Flavor-Breaking Vacua as a Precision Test for Clock-Based State Preparation

Many quantum algorithms for physics start by preparing the lowest-energy state (the "vacuum") of a model on a quantum computer and then measuring it. Simulations of particle physics often study models with several fermion types, called flavors, that are nearly interchangeable. Real flavors have slightly different masses, so that interchangeability, flavor symmetry, is slightly broken. The physics of interest is usually in the small effects this breaking produces. This note argues that preparing flavor-broken vacua has a specific, measurable difficulty. A vacuum with exact flavor symmetry is easy to prepare. Its overlap with the broken vacuum, however, shrinks exponentially with lattice size at any fixed mass splitting, and the observables of interest are only as large as the splitting. Past a crossover lattice size, one must therefore drive the system from the symmetric vacuum to the broken one with a time-dependent Hamiltonian, and that drive must be simulated to an accuracy set by the small splitting.

High-accuracy simulation of time-dependent Hamiltonians is what "clock" constructions were built for. A clock construction adds a register that stores time as a quantum variable, which turns a time-dependent problem into a fixed, time-independent one. So far these methods have been tested only on toy spin chains. We propose to measure this flavor-breaking overlap decay in the two-flavor lattice Schwinger model, a one-dimensional analogue of quantum chromodynamics (QCD). We will then find the crossover lattice size beyond which ramping is required, and count gates for clock-based ramps against the best plain alternatives. The goal is to settle whether clock methods buy anything in the regime where small flavor-breaking signals must be resolved.

Core Claim

Let $|\Omega_0\rangle$ be the vacuum of the two-flavor model at equal masses, and let $|\Omega_\delta\rangle$ be the vacuum with mass splitting $\delta = m_d - m_u$ on a lattice of $N$ sites. Second-order perturbation theory gives the overlap

$$\log |\langle \Omega_0 | \Omega_\delta \rangle|^2 \approx - N \chi_F \, \delta^2 .$$

Here $\chi_F$ is the fidelity susceptibility per site, meaning the per-site sum of squared matrix elements of the isospin-breaking mass density, which takes the vacuum to isospin-one excited states, each divided by the squared excitation energy. The form follows from the fact that the perturbation is a sum of local terms, so the log-overlap is extensive in $N$.

Two consequences follow. Using $|\Omega_0\rangle$ directly as the starting state for phase estimation costs on the order of $e^{N\chi_F\delta^2}$ repetitions. Beyond roughly $N_* \sim 1/(\chi_F \delta^2)$ sites, a ramp in $\delta$ is unavoidable. Moreover, $\chi_F$ grows as the isospin-one gap shrinks. That happens near topological angle $\theta \to \pi$, where flavor breaking acts non-perturbatively. In that region $N_*$ becomes small even for modest splittings.

Along the ramp, the state error must stay well below the size of the isospin-breaking observables, which are themselves of order $\delta$. Our hypothesis is that in this small-$\delta$, large-$N$ regime the needed accuracy favors clock-based methods with polylogarithmic error scaling over product formulas, provided both are compared at equal gate cost. This is a falsifiable claim about resource counts, not a claim that clocks are always better.

Where This Sits

Three lines of prior work motivate the problem but stop short of it.

The multi-flavor Schwinger studies by Kühn, Funcke and coauthors prepare vacua with symmetry-restricted variational circuits (multi-flavor phase structure; three-flavor phase diagram with chemical potential). These variational circuits are the variational quantum eigensolver (VQE): a parameterized circuit tuned by a classical optimizer to minimize energy. They reach overlaps above 95% for up to 6 sites. However, their parameter reduction from $6N-1$ to $3N$ per layer relies on symmetries that a mass splitting destroys, and they report optimizer runs that land in unphysical states.

The MPS study of fermion discretizations in the two-flavor Schwinger model computes accurate vacua classically and shows that some lattice fermion formulations add their own artificial isospin splitting. That work supplies our ground truth and our classical tools. MPS here means matrix product states, a compressed classical representation of one-dimensional quantum states, optimized with the DMRG algorithm.

Kühn, Cirac and Bañuls show that a slow unitary sweep reaches the single-flavor vacuum at overlap above 0.99 with total time $T \ge 60$, independent of size up to 50 sites. That result is the reference our ramps must match or beat. It does not address flavor splittings.

The clock methods supply the technique. Watkins et al.'s discrete clock constructions store time bins in qubits and start the clock in a Gaussian wavepacket to suppress edge errors. This enables time-dependent qubitization, a near-optimal simulation method whose cost grows only polylogarithmically with inverse error. It also enables time-dependent multiproduct formulas, which combine several product-formula runs to cancel their errors and keep favorable commutator scaling. On spin chains these reach errors "orders of magnitude lower" than plain Trotter steps at equal cost. Cao, Shi et al.'s Sambe–Howland treatment derives high-order product formulas for digitized adiabatic sweeps. Neither has been run on a fermionic gauge model, and neither asks when the extra precision actually matters for preparing a state.

One framing in the research file we reject on reflection. Coupling a clock to a gauge theory does not by itself threaten Gauss's law, the local constraint that physical states of a gauge theory must satisfy. Each $H(t_k)$ commutes with the Gauss-law generators, and the clock-shift term acts only on the clock, so the enlarged Hamiltonian preserves the physical sector exactly. Our contribution is therefore the precision question, not a gauge-invariant clock.

Research Program

The first stage is classical. Using 2-site DMRG in ITensors, as in the discretization study, we will compute $|\langle\Omega_0|\Omega_\delta\rangle|^2$ for $N$ up to about 100. We will cover a grid of splittings $\delta/m$, angles $\theta \in [0,\pi]$, and staggered and Wilson discretizations. We will fit $\chi_F(\theta)$ and test the extensive form of the equation above. We will also locate $N_*$ for each setting. This stage also checks the physics of the claim: if $\chi_F$ turns out small enough that $N_*$ exceeds any foreseeable device size, the ramp is unnecessary and the proposal reduces to a negative result, which we would report.

The second stage builds ramps $H(t) = H(\delta = 0) + \delta\, s(t/T)\, O_3$. Here $O_3$ is the up-minus-down mass operator and $s$ is a smooth schedule rising from 0 to 1 over total time $T$. Each ramp is compiled four ways: product formulas with terms grouped so that every factor respects isospin symmetry (described below), the discrete-clock multiproduct formula, discrete-clock time-dependent qubitization, and the Sambe–Howland high-order product formulas.

For each compilation and each target splitting we will find the cheapest settings that reach a fixed final infidelity. We will then measure isospin-breaking observables, namely the condensate difference $\langle \bar\psi_u\psi_u - \bar\psi_d\psi_d\rangle$ and the splitting between charged and neutral pions, to a relative error of 10% of their DMRG values.

Costs are reported as two-qubit gate counts and, for fault-tolerant variants, non-Clifford gate counts, which are the expensive gates on error-corrected hardware. We include the clock-register qubits in all counts. We will also study the clock register's own starting state, comparing Gaussian against flat wavepackets and varying their width. This is the knob Watkins et al. fixed without optimizing.

The third stage goes to (2+1) dimensions: two-flavor Wilson lattice QED$_3$ with truncated links (Hamiltonian QED$_3$ with Wilson fermions) on $2\times2$ lattices by exact diagonalization. The purpose is to check whether $\chi_F$ and the cost ranking carry over to the setting where classical tensor methods stop scaling.

Baselines

The naive reference point uses the symmetric vacuum $|\Omega_0\rangle$, prepared by the size-independent sweep of Kühn, Cirac and Bañuls, directly as the guiding state for phase estimation or the rodeo algorithm, an energy filter that projects onto eigenstates by averaging evolutions over random times. No ramp is used. Its cost is the preparation cost times $1/|\langle\Omega_0|\Omega_\delta\rangle|^2$, computed from stage one. A ramp method beats this baseline only for $N > N_*$, and we will report the measured crossover rather than assume it. The paper reports overlap above 0.99 for $T \ge 60$ at $N = 50$ in the single-flavor model. No reported result was found on Lacuna for the two-flavor broken case.

The strongest existing competitor is a product formula built so that it does not break isospin. Up-quark and down-quark hopping on the same link commute, so the even-link and odd-link groups, the mass term and the long-range charge term are each isospin-invariant. Their Trotter errors are therefore invariant too. This baseline removes any spurious flavor breaking from the integrator, needs no clock qubits, and is what a careful practitioner would actually run. It receives the same schedules, the same target infidelity and a search over step count and Trotter order (2 and 4) of the same size as the clock methods' parameter search. Clock methods beat it only if they reach the target at lower total gate count, clock qubits included, at the splittings where $N > N_*$.

The variational baseline is the symmetry-restricted ansatz of Kühn et al., extended with unrestricted parameters for $\delta \neq 0$. Its reported reference number is overlap exceeding 95% for $N = 2, 4, 6$. It is run at the same sizes with a fixed budget of optimizer restarts, and we count how often it lands in unphysical states.

Two ablations isolate the key ingredient, which is high precision in the ramp simulation. The first compares ungrouped against isospin-grouped Trotter, which separates spurious-symmetry effects from raw accuracy. The second compares clock multiproduct formulas with and without the Gaussian clock wavepacket, which isolates boundary errors.

Why not just…?

Why not just use the symmetric vacuum and let phase estimation do the work? For small lattices or tiny splittings this is the right answer, and the proposal is built to say so. The equation above predicts exponential cost in $N\delta^2$. Stage one measures $\chi_F$ directly and turns this into a concrete crossover $N_*$. If $N_*$ is large for physically reasonable splittings away from $\theta \approx \pi$, we will restrict our claims to the near-$\pi$ region where $\chi_F$ is large.

Why not just use the isospin-grouped Trotter ramp? This is the real threat. Because grouping makes Trotter errors isospin-invariant, the clock methods' only remaining advantage is error scaling. That may not pay for the clock qubits and controlled operations at the modest accuracies preparation needs; Watkins et al. themselves note substantial qubit overhead. The stage-two cost curves settle this. If grouped Trotter matches clock methods at the 10%-of-signal accuracy target, we report that clock constructions are unnecessary for this task.

Why not just compute flavor-breaking effects as linear response around the symmetric vacuum? For observables that depend linearly on small $\delta$, perturbation theory around $|\Omega_0\rangle$ needs no broken-vacuum preparation and is cheaper. The proposal targets splittings comparable to the mass, for example $m_u/m_d \approx 0.5$, and angles near $\pi$, where DMRG can show directly whether the response is nonlinear. We will report the $\delta$ at which first-order predictions deviate by more than 10% and confine claims to beyond that point.

Why not just stay classical with MPS? In one dimension MPS handles this model, as the discretization study shows, with twisted-mass pion mass 0.2117(11). The Schwinger model is our controlled testbed. The transferable results are $\chi_F$ and the cost ranking, which stage three checks in QED$_3$, where classical study is confined to $4\times4$ lattices.

Proposed Experiments

The first decisive check tests whether the log-overlap scales linearly in $N$ and quadratically in $\delta$ across the DMRG grid, and whether $\chi_F$ rises sharply toward $\theta = \pi$. Failure of the extensive form would undercut the motivation.

The second check produces gate-count curves for all four ramp compilations against the naive baseline, as functions of $N$ and $\delta$, at fixed final infidelity and fixed observable accuracy. The outcome supporting the proposal is a region at $N > N_*$ where discrete-clock multiproduct formulas or qubitization use fewer total gates than grouped Trotter, including clock overhead. The outcome refuting it is grouped Trotter dominating everywhere.

The third check repeats the comparison under simulated depolarizing noise for the near-term variants. Here we test whether the higher precision of clock methods survives their larger qubit count.

The fourth check verifies on Wilson QED$_3$ at $2\times2$ that $\chi_F$ and the cost ranking qualitatively persist, before any claim is extended beyond one dimension.