Resource-Scalable Fully Quantum Metropolis-Hastings for Integer Linear Programming

Abstract: Integer linear programming (ILP) remains computationally challenging due to its NP-complete nature despite its central role in scheduling, logistics, and design optimization. We introduce a fully quantum Metropolis-Hastings algorithm for ILP that implements a coherent random walk over the discrete feasible region using only reversible quantum circuits, without quantum-RAM assumptions or classical pre/post-processing. Each walk step is a unitary update that prepares coherent candidate moves, evaluates the objective and constraints reversibly -- including a constraint-satisfaction counter to enforce feasibility -- and encodes Metropolis acceptance amplitudes via a low-overhead linearized rule. At the logical level, the construction uses $\mathcal{O}(n\log_2 N)$ qubits to represent $n$ integer variables over the interval $[-N,\,N-1]$, and the Toffoli-equivalent cost per Metropolis step grows linearly with the total logical qubit count. Using explicit ripple-carry adder constructions, we support linear objectives and mixed equality/inequality constraints. Numerical circuit-level simulations on a broad ensemble of randomly generated instances validate the predicted linear resource scaling and exhibit progressive thermalization toward low-cost feasible solutions under the annealing schedule. Overall, the method provides a coherent, resource-characterized baseline for fully quantum constraint programming and a foundation for incorporating additional quantum speedups in combinatorial optimization.

• The paper presents a fully coherent quantum Metropolis-Hastings algorithm for integer linear programming that generates proposals, evaluates objectives, checks constraints, and filters acceptance without QRAM or classical processing.
• The circuit uses O(n log₂ N) logical qubits and a per-step Toffoli-equivalent cost linear in total qubit count, with simulations of more than 1,500 instances supporting the resource-scaling claims and R² > 0.99 for qubit growth.
• The method reduces arithmetic overhead by replacing equality constraints with paired inequalities and approximating acceptance rotations linearly, but convergence remains instance-dependent and has only been demonstrated on small problems of up to nine variables.

Overview

The paper "Resource-Scalable Fully Quantum Metropolis–Hastings for Integer Linear Programming" (2602.11285) presents a fully coherent quantum algorithm for integer linear programming (ILP) in which every component of a Metropolis–Hastings sampler—proposal generation, objective evaluation, constraint checking, and acceptance filtering—is realized as a reversible quantum circuit. The construction requires no quantum RAM, no preloaded data structures, and no classical pre- or post-processing, distinguishing it from quantum interior-point methods for LP and SDP that depend on sparsity and QRAM-style data access (2602.11285). The walk operates on a fully connected transition graph over the discretized search space, with feasibility enforced dynamically by the acceptance filter rather than by pruning edges.

The central quantitative claims are two scaling results: the logical qubit count scales as $\mathcal{O}(n\log_2 N)$ for $n$ variables over $[-N, N-1]$, and the per-step cost in Toffoli-equivalent gates scales linearly in the total qubit count $k$. Both are validated empirically on more than 1500 randomly generated instances via gate-level simulation in Qiskit, with linear regression fits achieving $R^2 > 0.99$ for the qubit-count relation.

Algorithmic structure

The one-step walk operator is

$W = R\,P^{\dagger}\,F\,P,$

acting on six registers: the system register $S$ holding the current assignment $\mathbf{x}$, the proposal register $S'$, function registers $F$ and $n$0 storing $n$1 and $n$2, a constraint counter $n$3, and a single coin qubit $n$4. The preparation stage $n$5 generates a uniform superposition of candidates in $n$6, evaluates all constraints into $n$7 and the candidate cost into $n$8 (operator $n$9), then encodes the Metropolis acceptance amplitude $[-N, N-1]$0 onto the coin via controlled rotations (operator $[-N, N-1]$1), where $[-N, N-1]$2. The shift $[-N, N-1]$3 is a controlled-SWAP between $[-N, N-1]$4 and $[-N, N-1]$5, conditioned jointly on the coin being $[-N, N-1]$6 and the counter reaching $[-N, N-1]$7, so only feasible, accepted moves execute. The reflection $[-N, N-1]$8 supplies the Szegedy-type spectral structure.

The annealing loop follows the quantum simulated annealing framework of Somma et al., with a randomized number of walk applications per temperature step and partial measurements on the coin and movement registers between temperature stages. Convergence is governed not by classical mixing times but by phase-estimation resolution of the spectral gap $[-N, N-1]$9 of $k$0, which inherits the quadratic Szegedy speedup $k$1 relative to the classical gap $k$2. A notable robustness claim is that, unlike Grover-type amplitude amplification, extending the schedule does not cause oscillatory overshoot; the authors observe non-oscillatory, cumulative probability concentration across all tested instances, while conceding that a strict proof of monotonicity is outside the scope of the paper.

Reversible arithmetic and constraint encoding

All arithmetic uses two's-complement encoding with the CDKM ripple-carry adder as the primitive; multiplication by an integer coefficient $k$3 costs $k$4 adder invocations, giving $k$5 Toffoli gates for evaluating a linear form. Equality constraints are checked by full multi-qubit comparison against zero, incurring quadratic Toffoli overhead in the workspace bit width, whereas inequality constraints reduce to a sign-bit test with only linear cost. The paper formalizes this asymmetry in a theorem on "resource-efficient equality elimination": each equality $k$6 can be replaced by the pair $k$7 and $k$8, preserving the feasible region exactly and reducing the total constraint count to $k$9 pure inequalities with strictly lower asymptotic Toffoli overhead. This reformulation is used throughout the numerical experiments.

A second efficiency device avoids computing exponentials entirely. The acceptance rotation angle $R^2 > 0.99$0 is replaced by a monotone linear interpolation over the binary-encoded bits of $R^2 > 0.99$1, with fit coefficients depending only on register width and $R^2 > 0.99$2, not on instance data. An appendix corollary proves that any target distribution satisfying the detailed-balance condition yields the Gibbs-like state as a unique eigenvalue-1 eigenstate of $R^2 > 0.99$3, so the approximation preserves thermal ordering even though the stationary distribution deviates slightly from the exact Boltzmann form. This is a deliberate trade-off: exact Gibbs sampling is sacrificed for optimization-oriented behavior, and the ground-state support remains intact under monotone reparameterizations.

Resource characterization and numerical validation

The register budget gives $R^2 > 0.99$4 with logarithmic corrections from the constraint counter and additive terms from the arithmetic registers. Gate counts use a Toffoli-equivalent cost model in which an MCX with $R^2 > 0.99$5 controls contributes $R^2 > 0.99$6 equivalents, controlled unitaries contribute $R^2 > 0.99$7, and $R^2 > 0.99$8 gates are charged at one Toffoli-equivalent per seven gates to reflect distillation overhead.

Three simulation campaigns support the claims:

Study	Instances	Finding
Toffoli count vs. total qubits, varying $R^2 > 0.99$9	1000	Strictly linear; slope grows systematically with constraint density
Toffoli count vs. total qubits, varying $W = R\,P^{\dagger}\,F\,P,$0	1000	Linear with invariant slope; variable count shifts the offset only
Total qubits vs. $W = R\,P^{\dagger}\,F\,P,$1	1500	Linear ($W = R\,P^{\dagger}\,F\,P,$2), confirming $W = R\,P^{\dagger}\,F\,P,$3 spatial scaling

Constraint density thus contributes additively to the cost coefficient without altering the linear scaling law. End-to-end runs at up to three discretization qubits per variable show progressive thermalization toward low-cost feasible configurations under a hybrid linear-exponential $W = R\,P^{\dagger}\,F\,P,$4 schedule with randomized repetition depth, with the search domain spanning $W = R\,P^{\dagger}\,F\,P,$5 in the largest demonstrated case.

Limitations and open questions

The paper is explicit about several boundaries of its results. All scaling estimates are logical-circuit complexities assuming a fault-tolerant machine; error-correction and connectivity overheads are excluded. The number of annealing steps required for convergence depends on the instance-specific spectral gap and temperature schedule, introducing a heuristic element typical of annealing methods—the rigorous $W = R\,P^{\dagger}\,F\,P,$6 bound applies to exact sampling, and the observed favorable empirical convergence for optimization is not backed by a worst-case guarantee. The claimed monotone improvement of success probability with schedule length is supported numerically but lacks a formal proof. Coefficient magnitudes enter the gate count through $W = R\,P^{\dagger}\,F\,P,$7 repeated adder invocations, so instances with large coefficients inflate the per-step cost; no coefficient-compression technique is proposed. Finally, demonstrations are limited to very small instances (at most nine variables, seven constraints, three qubits of discretization), leaving open whether the thermalization behavior persists at scales where the exponential classical search space becomes practically relevant.

Conclusion

This work provides a complete, circuit-level realization of a quantum Metropolis–Hastings sampler specialized to ILP, with all arithmetic performed coherently and resource requirements characterized both analytically and empirically. Its main contributions are the proven linear scaling of logical resources per walk step, the equality-to-inequality reformulation that strictly reduces asymptotic Toffoli cost, and the demonstration that a monotone linearized acceptance rule suffices for correct thermalization. The method positions itself as a resource-characterized baseline for fully quantum constraint programming rather than a claim of end-to-end speedup; whether the empirically observed convergence advantages survive at problem sizes beyond the simulated regime remains the principal open question.