Spending Bits and Exact Arithmetic Only Where Rare Events Live

Many rare-event searches spread their detection across several devices. Examples are subdetectors in a collider trigger, sensor arrays, and observatories on different continents, like gravitational-wave and neutrino observatories combining weak alerts. Each device sees evidence that is too weak to claim a discovery alone. The discovery only appears when that weak evidence is combined. Devices are limited in two ways. They can send only so many bits to a central combiner, and they have tight power and time budgets for computing their local evidence. The usual tools answer these limits with the wrong goal. Compression schemes minimize average reconstruction error. Approximate computing, meaning hardware that saves energy by allowing small arithmetic errors, keeps average accuracy close to exact. Rare-event detection cares about neither average. Its outcome depends almost entirely on a thin upper tail of candidates whose combined score lands near the decision threshold. Everything else is background whose exact value does not matter.

We propose a two-tier scheme for each device, which we call "score-then-refine." Every candidate is first scored cheaply on approximate, low-voltage or low-precision hardware and reported with one or two bits. Only candidates whose rough score lands inside a calibrated escalation band are recomputed exactly and sent at full precision. The band is set wide enough, using the measured error distribution of the approximate hardware, that a true signal is rarely pushed out of it. The claim is that this recovers nearly the discovery significance of a centralized, bit-exact analysis for a small fraction of the bits and arithmetic energy. It also avoids two failures that the parts suffer when used separately: signals lost to compression, and spurious tails created by approximate hardware.

Core Claim

For distributed rare-event search, a single budget should govern both communication and computational precision. That budget is the probability that a value can change the global detection decision at a fixed background false-alarm rate. Under this objective, the best allocation is extremely lopsided. The best-known distributed compression methods are designed for mean-squared error, not for a tail-dominated test, and they are the wrong tool here. Randomized sparsification keeps random coordinates. Server-side temporal fill-in, as in Rand-k-Temporal from Leveraging Spatial and Temporal Correlations in Sparsified Mean Estimation, fills missing coordinates from history, which tends to smooth away a transient spike. Uniform quantization spends equal effort on background and signal.

The method centres on one escalation rule. Let $s$ be a device's exact local score for a candidate, for instance a log-likelihood ratio or anomaly score. Let $\hat s$ be the score computed on approximate hardware, and let $\tau$ be the lowest local score that could still matter for the global test given the other devices. The device escalates, meaning it recomputes exactly and sends at full precision, when

$$\hat s \ge \tau - m, \qquad m = \min\{m' : \Pr(s \ge \tau,\ \hat s < \tau - m') \le \delta\},$$

where $m$ is the escalation margin and $\delta$ is the tolerated miss rate. The margin is estimated as an empirical quantile from held-out candidates run through the hardware's error model. The hardware adds random errors that can push a score in either direction. The exact recompute removes upward errors, which are background falsely lifted into the tail. The margin controls downward errors, which are signal pushed out of the tail.

This design responds to the heavy-tailed error that approximate arithmetic produces. That tail risk is exactly why physicists distrust approximate computing in discovery pipelines.

Relation to Prior Work

Three papers motivate the gap. Enhancing the significance of astrophysical events with multimessenger coincidences combines compact per-observatory likelihoods into an exact Bayesian update. It shows the gain is dominated by rare high-value triggers: one coincident 100 TeV neutrino lifts the probability of astrophysical origin from 7.7% to 86%. It assumes lossless, exact messages. We keep its combination rule but ask how few bits and how much inexact arithmetic it can tolerate.

Intelligent experiments through real-time AI runs a graph-network trigger for sPHENIX on a field-programmable gate array (FPGA). It meets a sub-10 μs deadline with static fixed-point quantization. It is centralized and makes no use of the fact that most inputs are background.

Distributed Expectation Propagation for Multi-Object Tracking over Sensor Networks saves bandwidth by exchanging only means and covariances. Every sensor transmits uniformly every round, regardless of what it sees.

Three papers constrain the design. Communication lower bounds for statistical estimation problems via a distributed data processing inequality proves that estimating even sparse high-dimensional signals needs communication that grows with dimension. We sidestep rather than contradict this result. Each device first reduces its data to a per-candidate score, turning estimation into detection. Our budget then scales with the number of escalated candidates, not with dimension. The spatial estimator of Leveraging Spatial and Temporal Correlations in Sparsified Mean Estimation is brittle when devices see differently distributed data. Heterogeneous subdetectors are exactly that case, so we do not rely on correlation between devices. The outlier-propagation problem identified in approximate-computing surveys (Approximate Computing Survey, Part I) is the failure our exact recompute is built to remove.

Three papers supply methods. AxTrain provides noise-aware training and low-voltage memory bit-flip noise models for the cheap scorer. AXNet provides the idea of a learned "safe to approximate" decision. We replace its learned predictor with a calibrated margin, because a rare-event search needs a stated miss rate. A learned gate offers no such guarantee. Error-Efficient Computing Systems supplies the principle of protecting only critical operations. We apply that principle at the level of candidates rather than bits.

Research Program

We will build three things.

The first is a simulator of distributed detection. Candidates are split across $K$ devices. Each device computes a score from its own subset of detector information on emulated approximate hardware. This hardware has AxTrain-style low-voltage bit-flip noise and inexact multipliers at several aggressiveness levels, each mapped to an estimated energy saving. A central combiner then runs the global test.

The second is the score-then-refine protocol. It has a 1–2 bit coarse report, an escalation band with the calibrated margin $m$, and full-precision escalated messages. The threshold $\tau$ is broadcast by the combiner. It is either static or updated per time window from the other devices' coarse reports.

The third is an analysis of the regime where the scheme pays off. The expected cost per candidate is roughly the cost of an approximate score, plus the escalation rate times the cost of an exact recompute and a full-precision message. The escalation rate is the fraction of background whose approximate score exceeds $\tau - m$. This fraction grows with hardware aggressiveness through $m$, so there is an interior optimum that we will locate empirically.

We will also state the case where the method is not worth it. At low candidate rates, as in the multimessenger setting, sending exact scalar likelihoods costs almost nothing. The contribution then reduces to saving energy on the approximate scorer.

Baselines

The naive reference is local triggering. Each device applies its own threshold and sends one bit, or nothing, to the combiner. This is standard practice and the cheapest possible scheme. It is expected to lose subthreshold coincidences. We run it with thresholds chosen, on the same validation data, to maximize global significance.

The upper bound is centralized bit-exact analysis. All device information goes to one combiner with 32-bit arithmetic. On CMS Open Data this corresponds to the CATHODE result in Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data. That paper reports a dimuon-resonance significance of 6.4σ against 1.6σ for the raw spectrum. We will reproduce this pipeline and treat our own reproduced number as the ceiling.

The strongest communication competitor is exact scalar-score transmission with uniform b-bit quantization, swept over b. This is the Veske-style message made cheaper. We pair it with magnitude-based top-k selection of candidates per window. Both use the same exact-arithmetic scorer and the same total bit budget.

The strongest compression-estimator competitor is Rand-k-Spatial and Rand-k-Temporal from Leveraging Spatial and Temporal Correlations in Sparsified Mean Estimation. These are applied to device feature vectors at matched bits.

The strongest approximate-computing competitor is an AxTrain-trained scorer with no escalation, which sends its approximate score at matched bits. We also run a post-training fixed-point quantized scorer in the manner of hls4ml, as in the sPHENIX work.

Two ablations isolate the key ingredients. The first keeps the band but removes the exact recompute, so escalated candidates send their approximate score at full precision. The second keeps the recompute but sets the margin to zero.

Every method gets the same training data, the same scorer architecture, and the same hyperparameter search budget, and every method is evaluated at a fixed background false-alarm rate. The proposal beats the field if it reaches at least 90% of the centralized significance at no more than one-fifth of the bits and arithmetic energy of the best matched-budget baseline. Separately, it must be the only method whose realized miss rate on injected signals stays within $\delta$ as hardware aggressiveness rises.

Why not just…?

Why not just send the top-k highest scores per window? Top-k keeps locally loud candidates, and a locally loud spike is exactly what it preserves. That makes it a genuine threat whenever signals are loud at a single device. Coincidence signals of the multimessenger kind are instead quiet everywhere and significant only jointly, and top-k gives no guarantee on them. It also inherits every upward error of approximate hardware, because a bit flip into the tail makes a background candidate "loud." We settle this by sweeping how evenly the signal's evidence is spread across devices. If top-k matches us when evidence is concentrated, we will report that regime honestly and claim only the spread-evidence regime.

Why not just trigger locally and send decisions? On CMS data, a simple selection gave only 1.6σ where multivariate combination gave 6.4σ (Isolating Unisolated Upsilons). Local hard cuts throw away the joint structure that discovery depends on. Our local-trigger baseline measures this loss directly in the distributed setting.

Why not just train the scorer with AxTrain and skip escalation? AxTrain reduces average error by 25–40% and keeps accuracy within 2% of ideal (AxTrain). But its guarantees concern average accuracy on balanced benchmarks. A 2% average degradation can be concentrated entirely in the one-in-a-million tail that decides a discovery. The no-escalation ablation, evaluated on injected-signal miss rate rather than accuracy, tests this directly.

Why not just send exact likelihoods, since scalars are cheap? At low candidate rates this is correct, and we concede it. The communication saving matters only at trigger-like rates, such as sPHENIX's 3 MHz collisions against a 15 kHz calorimeter trigger ceiling (Intelligent experiments through real-time AI). At those rates 32 bits per candidate per device is significant, and exact scoring of every candidate is the dominant power cost. We report results as a function of candidate rate so the crossover point is visible.

Proposed Experiments

The first testbed partitions the LHC Olympics 2020 R&D dataset and Black Box 1 (The LHC Olympics 2020) across simulated subsystems. Each device holds the substructure features of one jet or one feature group and computes a weakly supervised score. Black Box 2, which contains background only, serves as a false-discovery check. The measures are significance at fixed false-alarm rate, bits per event, and estimated arithmetic energy. For this benchmark the file reports only qualitative outcomes: success on Black Box 1, correct rejection on Black Box 2, and widespread failure on Black Box 3. No reported numerical result was found on Lacuna.

The second testbed is CMS Open Data 2016, split the same way. Here the centralized ceiling is anchored to the reported 6.4σ. We will also repeat the paper's same-sign muon-pair control, where no signal should appear, to confirm that approximate hardware plus escalation does not manufacture a peak.

The third testbed is a synthetic coincidence network modeled on the O4a multimessenger setup and on the DEP sensor datasets. Rare transients are injected at controlled strength and spread across devices. This lets us vary candidate rate, heterogeneity across devices, and how evenly evidence is spread. For reference, DEP reports a tracking-accuracy score (generalized optimal subpattern assignment, GOSPA) of 11.5 on its stable dataset, matching its centralized Gibbs sampler (Distributed Expectation Propagation). That is a tracking metric, and we cite it only to calibrate the simulator, not as a detection baseline.

The decisive checks are three.

First, calibration. Across hardware aggressiveness levels, the realized miss rate on injected signals must stay at or below $\delta$ for our method, while exceeding $\delta$ for the zero-margin ablation and for no-escalation AxTrain.

Second, the frontier. Significance plotted against total cost, meaning bits plus arithmetic energy, should lie strictly above every matched-budget baseline in the spread-evidence regime.

Third, the concession test. The spread-evidence sweep must show where top-k and exact-scalar transmission catch up, so the claimed advantage is bounded rather than asserted.

If the escalation rate needed to hold $\delta$ turns out to exceed a few percent of background at useful energy savings, the lopsided-allocation premise fails. We will report that as the main negative result.