Searches for very rare particle decays, such as a B meson turning into a kaon plus two invisible neutrinos, must find a handful of signal events among hundreds of millions of ordinary collisions. The classifiers that do this are trained on simulation and then applied to real detector data. When the simulation is imperfect, the classifier's performance shifts, and that shift often costs more than the classifier gained. This note proposes a neural event classifier whose filters adapt to each event's overall kinematics, meaning the direction and momentum of the partner B meson and the amount of missing energy. The ability to adapt is added one direction at a time, starting from a plain fixed network. A new direction is admitted only if it improves separation of signal from background and does not widen the disagreement between simulation and real control data.
In two sentences: a dynamic kernel is a filter whose weights are computed from the input. We write it as a fixed filter plus a low-rank, event-dependent correction, and grow the rank of that correction from zero using the gradient signal that the current model cannot absorb. The number of adaptive directions then becomes a measured quantity, chosen by physics sensitivity under systematic uncertainty. It is no longer a hyperparameter such as the number of expert kernels.
Core Claim
For each layer, the kernel applied to event $i$ is
$$W(c_i) = W_0 + \sum_{j=1}^{r} a_j(c_i)\, u_j v_j^\top ,$$
where $W_0$ is a static weight matrix shared by all events and $c_i$ is a short vector of event-level context: tag-side B direction and momentum, thrust axis, missing momentum, and the di-neutrino mass squared $q^2$. Each pair $u_j, v_j$ is a fixed rank-one direction in weight space, $a_j$ is a small learned function of the context, and $r$ is the dynamic rank. At $r=0$ the model is an ordinary static network.
The central claim has three parts. First, the rank can be grown constructively. Second, the right stopping point is set by expected upper-limit sensitivity with data-versus-simulation systematics included, not by validation accuracy. Third, the resulting model beats both the fixed-bank dynamic convolutions from vision and the static networks that suffer most from simulation mismatch. The claim is falsifiable in a specific way. If data–simulation disagreement in control samples does not grow with $r$, or if a boosted-decision-tree analysis on engineered variables matches the grown model on expected limits, the proposal fails.
Where This Sits
Three lines of work supply the method.
The Riemannian Staircase in The non-convex Burer-Monteiro approach works on smooth semidefinite programs optimizes at a fixed rank, checks an optimality gap, and lifts the rank along a descent direction when the gap is nonzero. Low-rank optimization for distance matrix completion uses the same restart logic, embedding a stalled rank-$r$ solution into rank $r+1$ along the escape direction. Low-Rank Matrix Regression via Least-Angle Regression treats rank as sparse selection over a dictionary of rank-one matrices and traces a nested solution path. We borrow the lifting rule and the nested-path idea. All three work on static matrix recovery, where the low-rank object is the answer. Here the low-rank object is an input-conditioned correction inside a classifier, and none of their guarantees carry over. We treat the lifting as a well-motivated heuristic, not a theorem.
Dynamic kernels themselves come from CondConv, which mixes $n$ expert kernels with sigmoid routing weights computed from pooled features. Dynamic Convolution: Attention Over Convolution Kernels mixes $K=4$ kernels with softmax weights and needs temperature annealing to stop all weight collapsing onto one expert. When Sparsity Meets Dynamic Convolution tames the resulting parameter growth by pruning an oversized bank with learned masks. All three fix the size of the bank in advance and work top-down.
Our formulation differs in two ways. The bank size is grown from zero, and the adaptive part is a correction to a static kernel rather than a convex mixture of full kernels. That second choice speaks to the collapse problem. Understanding Incremental Learning of Gradient Descent proves that small initialization makes gradient descent discover a matrix one rank at a time, but only for squared loss under restricted isometry. Vision dynamic networks instead collapse. We read that as a warning that implicit incremental learning cannot be relied on in deep networks, which is why we make the growth explicit rather than hoping it emerges.
The physics setting motivates the work.
- BaBar analysis. Search for the rare decay B → Kνν̄ uses 459 million B-meson pairs with semileptonic tagging, meaning the partner B is reconstructed from a charmed meson plus a lepton. It separates signal with an ensemble of 20 bagged decision trees on 38 engineered variables. It reports 90% confidence upper limits of $1.3\times10^{-5}$ for the charged mode and $5.6\times10^{-5}$ for the neutral mode. Systematics include 14% on signal efficiency and 5–25% on background predictions.
- Belle II flavor tagging. B-flavor tagging at Belle II supplies the key constraint. On real data, a decision-tree tagger reached an effective tagging efficiency of $(30.0\pm1.2\pm0.4)\%$ against $(28.8\pm1.2\pm0.4)\%$ for a static deep network. The deep network degraded more because particle-identification inputs were mismodeled.
- What this implies. Any deep model we propose must be judged on real data under systematics, not on simulated area-under-curve.
- Jet taggers. Particle Transformer for Jet Tagging and Does Lorentz-symmetric design boost network performance in jet physics? show that pairwise physics features and raw particle inputs beat engineered summaries on large simulated jet samples. Their interaction terms are static across events, whereas our correction is conditioned on the event.
Research Program
Event representation. Each event's rest-of-event is represented as a set of reconstructed tracks and calorimeter clusters in the collision rest frame. The rest-of-event is everything not assigned to the reconstructed tag B. Each particle carries four-momentum, charge, particle-identification likelihoods and cluster shape. The backbone is a particle-cloud network with edge convolutions using pairwise invariant mass and angular separation, following the Lorentz-motivated features above. Every edge and node filter uses the kernel form in the Core Claim.
Growing the rank. Training at rank $r$ runs until validation loss plateaus. We then compute, on a held-out batch, the gradient of each event's loss with respect to each layer's weights, $G_i$.
- The average of $G_i$ over events is what the static kernel $W_0$ can already absorb, so we subtract it and project out the existing directions $u_j v_j^\top$.
- The top singular pair of the leading principal component of the remaining event-to-event variation defines $u_{r+1} v_{r+1}^\top$. This is the weight direction whose ideal size differs most across events.
- The new coefficient function $a_{r+1}$ is initialized by regressing each event's projection onto that direction against its context $c_i$.
- The explained variance of that regression is the growth score. A low score means the variation is not predictable from event context, and a static network would do as well.
This rule is the dynamic-kernel counterpart of the escape step in the staircase papers. The layer with the largest score receives the new direction.
Stopping rule. Growth continues until the gain in simulated sensitivity is outweighed by the increase in data–simulation disagreement on control samples. The control samples are:
- off-resonance data collected below the B-pair threshold, which contains only continuum background (the scan in Measurement of the e+e− → bb̄ Cross Section shows BaBar recorded such energies);
- signal-depleted sidebands in extra calorimeter energy.
Disagreement is measured as a two-sample distance between the classifier-output distributions in data and simulation. It is propagated as a background-shape systematic into the expected upper limit. The final rank is the one minimizing the expected limit including this systematic.
Analyses. We will record the full nested path of models from $r=0$ upward. For each rank we report expected limit, control-sample disagreement and parameter count. We also inspect which context variables drive each $a_j$ using Shapley attributions, following the interpretability practice in Interpretable boosted-decision-tree analysis for the Majorana Demonstrator.
Data access. The BaBar samples require collaboration or data-preservation access, which the research file does not document. If access fails, the mechanism study runs only on the open proxy described below, and the physics claim remains untested.
Baselines
Existing decision-tree analysis. The strongest existing competitor is the BaBar analysis itself: 20 bagged decision trees on the 38 kinematic and event-shape variables of Search for the rare decay B → Kνν̄. Its published numbers are the upper limits of $1.3\times10^{-5}$ (charged) and $5.6\times10^{-5}$ (neutral). The file reports no signal efficiency or classifier-level figure, so we found no reported result on Lacuna for per-event discrimination on this dataset.
We will retrain this ensemble on exactly our simulated samples, with the same tag selection and an equal tuning budget over depth and tree count. We will also train a gradient-boosted variant. It counts as beaten only if our model gives a lower expected limit with the control-sample systematic included, in both decay modes. It is the right bar because the Belle II comparison shows trees hold up better on real data.
Static particle-cloud network. The naive reference is the same backbone at $r=0$, trained with identical data, epochs and hyperparameter search. A second static variant receives $c_i$ concatenated to every particle's features. This separates "access to context" from "context reshaping the filters".
Fixed-bank dynamic kernels. These baselines use the same backbone. One is a softmax bank with $K=4$ and temperature annealing as in Dynamic Convolution. Another is a sigmoid-routed bank with 8 experts as in CondConv. The third is a pruned oversized bank as in When Sparsity Meets Dynamic Convolution. All are routed from $c_i$ rather than pooled features, and their parameter counts are matched to our final model.
Ablations of the key ingredient.
- Fixed-rank control. This model is trained from scratch at our final rank with random directions. It tests whether growing matters or only the final rank does.
- Accuracy-stopped growth. Growth stops on simulated validation loss instead of the systematic-aware rule. It tests whether the stopping criterion is what protects real-data performance.
Why not just…?
Why not just keep the boosted decision trees? This is the real threat. The Belle II result shows a static deep network losing to trees on data by 1.2 points of effective tagging efficiency, within one statistical error but in the tree's favour. On the other side, Particle Transformer for Jet Tagging shows raw-particle models improving background rejection by factors of 2 to 3 over earlier approaches on simulated jets. What neither settles is whether that gain survives on B-factory data once systematics are included.
The rank path answers this directly. If the rank-0 network already loses to the trees on expected limit, and every added rank raises the systematic faster than it improves separation, the correct conclusion is that trees win in this regime, and we will report it. A secondary test feeds the network score as one more tree input, to check whether the network contributes information the 38 variables miss.
Why not just tune the number of experts in a fixed dynamic bank? A sweep over $K$ could in principle find a good capacity. Two differences remain.
- A mixture bank adds whole kernels, so capacity rises in large steps with $K$-fold parameters. Our correction adds one rank-one direction at a time on top of a static kernel.
- A sweep chooses $K$ by accuracy, while our path chooses rank by systematic-aware sensitivity.
The fixed-rank-from-scratch and accuracy-stopped ablations test exactly this. If a swept bank with a systematic-aware selection matches us, the contribution shrinks to the stopping criterion. We would then say so.
Why not just give a static network the context as extra inputs? A deep enough static network can represent the multiplicative interactions our kernel builds in, so this is a plausible match. The concatenation baseline tests it at matched parameters. The hypothesis is that confining context dependence to a few explicit directions produces fewer simulation-specific features than letting context enter everywhere. That predicts smaller control-sample disagreement at equal simulated performance. If concatenation shows equal disagreement, the hypothesis fails.
Why not just prune an oversized dynamic network? Pruning an oversized bank reached 73.3% ImageNet top-1 with 23.2M parameters, against 72.7% with 42.7M for the unpruned bank (When Sparsity Meets Dynamic Convolution). But it starts by training the full, most simulation-prone model, and its irregular sparsity gives no ordered path of capacities to evaluate against systematics. It is run as a baseline at matched parameters. Beating it requires lower real-data disagreement, not just fewer parameters.
Decisive Experiments
Open-data mechanism study. This runs first and needs no collaboration access. We train on JetClass with context $c$ set to jet transverse momentum and mass. Targets from Particle Transformer for Jet Tagging are 84.4% accuracy for ParticleNet and 86.1% for the Particle Transformer. The growth rule must reach ParticleNet-level accuracy at small rank to be worth continuing.
As a rough stand-in for data–simulation shift, we compare the distribution of the classifier's QCD score on simulated JetClass QCD jets with that on real CMS jets from Aspen Open Jets, after matching the transverse-momentum and pseudorapidity selection. This proxy has caveats: it mixes physics differences with detector mismodeling. It can only suggest whether disagreement grows with rank, not prove it.
Rare-decay test. On BaBar B→Kνν̄ samples with semileptonic tagging, we compare all models on:
- expected 90% upper limit with systematics;
- control-sample output disagreement in off-resonance data and in sidebands;
- signal efficiency at fixed background.
Decisive pattern.
- Confirmation requires three things together. Control-sample disagreement should rise with rank. The systematic-aware stopping point should give a better expected limit than the tree ensemble. Grown models should beat fixed-rank models trained from scratch at the same rank.
- Refutation is either of two outcomes. Trees remain best on expected limit, or disagreement shows no dependence on rank. The second would mean the stopping rule is doing nothing.
If Belle II flavor-tagging samples become available, the same comparison against the published 30.0% and 28.8% effective efficiencies would test whether the finding carries over to a second experiment.
References
- The non-convex Burer-Monteiro approach works on smooth semidefinite programs
- Low-rank optimization for distance matrix completion
- Low-Rank Matrix Regression via Least-Angle Regression
- CondConv: Conditionally Parameterized Convolutions for Efficient Inference
- Dynamic Convolution: Attention Over Convolution Kernels
- When Sparsity Meets Dynamic Convolution
- Understanding Incremental Learning of Gradient Descent: A Fine-grained Analysis of Matrix Sensing
- Search for the rare decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:mi>K</mml:mi><mml:mi>ν</mml:mi><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover></mml:math>
- B-flavor tagging at Belle II
- Particle Transformer for Jet Tagging
- Does Lorentz-symmetric design boost network performance in jet physics?
- Measurement of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:mi>b</mml:mi><mml:mover accent="true"><mml:mi>b</mml:mi><mml:mo>¯</mml:mo></mml:mover></mml:math>Cross Section between<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msqrt><mml:mi>s</mml:mi></mml:msqrt><mml:mo>=</mml:mo><mml:mn>10.54</mml:mn></mml:math>and 11.20 GeV
- Interpretable boosted-decision-tree analysis for the Majorana Demonstrator
- Aspen Open Jets: Unlocking LHC Data for Foundation Models in Particle Physics