Searches for very rare particle decays end in a likelihood fit. The analyst writes down a probability density for what signal events look like and another for background, then asks how many signal events best explain the handful of candidates seen. When only a few events are expected, every bit of separating power counts. Even so, the densities in these fits are usually built as products of one-variable shapes, which throws away correlations between measured quantities. The reason is practical: nobody has a flexible multivariate density that can also be normalized and integrated exactly over the analysis region. We propose to fill that role with a matrix product state (MPS), also called a tensor train. An MPS writes a high-dimensional function as a chain of small matrices, one per variable. It has a useful property: its total probability, its marginals over any subset of variables, and its integral over any box-shaped region can all be computed exactly by multiplying those matrices together.
The idea fits in two sentences. Represent the signal and background densities of an unbinned likelihood fit as MPS densities over continuous kinematic variables. The size of the matrices then becomes a single dial that runs from today's factorized fit (size one) to a fully correlated model, and the MPS's exact integrals let the background model be refit to real data in sidebands and transported into the signal window without Monte Carlo integration. The question is whether this beats neural density models, and calibrated classifiers, in the regime where it matters most. That regime is a tiny signal, a mismodeled background, and limited calibration data, with coverage of the resulting limits as the measure.
Core Claim
An MPS density for $K$ continuous variables can be written as
$$P(x) = \frac{1}{Z}\Big|\sum_{i_1\dots i_K} A^{(1)}_{i_1} A^{(2)}_{i_2}\cdots A^{(K)}_{i_K}\,\phi_{i_1}(x_1)\cdots\phi_{i_K}(x_K)\Big|^2,$$
where $\phi_i$ are fixed orthonormal one-dimensional basis functions. These are Legendre polynomials for bounded variables and Fourier modes for azimuthal angles, which makes periodicity exact by construction. Each $A^{(k)}_{i}$ is a $\chi\times\chi$ matrix, $\chi$ is the bond dimension, and $Z$ is the normalization. Squaring guarantees $P\ge 0$.
Because the basis is orthonormal, $Z$ is a chain contraction costing order $K d\chi^3$, where $d$ is the number of basis functions per variable. The integral of $P$ over a rectangular region $R$ comes from replacing the identity at each site by the Gram matrix of the basis on that interval, which is analytic. Marginalizing a variable amounts to leaving out its site. The ability to compute exact marginals and normalizations by contraction is the property established for discrete models in Matrix Product States for Inference in Discrete Probabilistic Models. The continuous feature-map construction comes from A matrix product state model for simultaneous classification and generation. Neither was used inside a statistical fit, and the latter stayed in two-dimensional toy problems.
Setting $\chi=1$ reduces the model exactly to a product of independent one-dimensional densities. That is the factorized form $P(\Delta E)P(M)P(\mathcal F)$ used in the CLEO unbinned extended-likelihood analyses (τ→μγ, B⁺→ωK⁺). The proposal's central claim is therefore falsifiable along a single axis. If correlations carry sensitivity, discovery significance and upper-limit strength should improve monotonically with $\chi$ before saturating, and limit coverage should stay correct throughout.
A caveat shapes the whole design. If the simulation is perfect and the background shape is fixed, the test of signal strength depends only on the ratio $p_s/p_b$. A well-calibrated classifier estimates that ratio directly, so a density model has no theoretical advantage there. A normalized density matters when the background model itself must be fitted or normalized. The common cases are refitting to data sidebands because simulation is wrong, integrating over a signal window or acceptance region, and handling events with missing features. The claim we test is that MPS densities are better in exactly those operations and match the strongest classifiers otherwise.
Research Program
The first component is a continuous MPS density estimator suited to collider kinematics. Two known failure modes of MPS training motivate the design choices.
The first is the ordering problem. An MPS places variables on a line, so strongly coupled variables placed far apart need a large $\chi$. Finding the best order is noted to be hard (Matrix Product States for Inference in Discrete Probabilistic Models). We will order variables greedily by estimated pairwise mutual information and report sensitivity to the ordering.
The second is the "causality trap" identified by Tang, Khoo and Ying. Plain gradient descent and two-site sweeps learn loop-like or long-range dependencies as if the chain were open, and increasing $\chi$ does not fix this. Collision events contain exactly such structure, in periodic angles and momentum balance across all objects. We will therefore compare three training schemes: plain gradient descent, the gauge-invariant natural gradient, and sketch-based warm-starts. Tang et al. report that the warm-start rescues training from as few as 128 samples on periodic Ising models. Whether that sample efficiency carries over to kinematics is itself a result, because calibration data in sidebands are scarce.
The second component is the statistical layer. We will plug signal and background MPS densities into the extended unbinned likelihood and build profile-likelihood test statistics for discovery significance and 90% confidence-level upper limits.
For data-driven background, the background MPS is first fit on simulation. It is then refit on observed sideband events, meaning events outside a window in a reconstructed-mass variable. The mass site uses a low-order Legendre basis, so the interpolation into the window is smooth by construction. The window background yield and shape follow from the exact region integral, with no sampling.
The third component is a characterization of truncation error. Existing MPS studies find that accuracy saturates at very small $\chi$ (about 4 on two-dimensional benchmarks, per the classification-and-generation paper). Rare-event sensitivity, however, lives in the far tail of the background distribution. We will measure how $\chi$ controls background density error in the most signal-like 1% and 0.1% of events. We will also measure how that error translates into bias in fitted signal yield and into coverage of upper limits.
Baselines
The naive reference is a cut-and-count analysis. It applies rectangular selections on high-level variables and computes a Poisson significance of roughly $S/\sqrt{B}$. Every method must beat it, and its gap to the others shows how much shape information is worth at all. It will use the same events and selections optimized on the same validation split.
The established practice is the factorized unbinned extended-likelihood fit of the CLEO type. CLEO used it to set $\mathcal B(\tau\to\mu\gamma)<1.0\times10^{-6}$ at 90% C.L. with 15.2% efficiency, and to observe $B^+\to\omega K^+$ at 3.9σ (τ→μγ, ωK⁺). Those results are on CLEO data we do not have, so they set context, not a target number. In our study this baseline is implemented exactly as the $\chi=1$ MPS. That makes it the key ablation: any gain at larger $\chi$ comes from captured correlations, with the basis, training data and fit held fixed.
The strongest discriminative competitor is a deep network on low-level four-momenta, following Baldi, Sadowski and Whiteson. On the HIGGS benchmark they report 4.9σ for a five-layer network against 2.5σ for a shallow network on low-level features, and an 8% AUC improvement over traditional methods. We will reproduce their architecture and use it in two ways. In the first, its score is histogrammed and fit as a binned template, which is standard practice. In the second, the calibrated score is turned into an unbinned likelihood ratio, which is the theoretically sufficient statistic when simulation is correct. A boosted decision tree is included as well; the quantum machine learning LHC study reports a BDT AUC of 0.80 on $H\to\mu\mu$. Our significances will be computed under a single luminosity and signal-count convention applied to all methods. We will first confirm that our reproduction of the deep network approaches the 4.9σ figure under Baldi et al.'s setup before comparing anything.
The strongest density competitor is a normalizing flow, an invertible neural network that gives exact likelihoods for continuous data. Flows have been used on LHC data for anomaly detection (Improving Variational Autoencoders for New Physics Detection at the LHC With Normalizing Flows, which reports 72% median signal identification at 1% background acceptance on DarkMachines). We will use a mass-conditional flow in the sideband setting. Its window integrals will be estimated by Monte Carlo sampling, and missing features will be handled by imputation or a separate model.
All methods get the same training events, an equal hyperparameter search budget of a fixed number of trials on a shared validation split, and comparable GPU-hours. Beating a baseline means a higher median discovery significance, or a lower median upper limit, on the same toy ensembles, with empirical coverage of 90% limits within statistical error of 90%. A method that gains significance while undercovering does not count.
Why not just…?
Why not just histogram a deep-network score and do a binned fit? It is the field's default and the HIGGS results show deep networks extract most of the correlation information. Binning costs resolution in narrow features and fails when bins are nearly empty. The bigger problem is that a score trained on mismodeled simulation cannot be refit to data. Our perfect-simulation experiment may show the binned score close to MPS, and we will report that plainly. The mismodeled-background experiment is where we expect it to show bias or undercoverage.
Why not just use the calibrated classifier as an unbinned likelihood ratio? This is the most serious threat. With correct simulation and a fixed background shape, it is the optimal statistic, and MPS can at best tie it. We include it as a first-class baseline. Our claim survives only if MPS wins when the background must be fitted to sideband data or integrated over a window, and if not, the honest conclusion is that MPS adds nothing for rare-event fits beyond the factorized dial.
Why not just use a normalizing flow? Flows give exact normalized densities and can be refit to sidebands, so they are a genuine competitor. What they lack is exact marginals and region integrals. The flow paper above also reports sensitivity to hyperparameters and unstable results at small sample sizes. We test both weaknesses directly. Sideband refits will use calibration sets from a few hundred to a few thousand events. Feature-masking experiments will require marginalization. If flows match MPS on bias, coverage and significance in those regimes, the MPS advantage reduces to computational convenience, and we will say so.
Proposed Experiments
The primary testbed is the HIGGS benchmark from Baldi et al.: 11 million simulated events with 21 low-level and 7 high-level features. SUSY, with 5 million events and 18 features, is a second benchmark; we found no reported result on Lacuna for SUSY significance or AUC, so it serves as a replication set rather than a comparison to published figures.
Rare-event conditions are created by drawing toy datasets from held-out events. Each contains a fixed background expectation and an injected signal of zero to a few tens of events. Every method is run on thousands of toys.
The first experiment holds simulation correct. It sweeps $\chi$ from 1 to 64 and reports median significance, median 90% upper limit, empirical coverage, and tail density error in the most signal-like 1% and 0.1% of background. The decisive check is a curve rising from the $\chi=1$ factorized fit and approaching the unbinned classifier-ratio result without loss of coverage. A flat curve would refute the premise that correlations matter for these fits.
The second experiment introduces controlled mismodeling. "Data" background is distorted relative to "simulation" by a jet-momentum scale shift and a reweighting correlated with $m_{wwbb}$. Each method trains on simulation; density methods are then refit on data sidebands outside an $m_{wwbb}$ window, and the window background is predicted. We measure bias in fitted signal yield on signal-free toys, upper-limit coverage, and significance. The decisive check is that MPS window predictions stay unbiased with sideband sets of a few hundred events, while classifier-based fits are biased and flows show larger variance at the same sample size.
The third experiment randomly masks features, such as one jet's b-tag or angle, in a fraction of events. MPS marginalizes exactly, while the alternatives impute. Success means no loss of coverage and a significance decline proportional to the information removed.
A cross-cutting check addresses the training traps. We train with plain gradient descent, natural gradient, and sketch warm-start, using raw angles and momentum-balanced features. Following Tang et al., we test whether held-out log-likelihood of the angular and momentum-sum marginals saturates under plain gradient descent but not under the corrected schemes. If it does, that is the first evidence of the causality trap in real kinematics.
References
- Matrix Product States for Inference in Discrete Probabilistic Models
- A matrix product state model for simultaneous classification and generation
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">B</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mo>→</mml:mo><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">ω</mml:mi><mml:mi mathvariant="italic">K</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and Search for Related<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="italic">B</mml:mi></mml:math>Decays Modes
- Initialization and training of matrix product state probabilistic models
- Searching for exotic particles in high-energy physics with deep learning
- Application of Quantum Machine Learning to High Energy Physics Analysis at LHC using IBM Quantum Computer Simulators and IBM Quantum Computer Hardware
- Improving Variational Autoencoders for New Physics Detection at the LHC With Normalizing Flows