Plasma physicists often need to answer a narrow question: among a few dozen candidate laser and target settings, which one gives the best outcome, such as the most energetic and narrowest-spectrum proton beam? Each candidate is scored by running a particle-in-cell (PIC) simulation, which tracks millions of charged "macro-particles" through self-consistent electromagnetic fields. Each run is expensive and returns a noisy answer, because the particles are placed randomly at the start. Cheaper, lower-resolution runs exist, but they can be badly wrong, and sometimes they miss the physics that matters. We propose an algorithm that decides which run to launch next and when to stop. It declares a winner with a guaranteed error probability of at most $\delta$, and its total compute cost is provably close to the minimum any method could achieve.
The key observation is that a cheap run and an expensive run of the same setting can share a random seed, meaning the same initial particle draw and the same noise realization. Their outputs are then correlated. The cheap run becomes a control variate, a correlated side measurement that cancels part of the expensive run's noise, rather than a biased stand-in for it. This changes the theory of multi-fidelity best-arm identification in a basic way. It removes the need to know in advance how wrong the cheap simulator is. It also predicts that the optimal policy mixes fidelities per candidate, where existing theory says the optimal policy concentrates on a single fidelity.
Core Claim
Best-arm identification (BAI) is the bandit problem of finding the best of $K$ options, called "arms," using as few noisy evaluations as possible. In its fixed-confidence version, it has a sharp theory. Optimal Best Arm Identification with Fixed Confidence proves that expected sample count must grow as $T^*(\mu)\log(1/\delta)$, where $T^*(\mu)$ is an instance-dependent "characteristic time." It gives Track-and-Stop, which matches that bound: the algorithm samples toward the optimal proportions computed from current estimates and stops with a likelihood-ratio test.
Optimal Multi-Fidelity Best-Arm Identification extends this to settings where cheaper, biased fidelity levels exist. Its algorithm, MF-GRAD, is cost-optimal as $\delta \to 0$. It also proves "fidelity sparsity": each arm's optimal budget goes to a single fidelity level. Two assumptions underlie that result. Observations at different fidelities are independent, and each fidelity's bias from the top level is bounded by a known constant.
Both assumptions fail for PIC. The research file's Gap 4 names the danger: a reduced-mass-ratio or fluid model can suppress a micro-instability entirely, as in the lower-hybrid-drift saturation mechanism of Kinetic simulations of anomalous resistivity in high-temperature current carrying plasmas. No finite bias bound then holds, and trusting the cheap level is fatal.
Our claim is that seed-paired evaluation removes the bias problem entirely. The target is always the high-fidelity mean $\mu_k^H$ of arm $k$. Low-fidelity outputs enter only through their estimated correlation $\rho_k$ with high-fidelity outputs. Correctness therefore never depends on the low fidelity being accurate, only on its being measured. If a low-fidelity model misses the physics, its correlation with the high-fidelity output is near zero and the algorithm stops paying for it.
For a Gaussian observation model, the per-arm effect of fidelity choice can be isolated. Suppose arm $k$ receives $n$ paired runs, each costing $c_H + c_L$ where $c_H$ and $c_L$ are the high- and low-fidelity costs, plus $m$ extra low-only runs. The control-variate estimator of $\mu_k^H$ then has variance
$$ V_k(n,m) = \sigma_k^2\left[\frac{1-\rho_k^2}{n} + \frac{\rho_k^2}{n+m}\right]. $$
Here $\sigma_k^2$ is the high-fidelity noise variance. The formula is the standard variance of a regression estimator whose control has unknown mean. Multiplying $V_k$ by the cost spent gives a cost-weighted variance $v_k$, and the arm-local problem is to minimize it over the ratio $m/n$ and over the high-only option:
$$ v_k = \min\Big\{\, c_H\sigma_k^2,\ \min_{r \ge 0}\big(c_H + (1+r)c_L\big)\,\sigma_k^2\Big[(1-\rho_k^2) + \tfrac{\rho_k^2}{1+r}\Big] \Big\}. $$
Here $r = m/n$ is the ratio of extra low-only runs to paired runs.
We conjecture, and will prove, that the multi-fidelity lower bound then reduces exactly to Gaussian BAI with arm-wise effective variance $v_k$. The global allocation is found by the Track-and-Stop max-min problem over arms. The within-arm fidelity mix has the closed form above. The resulting optimum is generally not sparse: whenever $\rho_k^2 > c_L/(c_H + c_L)$ it uses two modes at once, paired runs and extra cheap runs. That is a direct, testable contradiction of the independent-observation sparsity theorem, and it resolves the Gap 4 tension. Sparsity is an artifact of independence, not a property of multi-fidelity identification.
Research Program
The theory work has three parts. First, prove the lower bound for $K$ arms, each with jointly Gaussian (high, low) observations under shared seeds, with unknown low-fidelity mean treated as a nuisance parameter, and show the separation into arm-wise $v_k$. Second, extend it to $M > 2$ ordered fidelities with a full per-arm covariance, where the within-arm problem becomes a small convex program. Third, design Paired Track-and-Stop. This algorithm tracks the plug-in optimal cost allocation, chooses the sampling mode inside each arm using estimated $(\sigma_k, \rho_k)$, and forces a minimum number of paired runs so that $\rho_k$ stays consistently estimated. It stops with a generalized likelihood-ratio test on the control-variate estimates.
Unknown covariances make exact finite-$\delta$ thresholds delicate. We will use a conservative self-normalized threshold and verify error rates empirically instead of claiming more than we prove. We restrict to finite candidate sets, 16 to 25 configurations. Continuous design, the file's Gap 2, is out of scope.
The engineering work is to make seed pairing real inside a PIC code. The PIC workflow is: deposit particle charge onto a grid, solve Maxwell's equations, and push particles with the Lorentz force. Pairing lives in the first step. We will define three pairings and measure the correlation each produces:
- Particle-count fidelity. The low-fidelity run uses a subsample of the high-fidelity run's initial macro-particles, taking advantage of the $1/\sqrt{N_{ppc}}$ shot-noise scaling, where $N_{ppc}$ is the number of particles per grid cell.
- Resolution fidelity. The same initial particles are run on a coarser grid with a larger time step.
- Mass-ratio fidelity. Ion and electron initial positions are shared across mass ratios $m_i/m_e = 184$ and $1836$.
How strongly these correlations persist through nonlinear, chaotic dynamics is an empirical question, and measuring it is itself a deliverable.
Baselines
The naive reference point is high-fidelity-only Track-and-Stop Optimal Best Arm Identification with Fixed Confidence, with Gaussian arms and estimated variances. It is the correct answer if cheap runs carry no usable information. Our method must beat it on total cost whenever the measured $\rho_k$ is large, and must never lose by more than the forced-pairing overhead when $\rho_k$ is small. The file reports that on a 5-arm synthetic benchmark Track-and-Stop needed about 4,000 draws versus over 8,000 for KL-LUCB. This anchors the claim that optimal tracking matters at all.
The strongest existing method is MF-GRAD Optimal Multi-Fidelity Best-Arm Identification, run on the same PIC fidelity levels with independent seeds. Its required bias bounds will be set from a pilot study whose cost is charged to MF-GRAD, and tuned from a small grid of inflation factors under the same budget we use to tune our forced-pairing rate. On its designed "Hard" 5×2 synthetic instance, MF-GRAD is reported to be "nearly 3x more cost-effective than the next best baseline." Beating it means lower median cost-to-stop at matched empirical error on the PIC testbeds. Where our bias-bound-free guarantee matters, it also means MF-GRAD's error rate exceeding $\delta$ when bias bounds are set too tight. On MF-GRAD's own synthetic instances with $\rho = 0$, we expect to tie, and we will report that.
The decisive ablation is Paired Track-and-Stop with seeds deliberately decorrelated, so $\rho_k$ is forced to about 0. This isolates pairing as the source of any gain. A second ablation fixes the within-arm ratio $r$ at a single global value, which tests whether per-arm adaptation of the fidelity mix matters.
We also run a pilot-then-fixed control-variate scheme. It estimates $\rho$ once from pilot runs, fixes the allocation, and then runs Track-and-Stop. Successive Halving and a multi-fidelity Gaussian-process Bayesian optimizer serve as practitioner references. All methods share the same simulator, seeds pool, arm set, cost accounting in core-hours, and 200 independent replications per $\delta$ value.
Why not just…?
Why not just run MF-GRAD on the PIC fidelity ladder? MF-GRAD is the closest prior work and its optimality proof is sound under its assumptions. The problem is that it needs a known bias bound per fidelity. It also treats cross-fidelity observations as independent, so it cannot exploit shared seeds. If the true high-to-low correlation is near zero, MF-GRAD with honest bias bounds and our method should coincide. The seed-decorrelation ablation and the synthetic replication will show this directly. Where a cheap model misses an instability, MF-GRAD either wastes budget because its bounds are inflated or errs because its bounds are too tight. The mass-ratio testbed is built to expose exactly this.
Why not just estimate the correlation once and run a fixed control-variate allocation? This is a real threat. If $\rho$ is roughly the same across candidates, the pilot-then-fixed baseline will capture most of the gain with far less machinery, and we will say so. The adaptive method should only win when $\rho_k$ varies across arms, for instance when some laser-target matchings drive the foil into a regime where the coarse grid diverges from the fine one. We will report the spread of measured $\rho_k$ per testbed and the gap between the two methods as a function of that spread.
Why not just use multi-fidelity Bayesian optimization? It handles fidelity naturally but offers no stopping rule with a $\delta$ error guarantee. Its optimism-based acquisition also over-samples noisy arms, as argued in Variance Driven Exploration, which proves that optimism wastes samples under heteroscedastic noise. We include it with a posterior-probability stopping heuristic and report its empirical error rate alongside its cost.
Why not just train a surrogate like Diff-PIC and search it? Diff-PIC generates electric-field snapshots conditioned on physical parameters with roughly 1–2% error in-distribution. However, it needed 6,615 OSIRIS runs to train, and its error grows 10–20% outside the training range. A deterministic surrogate is also uncorrelated with seed noise, so in our framework it could only enter as a biased low fidelity. That is exactly the case pairing is designed to avoid. Our method targets the new-regime setting, where no such training corpus exists.
Proposed Experiments
The first testbed is a synthetic sanity check. We replicate MF-GRAD's 4×5 and Hard 5×2 Gaussian instances, then add correlated variants with $\rho \in \{0, 0.5, 0.8, 0.95\}$. We measure cost-to-stop relative to the computed lower bound $T^*\log(1/\delta)$ and check empirical error against $\delta \in \{0.1, 0.01, 0.001\}$. The prediction is non-sparse optimal allocations exactly when $\rho^2 > c_L/(c_H + c_L)$.
The second testbed is 1D PIC radiation-pressure acceleration, following the sharp-front laser-foil setup of Efficient acceleration of monoenergetic proton beam by sharp front laser pulse. The arms are a 5×5 grid over front steepness $a_0/t_{up}$ and areal density $n_0 l_0$, and the score is the peak proton energy within a spectral-width cut. Low fidelity uses subsampled particles and a coarser grid. Ground truth comes from 500 high-fidelity seeds per arm. No reported result was found on Lacuna for any bandit method on this task.
The third testbed is a reduced 2D current-sheet lower-hybrid-drift scan in the style of Kinetic simulations of anomalous resistivity in high-temperature current carrying plasmas, with $m_i/m_e = 184$ as the cheap level and $1836$ as the target. Arms are drift and temperature-ratio settings, and the score is the saturated resistivity. Here some arms should show collapsed correlation. No reported result was found on Lacuna for fixed-confidence identification on this scan.
The decisive checks are four:
- Empirical error must stay at or below $\delta$ on every testbed, including arms where the cheap model misses the instability.
- On the radiation-pressure testbed, cost-to-stop must fall below both high-fidelity-only Track-and-Stop and MF-GRAD whenever the measured $\rho$ exceeds the predicted threshold.
- The decorrelated ablation must erase the gain.
- The chosen allocations must match the closed-form mix predicted from measured $(\sigma_k, \rho_k)$.
If the measured PIC correlations turn out uniformly low, that negative result still settles which cheap PIC runs are worth anything for design decisions, which the field currently decides by intuition.
References
- Optimal Best Arm Identification with Fixed Confidence
- Optimal Multi-Fidelity Best-Arm Identification
- Kinetic simulations of anomalous resistivity in high-temperature current carrying plasmas
- Variance Driven Exploration: A Provable and Efficient Methodology for Pure Exploration in Highly Stochastic Environments
- Diff-PIC: Revolutionizing Particle-In-Cell Nuclear Fusion Simulation with Diffusion Models
- Efficient acceleration of monoenergetic proton beam by sharp front laser pulse