When a quantum field theory is simulated on a quantum computer or with tensor networks, the answer you want is never computed directly. You compute it on a grid with spacing $a$, with each continuous field squeezed into a finite number of states (a "Hilbert space truncation", for example keeping only electric-field values $|E| \le \Lambda$). You then extrapolate to zero spacing and infinite cutoff. On a quantum device, every number entering that extrapolation is also an average over finitely many measurement repetitions ("shots"), so it carries noise that shrinks only as $1/\sqrt{M}$. Today these three error sources are handled one at a time. The cutoff is either fixed generously or extrapolated away first at each spacing. The continuum fit comes second. Shot noise, if it is considered at all, is driven down by brute force. Nobody asks how a finite shot budget should be split across spacings and cutoffs so that the final continuum number is as accurate as possible.
We propose to treat the cutoff as a design variable chosen jointly with spacing and shot count. The method minimizes the mean squared error of the extrapolated continuum value under a fixed total cost. Because truncation bias falls very fast with cutoff, and because measurement cost grows with it, we expect the best design to use a different, often surprisingly small, cutoff at each spacing. The design should spend shots unevenly, and its validation step should certify that leftover truncation bias sits below the statistical error. The test is whether this design resolves a physics question that existing deterministic studies settled only classically: the 0.015 gap between two competing predictions for the Schwinger-model pion mass. We ask whether it does so at a fraction of the shots that naive designs need.
Core Claim
Consider a continuum observable $c_0$, estimated as a weighted combination of measured averages $\bar y_i$ at design points $(a_i, \Lambda_i)$, with $M_i$ shots each. Its error decomposes as
$$\mathrm{MSE} = \Big(\sum_i w_i\, b_i(a_i,\Lambda_i)\Big)^2 + \sum_i \frac{w_i^2\, \sigma_i^2}{M_i}, \qquad \text{subject to } \sum_i M_i\, C(a_i,\Lambda_i) \le \mathcal{C}_{\text{tot}}.$$
Here $w_i$ are the extrapolation weights implied by the continuum ansatz (for instance $c_0 + c_1 a^2$). The bias $b_i$ collects truncation error and any part of the true $a$-dependence the ansatz misses. $\sigma_i^2$ is the single-shot variance of the measured operator. $C$ is the cost per shot, which grows with $\Lambda$ through qubit count and the number of Pauli strings to measure, as Toward scalable simulations of lattice gauge theories on quantum computers quantifies for spin-$S$ link models.
For fixed design points, the shot split $M_i \propto |w_i|\sigma_i/\sqrt{C_i}$ is the standard Lagrange-multiplier answer, and we do not claim it. Our claim is about two things that standard designs leave out. The first is that $\Lambda_i$ enters both the bias and the cost. The second is that truncation bias decays roughly exponentially, which Quantum simulation of the Schwinger model: A study of feasibility shows for the Schwinger model, where a three-state link already reproduces the continuous-group energy density. Together these mean the optimal $\Lambda_i$ sits where the marginal bias reduction equals the marginal noise reduction that the saved cost would buy elsewhere. We will show that choosing $\Lambda_i$ this way, with bias estimated from cheap pilot runs, cuts the shots needed for a target continuum error substantially compared with fixed-cutoff designs. It must do so without introducing detectable bias.
Relation to Prior Work
Several studies motivate the gap because they already perform two of the three extrapolations, but deterministically. Dynamics in hamiltonian lattice gauge theory: approaching the continuum limit with partitionings of SU(2) discretizes the SU(2) group manifold into point sets with finite-element derivatives. It restricts the manifold to a cap whose radius scales with the coupling $g$, giving constant cost toward the continuum. It extrapolates mesh density first and coupling second, obtaining a single-plaquette mass gap $M_0 = 4.008(27)$ against the exact value 4. Fermion Discretization Effects in the Two-Flavor Lattice Schwinger Model: A Study with Matrix Product States extrapolates tensor-network bond dimension first and spacing second. It reports continuum pion masses $M_\pi/g$ of 0.211(5) for staggered, 0.2103(22) for Wilson and 0.2117(11) for twisted-mass fermions, favoring Hosotani's 0.213 over Smilga's 0.198. Both papers are noise-free, and both use the sequential "truncation first, continuum second" pipeline. We keep their systems and their Hamiltonians but replace exact expectation values with sampled ones, and we make the order of extrapolation a design question.
Matching Lagrangian and Hamiltonian simulations in (2+1)-dimensional U(1) gauge theory finds that Monte Carlo results extrapolated to zero temporal spacing lie consistently slightly below truncated-Hamiltonian exact diagonalization, within about $2\sigma$, with the cause unknown. This motivates our certification step: a design that budgets truncation bias explicitly can say whether a residual of that size is attributable to the cutoff. Review on Quantum Computing for Lattice Field Theory notes that $a \to 0$ and $\Lambda \to \infty$ must be taken together but gives no allocation rule. The Mathis et al. resource analysis supplies our cost model $C(a,\Lambda)$ but stops before extrapolation or shot counting.
Two works constrain the design. Chiral and continuum extrapolation of partially-quenched lattice results shows that unconstrained polynomial continuum fits chase statistical fluctuations. This warns us that an optimizer allowed to pick extreme spacings will exploit an ansatz it cannot verify, so our objective includes a misspecification term in $b_i$ rather than trusting the fitted form. Limits on non-Gaussianities from WMAP data recovers minimum-variance sensitivity by adding a correction term that removes spurious couplings from a sky mask. It supplies the analogy we use for truncation boundaries, which act like a mask whose induced shift must be subtracted or bounded rather than ignored.
Research Program
The work has three parts. First, we build a shot-noise emulator for each benchmark Hamiltonian. We prepare the exact or high-accuracy tensor-network ground and first excited states at each $(a,\Lambda)$. We then sample measurement outcomes for each Pauli-string group with the true outcome distributions. This gives realistic $\sigma_i^2$ and lets us run thousands of synthetic experiments with known ground truth. We deliberately idealize state preparation so that the study isolates the measurement budget. Hardware noise would add a fourth error source and is out of scope.
Second, we build the design procedure. It spends a small pilot fraction of the budget at two or three cutoffs per spacing, fits the exponential truncation tail to estimate $b_i(\Lambda)$ with uncertainty, and estimates misspecification by comparing $a^2$ and $a^2{+}a^4$ ansätze. It then solves the constrained problem above for $(\Lambda_i, M_i)$ and, optionally, the set of spacings. Third, we add a certification test. It checks, with a stated false-alarm rate, whether the data at the chosen cutoffs are consistent with zero residual truncation bias relative to the final error bar. It uses a held-out higher-$\Lambda$ check run at a few points, and flags designs that should be redone.
Baselines
The naive reference point is uniform allocation. It uses the same spacings as the reference papers and the largest cutoff the qubit budget allows at every spacing, with shots split equally and an ordinary least-squares $a^2$ fit. This is what one does without thinking about the problem. It defines the cost to reach a given error that we must undercut.
The strongest existing method is the sequential pipeline of Jakobs et al. and Schwägerl et al., transplanted to the noisy setting. At each spacing, measure several cutoffs and extrapolate to infinite cutoff, then fit the continuum limit to those extrapolated points with propagated errors. It is the strongest competitor because it is what the field currently does to remove truncation bias, and it is unbiased in principle. Its deterministic reported results are $M_0 = 4.008(27)$ for single-plaquette SU(2) (Jakobs et al.) and $M_\pi/g = 0.2117(11)$ for twisted-mass Schwinger fermions (Schwägerl et al.). Those error bars come from variational and extrapolation error, not shots. No reported result was found on Lacuna for shot-budget-constrained continuum extrapolation on any of these systems, so the noisy sequential baseline is something we must run ourselves. We give the sequential pipeline its own optimal shot split within its structure, so it is not handicapped.
The key ablation is a textbook optimal design for the intercept, with the cutoff frozen at its maximum and shots placed by the Lagrange rule. Its spacings are also optimized. This isolates whether the gain comes from treating the cutoff as a variable or merely from good shot allocation.
All methods share the same Hamiltonians, the same emulator, the same candidate spacings and cutoffs, and the same total cost $\mathcal{C}_{\text{tot}}$ measured in Pauli-string shots weighted by circuit width. Each method gets the same pilot budget, counted against its total. Beating a baseline means a lower root-mean-squared continuum error at equal cost across 1,000 synthetic replicates, or equivalently equal error at lower cost. Confidence intervals must also cover the truth at the nominal rate. A gain that comes with undercoverage does not count.
Why not just…?
Why not just keep the cutoff large everywhere and only optimize shots? If truncation bias is already negligible at affordable cutoffs and the cost of extra cutoff states is small, this is exactly as good as our method. The fixed-maximum-cutoff ablation is built to detect that case. Mathis et al. show that Pauli-term counts and qubit numbers grow with the link dimension. Jakobs et al. note that electric-basis truncations need rapidly growing state counts near weak coupling. So in SU(2) and (2+1)D U(1) we expect a real cost penalty, but in the 1+1D Schwinger model the Kühn et al. result suggests the gap may be small. We will report it honestly per system rather than claim a universal gain.
Why not just extrapolate the cutoff away first at each spacing, as current papers do? This is the sequential baseline. Its weakness under noise is that an extrapolation in $\Lambda$ at each spacing amplifies that point's statistical error before the continuum fit amplifies it again. It also spends shots on high cutoffs whose bias contribution is already below noise. If our experiments show the sequential pipeline matching the joint design at equal cost, the joint formulation adds nothing, and we will say so.
Why not just run at the finest affordable spacing? A single spacing cannot separate physics from grid artifacts. The twisted-mass isospin splitting reported by Schwägerl et al. disappears only as $a \to 0$. We nonetheless include a finest-spacing-only design in the emulator to show its bias directly against known truth.
Why not just use a more constrained, theory-guided continuum fit? Regularized fits in the style of Allton et al. reduce variance, and they are complementary rather than competing. Any fit form can be plugged into our weights $w_i$. We will run the joint design with both plain and constrained ansätze to show the cutoff-allocation gain persists independently of the fit.
Proposed Experiments
The first benchmark is the single-plaquette SU(2) system of Jakobs et al. at the same couplings and cap radii. The target is to recover $M_0$ within the reported 0.027 uncertainty under a shot budget. We will report shots-to-target for each design. The second is the (2+1)D U(1) truncated Hamiltonian on the small lattices of Groß et al., measuring the plaquette expectation value. Here we ask whether a budgeted, certified truncation can bound the cutoff's share of the reported Lagrangian–Hamiltonian offset. A certified bound well below the offset would point the cause toward temporal discretization or scale setting. The third is the two-flavor Schwinger model with truncated link dimension, using twisted-mass fermions.
The decisive check is a discrimination experiment on the Schwinger model. We ask how many total shots each design needs to separate 0.213 from 0.198 at $5\sigma$ with nominal coverage, given that the ground truth from Schwägerl et al. is near 0.211. A result where the joint design reaches that separation at a several-fold smaller budget than the sequential pipeline, with coverage intact, supports the claim. Equal budgets refute it. The secondary check is calibration of the certification test: on replicates where we deliberately set cutoffs too low, it should flag bias at its stated rate.
References
- Toward scalable simulations of lattice gauge theories on quantum computers
- Quantum simulation of the Schwinger model: A study of feasibility
- Dynamics in hamiltonian lattice gauge theory: approaching the continuum limit with partitionings of SU(2)
- Fermion Discretization Effects in the Two-Flavor Lattice Schwinger Model: A Study with Matrix Product States
- Matching Lagrangian and Hamiltonian simulations in (2+1)-dimensional U(1) gauge theory
- Review on Quantum Computing for Lattice Field Theory
- Chiral and continuum extrapolation of partially-quenched lattice results
- Limits on non-Gaussianities from WMAP data