systems would also face this fundamental barrier; even if memory and time allowed simulation of large $L$, predicting the infinite-size limit behavior remains undecidable.