Physicists who study rare decays of heavy-quark particles often have only a handful of signal events. Belle's search for $B_s^0 \to \eta\eta$ found $2.7 \pm 2.5$ events. To claim anything from samples that small, they cannot lean on textbook large-sample formulas. Instead they run thousands to millions of simulated repeat experiments, called pseudo-experiments or toys, to check that their error bars and upper limits mean what they say. Each toy is a full fit of an amplitude model, a quantum-mechanical description of how intermediate resonances interfere across the decay's kinematic space. When the model's resonance masses and widths are left free, every fit step must recompute a normalization integral over millions of simulated decays. The result is a workload of order $10^{11}$ amplitude evaluations per optimizer step across an ensemble, which runs today in 64-bit arithmetic.
This note proposes running those amplitude and normalization sums in 16- or 8-bit arithmetic on AI accelerators, using stochastic rounding. Stochastic rounding rounds a number up or down at random, with probabilities chosen so the expected result equals the exact value. The proposal does not claim that low precision is merely "fast enough." Its testable claim is narrower. Under stochastic rounding, arithmetic error behaves like extra Monte Carlo noise: it has zero mean and is independent across simulated points and across toys. That lets a fixed compute budget buy more simulated points and more toys, and reach lower total error than full precision with fewer points. Deterministic rounding instead leaves a bias that is shared by every toy, and no ensemble size removes it.
Core Claim
Write the normalization integral as $I(\theta)=\frac{1}{M}\sum_{m=1}^{M} w_m(\theta)$. Here $w_m=|\mathcal{M}(\mathbf{x}_m;\theta)|^2\epsilon(\mathbf{x}_m)$ is the squared amplitude times detector efficiency at simulated point $\mathbf{x}_m$, and $M$ is the number of simulated points.
With unit roundoff $u$ (the relative spacing of the number format), stochastic rounding of each term and of the running sum gives an estimate $\hat I$ that remains unbiased. Its variance is approximately
$$\mathrm{Var}(\hat I)\approx \frac{\sigma_w^2 + c\,u^2\,\mathbb{E}[w^2]}{M} + V_{\text{acc}}(M,u),$$
where $\sigma_w^2$ is the natural spread of the weights, $c$ is an order-one constant from the per-term bound $\mathrm{Var}\le\delta^2/4$, and $V_{\text{acc}}$ is the variance added by a low-precision accumulator. Deterministic rounding instead adds a squared bias term $B^2(M,u)$ that does not shrink with $M$ or with the number of toys. In a naive low-precision accumulator this bias stagnates, because once the partial sum is large, new terms round away entirely.
For peaked amplitudes, $\sigma_w^2$ is comparable to $\mathbb{E}[w^2]$, so the rounding term is suppressed by $u^2$. A low-precision run that is $k$ times faster then wins at matched compute whenever $k$ exceeds roughly $1 + c\,u^2\,\mathbb{E}[w^2]/\sigma_w^2$ plus the accumulator penalty.
The novelty is a quantitative precision budget for statistical inference on scarce many-body data. It covers point estimates, interval coverage and tail significances, and it identifies which parts of the computation can and cannot be stochastically rounded. Two parts cannot. The first is Breit–Wigner denominators near narrow poles. The second is the curvature (Hessian) matrix that is inverted to get error bars.
Research Program
Three ingredients are borrowed from earlier work.
The first is the case for stochastic rounding itself. Deep Learning with Limited Numerical Precision showed that it lets 16-bit fixed-point training reach the full-precision result where round-to-nearest stalls. A Statistical Framework for Low-bitwidth Training of Deep Neural Networks explained why: averaging procedures tolerate zero-mean noise but not bias. Our setting differs from both. The object is a statistical estimator of a physical integral, not a network trained for accuracy, and the target is calibrated intervals.
The second is the two-copy design of Neural Quantum States in Mixed Precision. That work used 64-bit parameters with low-precision Monte Carlo sampling and found that computing gradients or the ill-conditioned reconfiguration matrix in 16 bits failed. We take that negative result as a constraint. Parameters, gradients and the Hessian inversion stay in 64 bits, and only the bulk sums over simulated and data events are reduced. That paper used deterministic IEEE rounding on discrete spin lattices. We study stochastic rounding over continuous phase space.
The third is double sampling from ZipML. Squaring one rounded amplitude is biased, so we compute $|\mathcal{M}|^2$ as $\mathrm{Re}[Q^{(1)}(\mathcal{M})\,\overline{Q^{(2)}(\mathcal{M})}]$ from two independent roundings, which is unbiased.
Supporting evidence comes from Climate Modelling in Low-Precision, which found that deterministic rounding causes conserved-quantity drift and stochastic rounding does not. That motivates the claim about shared bias across toys.
The system to be built is a GPU fitter that runs whole toy ensembles in batch. Each event's resonance denominator, $m_0^2 - s - i m_0\Gamma$, is computed in 32 bits, where $m_0$ is the resonance mass, $s$ is the event's invariant mass squared and $\Gamma$ is the width. The complex amplitude products and the normalization and data sums use 16-bit or 8-bit tensor arithmetic with stochastic rounding, with accumulator precision as a configurable option. Each toy draws independent rounding seeds.
The analysis has three parts. It derives the matched-compute crossover above. It bounds the bias in $\log \hat I$ that comes from Jensen's inequality, which is of order $\mathrm{Var}(\hat I)/2I^2$, multiplied by the event count $N$ in the likelihood. It also maps which width-to-mass ratios break 16-bit denominator evaluation.
The scope is deliberately narrow: fits with floating resonance-shape parameters plus large toy ensembles. With fixed shapes, the normalization reduces to a precomputed matrix of overlap integrals between amplitude components, and nothing here is needed.
Baselines
The naive reference is the current practice: everything in 64-bit arithmetic with the same simulated sample size. It defines ground truth for each pseudo-data model, and every other configuration is scored by its deviation from it in units of the statistical uncertainty.
The strongest competitor is standard mixed precision. Products are computed in 16 bits with round-to-nearest, and results go into a 32-bit accumulator. This is the default on current hardware, and it removes most accumulation stagnation without random numbers. It runs on the same simulated sample, hardware and toy count, with the same 64-bit optimizer. Beating it means one of two things. Either stochastic rounding matches its accuracy at higher throughput by using lower-precision accumulators, or it removes a bias shared across toys that this baseline retains. If neither happens, stochastic rounding is unnecessary and we will say so.
A second competitor is low-precision accumulation with pairwise or compensated summation. These deterministic methods cut accumulation error to logarithmic growth.
A third competitor is 64-bit arithmetic with fewer simulated points at matched wall-clock time. It isolates whether extra points beat extra precision.
A fourth competitor applies to cases with one or two floating shape parameters. The normalization integral is precomputed on a parameter grid and interpolated.
No reported result was found on Lacuna for low-precision amplitude fits or toy ensembles in heavy-flavor analyses. The nearest reported figures are adjacent. Solinas et al. report up to 3.5× sampling speedup on an H100 and about 2.3× total speedup for a 2D ResCNN, with "no measurable loss in the accuracy of the final ground-state energy". Those figures set the order of speedup we must at least match.
Two ablations isolate the key ingredients. The first applies stochastic rounding only in the accumulator, or only in amplitude evaluation. The second shares one rounding seed across all toys instead of drawing independent seeds. The shared-seed version should re-create the toy-common bias and directly tests the decorrelation claim.
Why not just…?
Why not just use FP32 everywhere? This is a real option. If 32-bit throughput on the target hardware is close to 16-bit tensor throughput for complex amplitude kernels, the gain disappears. We therefore report measured wall-clock time and energy per toy for each precision on the same GPU and do not assume a speedup. The Solinas results show that reduced-precision Monte Carlo sums can pay off in a physics workload, but that was 16-bit against 64-bit, so the 32-bit comparison has to be measured here.
Why not just use round-to-nearest with a 32-bit accumulator? This is the most serious threat. Per-term rounding error is already suppressed by $u^2$ relative to Monte Carlo spread, so deterministic mixed precision may be accurate enough. The proposal's distinctive prediction is that it leaves a small bias shared by all toys, sitting in the extreme tail quantiles that determine a 3σ or 5σ significance. The shared-seed ablation and the tail-quantile comparison settle this. If the shared bias is below the toy-ensemble error at $10^6$ toys, stochastic rounding is justified only for 8-bit-accumulator or integer-only hardware, the setting Is Integer Arithmetic Enough for Deep Learning Training? addresses for training.
Why not just precompute and interpolate the normalization integral? For one or two floating shape parameters this probably works and may beat everything else, which is why it is a baseline. Grid cost grows exponentially with the number of floating masses, widths and coupled-channel parameters. The experiments include a configuration with four or more floating shape parameters, where the grid becomes infeasible.
Why not just use fewer simulated points in FP64? The matched-compute experiment answers this directly. It is the central decisive check, not an aside.
Proposed Experiments
Real collaboration data are not public, so all tests use pseudo-data generated from published models, with the published numbers as targets.
The high-statistics anchor is the CLEO-c isobar model of Dalitz plot analysis of the D+ → K− π+ π+ decay. That analysis has 67,086 candidates, about 1.1% background, and a Kπ S-wave carrying 50–60% of the rate. It tests whether interference fractions are reproduced when shapes float.
The scarce-data regimes reproduce four published results:
- Recoil-mass fit. Belle's Ξc measurement (Measurements of the branching fractions B(B−→Λ¯c−Ξc′0) , B(B−→Λ¯c−Ξc(2645)0) and B(B−→Λ¯c−Ξc(2790)0)) reports $\mathcal{B}(B^-\to\bar\Lambda_c^-\Xi_c(2790)^0)=(1.1\pm0.4\pm0.2)\times10^{-3}$ at 3.1σ and a 1.7σ $\Xi_c'$ signal with a 90% credibility limit below $6.5\times10^{-4}$. Narrow peaks here stress the denominator precision.
- Upper-limit search. The $B_s^0\to\eta\eta$ search found 2.7 events and set a limit of $<6.5\times10^{-5}$.
- Floating broad structure. $B^+\to K^+K^-\pi^+$ has 429 ± 43 events and a broad enhancement near 1.5 GeV/c² that serves as a floating-shape component.
- Threshold behaviour. The three-body $\chi_{cJ}\to\Lambda\bar\Lambda\phi$ analysis reported no threshold enhancement. Pseudo-data with and without an injected enhancement test whether low precision changes that qualitative conclusion.
Metrics are as follows. Parameter bias and RMS are measured relative to 64-bit results, in units of statistical error. Coverage of 68% and 90% intervals is measured over toy ensembles. Toy-derived significances and upper limits are measured against their 64-bit values. Total normalization error is measured against an independent high-statistics 64-bit integral, and energy is reported per toy.
There are four decisive checks.
- Matched compute. At equal wall-clock time, does 16-bit or 8-bit stochastic rounding with $k$ times more simulated points give lower total integral error and better interval coverage than 64-bit with $M$ points? The crossover predicted above is tested quantitatively.
- Shared bias across toys. As the toy count grows from $10^3$ to $10^6$, does the mean shift in the likelihood-ratio statistic vanish under independent-seed stochastic rounding but persist under deterministic rounding and shared seeds?
- Narrow-pole stress. Scanning $\Gamma/m_0$ downward should locate where 16-bit denominators fail and confirm that the 32-bit-denominator design fixes it.
- Curvature precision. Repeating the Solinas finding in this setting, we test whether a reduced-precision Hessian corrupts error bars while a 64-bit Hessian on top of stochastically rounded sums does not.
A clean negative outcome in check 2 is a publishable answer too. It would mean 32-bit-accumulator mixed precision is sufficient for heavy-flavor toy studies.
References
- Deep Learning with Limited Numerical Precision
- A Statistical Framework for Low-bitwidth Training of Deep Neural Networks
- Neural Quantum States in Mixed Precision
- ZipML: Training Linear Models with End-to-End Low Precision, and a Little Bit of Deep Learning
- Climate Modelling in Low-Precision: Effects of both Deterministic & Stochastic Rounding
- Is Integer Arithmetic Enough for Deep Learning Training?
- Dalitz plot analysis of the D+ ---> K- pi+ pi+ decay
- Measurements of the branching fractions B(B−→Λ¯c−Ξc′0) , B(B−→Λ¯c−Ξc(2645)0) and B(B−→Λ¯c−Ξc(2790)0)
- Search for the decay <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msubsup><mml:mi>B</mml:mi><mml:mi>s</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:mo stretchy="false">→</mml:mo><mml:mi>η</mml:mi><mml:mi>η</mml:mi></mml:math>
- Observation of the Decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:math>
- Study of the decays <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>χ</mml:mi><mml:mrow><mml:mi>c</mml:mi><mml:mi>J</mml:mi></mml:mrow></mml:msub><mml:mo stretchy="false">→</mml:mo><mml:mi mathvariant="normal">Λ</mml:mi><mml:mover accent="true"><mml:mi mathvariant="normal">Λ</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mi>ϕ</mml:mi></mml:math>