Quenched Kinematic Fits on Analog Crossbars

Particle physicists clean up noisy detector measurements with a kinematic fit. They nudge each measured track and photon by the smallest amount, weighted by how uncertain it is, until the event obeys conservation of energy and momentum and reproduces known particle masses. This makes reconstructed masses sharp and removes background. It is also an iterative solve with matrix inversions, repeated for every way of pairing photons in the event, and that is too slow for the microsecond decisions a collider trigger must make. Analog in-memory chips do matrix multiplications at very low energy: they store a matrix as device conductances in a crossbar, a grid of memory cells at every row–column crossing. Kösters et al. showed such a chip, built from phase-change memory, keeps up with the Large Hadron Collider's 40 MHz event rate at about 2 W. But they ran an unconstrained neural classifier, and they state that the effect of device noise on "physics-correctness" is open. Analog noise is exactly what a conservation law cannot tolerate.

We propose to borrow the logic of the lattice-QCD quenched approximation, in which an expensive, configuration-dependent factor is frozen to make the computation tractable and the resulting bias is then measured and partially undone. Here the frozen factor is the kinematic fit's projection matrix. We fix it at a reference kinematic point and store it in the crossbar as stationary weights. The cheap, exact part of the fit, evaluating how badly the current estimate violates conservation, stays in digital logic and feeds back each iteration. The result is that conservation is satisfied to digital precision even though the stored matrix is noisy and approximate. Device noise and the freezing only degrade how close the fitted point is to the true optimum, and we predict that this loss grows quadratically in the matrix error. The question then becomes how many frozen matrices a chip of fixed synaptic area must hold to keep that loss below the detector resolution.

Core Claim

A standard linearized kinematic fit updates the measured parameters $\alpha$ (track curvatures, angles, cluster energies) by

$$\alpha^{(k+1)} = \alpha^{(k)} - M\,H(\alpha^{(k)}), \qquad M^{*} = V D^{T}\left(D V D^{T}\right)^{-1},$$

where $H(\alpha)$ is the vector of constraint violations (four-momentum imbalance and mass-squared mismatches), $V$ is the measurement covariance, $D = \partial H/\partial\alpha$ is the constraint Jacobian, and $M^{*}$ is the exact per-event projector. In the conventional fit, $D$ and $V$ are recomputed and inverted every event and every iteration. A "quenched" fit replaces $M^{*}$ with a fixed $M$, computed offline at a reference kinematic point and programmed into conductances. $M$ is then further perturbed by device noise, drift, and 4–8 bit precision limits.

The update has a property that makes this safe. Any fixed point satisfies $H = 0$ exactly, as long as $\|I - D(\alpha)M\| < 1$ near the solution. So the error in $M$ affects only how fast the iteration converges and where on the constraint surface it lands. It does not affect whether conservation holds, because $H$ is evaluated digitally in full precision. The fit's $\chi^2$, the weighted sum of squared adjustments, is minimized on the constraint surface at the exact solution. A landing point displaced by $O(\|M - M^{*}\|)$ along that surface therefore raises $\chi^2$, and degrades resolution, only at $O(\|M - M^{*}\|^{2})$. This quadratic prediction is the central falsifiable claim.

The lattice analogy organizes the work. The fully quenched version uses a single reference $M$ and is cheapest. "Partial unquenching" uses a bank of $K$ projectors, each indexed by a coarse kinematic region, the way partially quenched simulations restore some sea flavors but not all (Allton et al.; Armour et al.). The synaptic area budget fixes $K$. The analogy is used to structure the error budget, not as a claim of mathematical equivalence.

Research Program

The first component is analysis. We will derive the contraction condition and the quadratic $\chi^2$ excess for 4C fits (four-momentum only), 5C fits (adding one intermediate mass, such as the $\pi^0$), and 6C fits (two intermediate masses). From these we will predict a minimum bank size $K$ as a function of how much $D$ varies across the phase space of a decay topology. The mass constraints are nonlinear, so $D$ varies with the opening angle of the photon pair. This is where quenching bias should concentrate, much as quenching errors in lattice QCD concentrate in quantities driven by vacuum fluctuations, such as the flavor-singlet sector (Fodor et al.).

The second component exploits combinatorics. For topologies with identical photons, every pairing hypothesis uses the same Jacobian up to a permutation of input columns. One stored projector can therefore serve all pairings by re-routing inputs rather than storing extra matrices. This turns the combinatorial explosion flagged for decays such as $\chi_{cJ}\to\eta\eta\eta'$ (BESIII) into a routing problem instead of a storage problem. Routing memory is the area bottleneck Qiao and Indiveri identify in mixed-signal chips, so the area accounting must include it explicitly.

The third component handles drift. Detector calibration drifts change $V$. Rather than rewriting conductances continuously, we adapt the error-triggered write rule of Payvand et al.. In their scheme, conductances are rewritten only when a local error crosses a threshold, which cut writes 100-fold on gesture data. Our trigger is the running mean iteration count, or the $\chi^2$ excess, of a bank entry. This protects write endurance, which Yang et al. list as a core limit of memristive crossbars.

The hardware substrate will be simulated. We will use phase-change-memory noise and drift models of the kind used by Kösters et al., and multi-device-per-synapse precision boosting as in Boybat et al.. A fabricated-chip test is a stretch goal.

Baselines

The naive reference point is no fit at all. Invariant masses would be computed from raw smeared momenta, with events selected by loose cuts. This establishes how much resolution and background rejection the fit buys, and the quenched fit must recover most of that gain.

The strongest existing method on the physics side is the exact iterative kinematic fit in double precision. This is the procedure used for 4C selection in BESIII's $J/\psi\to\gamma\pi^0,\gamma\eta,\gamma\eta'$ measurement and for 6C fits in Ablikim et al.. It defines the optimal mass resolution and the correct $\chi^2$ distribution. It is run on identical simulated events with the same covariance model, iterated to convergence, with no latency limit. Beating it is not the goal. The criterion is reaching within 5% of its mass resolution and signal efficiency at fixed background rejection. The file reports no resolution figure for such fits, so this baseline is computed in-house. For what precision kinematic anchoring ultimately supports, the CLEO $\eta'$ mass measurement reports $957.793 \pm 0.054 \pm 0.036$ MeV (CLEO).

The strongest existing method on the hardware side is the analog in-memory approach of Kösters et al.. It reports about 2 W at 40 MHz and up to 1,000× better energy efficiency than CPUs and GPUs, on anomaly detection with 76 kinematic inputs. We will reproduce their style of unconstrained network, trained to output fitted momenta, on the same crossbar noise model and the same parameter budget, and measure its constraint violation. Beating it means comparable energy and latency per event with constraint residuals at digital precision rather than at the device-noise floor. The file has no reported result on Lacuna for constraint violation or mass resolution of any analog-hardware kinematic fit.

Two ablations isolate the key ingredient. The first removes the digital residual feedback: one analog multiply, open loop. This should show conservation violated at the noise level, which tests whether the feedback loop is what does the work. The second runs the same frozen projector bank in noiseless 8-bit digital arithmetic. This separates the cost of quenching from the cost of analog noise.

Why not just…?

Why not just run the exact fit on FPGAs, which triggers already use? This is the real threat. The quenched-projector idea is not specific to analog hardware: a fixed $M$ times a residual is just as cheap on an FPGA, and removing the per-event inversion helps there too. The file contains no FPGA kinematic-fit benchmark, so we will cost both. Our noiseless-digital ablation is effectively the FPGA design. If it matches the analog version on energy per event at the projected bank sizes, we will say so: the contribution then becomes the quenching and bank-sizing analysis, with analog as one deployment target.

Why not just learn the fit with a neural network, as the Particle Transformer does for jets? Particle Transformer builds pairwise kinematics into attention and reports 86.1% accuracy on JetClass versus 84.4% for ParticleNet. However, it satisfies conservation only approximately, and its pairwise cost grows quadratically with particle count. Our open-loop learned-regressor baseline tests directly whether a learned map on noisy hardware can hold residuals below detector resolution. We expect it cannot, because nothing pins its output to the constraint surface.

Why not just fit only the events that survive a coarse trigger, in software? For low-multiplicity channels at BEPCII rates this may well be enough, and we will not claim otherwise. The case for the proposal rests on high-multiplicity topologies and LHC-scale rates. We will report the event rate and pairing count at which offline fitting stops fitting within a realistic trigger budget, and restrict our claims to that regime.

Proposed Experiments

We will generate simulated events for three topologies drawn from the file. The first is $J/\psi\to\gamma\eta'$-like three-photon final states with a 4C fit, after BESIII. The second is $\chi_{cJ}\to\eta\eta\eta'$-like multi-photon states with mass constraints and heavy pairing combinatorics. The third is hyperon-pair states with a 6C fit. Each uses Gaussian smearing and a variant with non-Gaussian energy-leakage tails, to probe the known failure of Gaussian-assumption fits. For an LHC-rate stress test of throughput, we will reuse the 76-variable event format of Kösters et al. We will report mass resolution, signal efficiency at fixed background, calibration of the $\chi^2$ distribution, the maximum constraint residual, iterations to convergence, and simulated energy and latency per event.

The first decisive check is whether the resolution loss scales quadratically with injected projector error. We will sweep conductance noise and bit precision and fit the slope on a log-log plot. A linear slope would falsify the stationarity argument and the error budget built on it. The second check is whether the closed-loop fit keeps residuals at digital precision while the open-loop ablation does not. The third is whether a bank size $K$ that fits a realistic crossbar area reaches 5% of exact-fit resolution on the high-combinatorics topology. The fourth is whether error-triggered rewrites track a simulated calibration drift with at least an order of magnitude fewer writes than continuous updating, as the Payvand rule achieved in its original setting.