Low-Rank Plus Sparse Acceptance Maps for Polarization Measurements Under Forest-Based Selection

When a spinless B meson decays into two spin-1 mesons, the fraction of decays in which both daughters carry zero spin along their flight direction, called the longitudinal fraction $f_L$, is a sharp test of how the weak interaction and the strong force combine. Simple theory predicts $f_L$ near one. Some penguin-dominated modes instead give values near one half, and the cause of this "polarization puzzle" is still open. These are rare decays, so the measurements are made on small samples. They depend on knowing the detector's acceptance, meaning the probability that an event is reconstructed and selected, as a function of the two helicity angles (the decay angles of each daughter in its own rest frame). Two trends make that acceptance harder to trust. The first is that selections increasingly use tree ensembles such as random forests, which can reshape the angular distribution in ways that are hard to see. The second is that the acceptance comes from simulation, and simulation fails in localized patches of the angular space, for example where slow tracks are mismodeled or detector regions are dead.

The proposal is to model the data-versus-simulation acceptance correction over binned helicity angles as a low-rank matrix plus a sparse error matrix, fitted with an $\ell_1$ loss so that corrupted bins are set aside instead of distorting the whole map. The low-rank assumption is not a statistical convenience; it follows from physics. If each daughter is reconstructed independently, the acceptance factorizes into a product of per-daughter terms, which is a rank-one matrix. The binned signal distribution has rank at most two. The measured yield matrix is the elementwise product of these, so it has rank at most two. Localized mismodeling violates this structure sparsely, and a selection classifier that couples the two daughters violates it smoothly. The goal is to separate these two kinds of violation and show which acceptance treatment keeps $f_L$ unbiased.

Core Claim

Write $C_{ij}$ for the ratio of control-sample data to simulation in helicity bin $(i,j)$, where $i$ indexes $\cos\theta_1$ and $j$ indexes $\cos\theta_2$. We estimate a correction $L = UV^\top$ of small rank $r$ and a sparse residual $S$ by solving

$$\min_{U,V}\ \sum_{(i,j)\in\Omega} w_{ij}\,\big|C_{ij} - (UV^\top)_{ij}\big|,$$

where $\Omega$ is the set of bins with enough control events, $w_{ij}$ is an inverse-uncertainty weight, and $S_{ij} = C_{ij} - (UV^\top)_{ij}$ is read off at the solution. Bins outside $\Omega$ are filled in from the low-rank part.

We make three claims, each testable on simulation. First, in the published regime of per-bin statistics, an $\ell_1$-fitted rank-$r$ correction yields $f_L$ with smaller bias than the current standard, which is smooth polynomial or binned efficiency tables. This should hold when localized corruptions are injected, and the method should cost essentially no precision when none are present. Second, the effective rank $r$ is a diagnostic of the selection itself. A random forest trained on features that couple both daughters, such as the minimum track momentum, raises the rank. The size of that increase predicts how much bias an uncorrected analysis would carry. Third, the $\ell_1$ loss rather than the low-rank constraint alone is what gives robustness to localized defects. A truncated singular value decomposition (SVD), the least-squares version, should not match it.

Research Program

The work rests on three prior results, each playing a distinct role. The measurements that motivate it are BABAR's multidimensional likelihood fits. Measurement of the branching fraction, polarization, and CP asymmetries in B0→ρ0ρ0 fits up to ten variables, including both helicity angles, and reports $f_L\approx0.75$. Observation and Polarization Measurements of B±→φK1± and B±→φK2*± reports $f_L = 0.46 \pm 0.13 \pm 0.06$ for $\phi K_1(1270)$ and $0.80 \pm 0.09 \pm 0.03$ for $\phi K_2^*(1430)$. Observation of B meson decays to ωK* and improved measurements for ωρ and ωf0 finds $f_L$ around 0.4 to 0.5. These analyses treat acceptance with simulation-derived tables or polynomials and use linear Fisher discriminants for background rejection. The systematic uncertainties, 0.06 and 0.03, set the scale at which acceptance bias matters.

The selection regime comes from Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC. That paper shows random forests beating hand-tuned cuts on correlated features. It also shows that such forests degrade under mismatch between simulation and data, which is the failure mode we target. The optimization method comes from Efficient computation of robust low-rank matrix approximations in the presence of missing data using the L1 norm. Its $L_1$-Wiberg algorithm handles missing entries and outliers together and reports RMS error 0.862 against 2.029 for the least-squares version on the Dinosaur sequence with 10% outliers. Our matrices are small, roughly 10 by 10 bins with some empty, which is exactly that algorithm's regime.

We deliberately do not use the self-expressive low-rank representation of Robust Subspace Segmentation by Low-Rank Representation. That method assumes data drawn from independent subspaces and uses the data matrix as its own dictionary. Neither assumption fits an acceptance map. We also do not use the transform-learning domain adaptation of Robust visual domain adaptation with low-rank reconstruction. That method aligns feature spaces, whereas we need a correction over a fixed, physically meaningful binning.

The build has four parts.

The first is a fast simulation of an Υ(4S) B-factory analysis. It includes a parametrized tracker whose efficiency depends on track momentum, continuum and $B\bar B$ backgrounds, and signal generated at the published $f_L$ values of 0.46, 0.75 and 0.80.

The second is a random forest selection trained on kinematic and event-shape features. We will train it in two versions: one given features that couple the two daughters, and one restricted to features that do not.

The third is a set of control channels in which one daughter is spin-0, such as a $B\to V P$ mode. There the vector's helicity distribution is exactly $\cos^2\theta$ and carries no free physics parameter. This avoids confusing an acceptance mismatch with a polarization mismatch. That confusion would be real in a $VV$ control channel, because a shift in $f_L$ is itself a rank-two change to the yield matrix. Per-daughter corrections measured this way, together with a vector-vector channel measured in dedicated simulation variants, give the two-dimensional map $C$.

The fourth is corruption injection into the "data" simulation. We inject localized efficiency holes, momentum-dependent slow-track mismodeling, and smooth global tilts, at controlled amplitude and extent. The corrected acceptance then enters the standard extended unbinned likelihood fit, and $f_L$ is extracted over many pseudo-experiments.

Baselines

The naive reference is to use the simulation acceptance with no data-driven correction. This is what an analysis gets if it trusts simulation. Any method that cannot beat it under injected mismodeling is useless, and any method that is worse than it when simulation is perfect costs precision for nothing.

The current standard is the approach of the BABAR analyses cited above: a per-bin data-to-simulation ratio table, and separately a low-order polynomial fitted by least squares to the control ratios. This is the strongest existing method because it is what published $f_L$ values rely on. For this acceptance-bias question, no reported result was found on Lacuna, so the benchmark is the systematic uncertainty those papers quote, 0.03 for $\phi K_2^*$ and 0.06 for $\phi K_1$. Both the table and the polynomial will use the same control samples, the same binning, and polynomial order chosen by cross-validation on held-out control pseudo-data.

The strongest machine-learning competitor is classifier-based reweighting. A random forest is trained to tell control-sample data from simulation, and its output odds are used as per-event weights on simulation. It is given the same control events, and its depth and size are tuned on the same held-out split.

The key ablation is a truncated SVD at the same rank. It isolates the $\ell_1$ loss from the low-rank constraint.

Beating these baselines means two things. Under injected localized corruption, the absolute bias on $f_L$ must fall below one third of the published systematic, about 0.01 for the $\phi K_2^*$ configuration, while the table, polynomial and SVD approaches exceed it. With no corruption, the statistical width on $f_L$ must stay within 5% of the best baseline, and pull distributions must have unit width.

Why not just…?

Why not just train the forest without angle-correlated features, so it cannot sculpt the helicity distribution? This helps with the selection-induced part and is included as the restricted forest variant. It does nothing about detector mismodeling of slow tracks or dead regions, which exists before any classifier is applied. It also gives up background rejection. The MAGIC work shows that removing informative features, such as timing, costs a factor of two in rejection power (Improving the performance of the single-dish Cherenkov telescope MAGIC through the use of signal timing). The experiment reports $f_L$ bias and background yield for both forest variants, so the trade-off is measured rather than assumed.

Why not just use finer efficiency tables or higher-order polynomials? Finer tables leave few control events per bin and become noisy. Higher-order polynomials let one bad region pull the whole surface, and that pull goes straight into $f_L$. The Wiberg comparison, 2.29 for alternating least squares against 1.01 for $L_1$-Wiberg on synthetic matrices with 10% outliers, suggests this failure is general. We test it directly with the injected-hole scans.

Why not just reweight simulation to data with a classifier? This is the real threat. It is non-parametric and handles smooth multidimensional mismatch well. Its weaknesses are that it has no notion that acceptance should factorize across daughters, and that it spreads a sharp defect into neighboring bins. With small control samples it may either overfit or oversmooth. If it matches the $\ell_1$ low-rank fit on bias and variance across the injection grid, we will conclude that the structural prior adds nothing, and we will report the rank diagnostic as the remaining contribution. A hybrid is also possible: forest reweighting followed by an $\ell_1$ low-rank check on the residual ratio map. We will test it.

Why not just use a robust but unstructured loss, such as a per-bin median? Such a loss downweights outliers but cannot fill empty bins or pool information across rows and columns. Filling bins from the low-rank part is how the method recovers acceptance in regions the control sample barely covers. Comparing results on the missing-bin fraction scan will show whether this matters.

Proposed Experiments

The decisive check is a two-dimensional scan: corruption amplitude, from 0 to 50% efficiency loss, against corruption extent, from one bin to a quarter of the map. For each configuration we run at least 1,000 pseudo-experiments per $f_L$ truth value of 0.46, 0.75 and 0.80, with signal yields matched to the published statistics. We report bias, width and pull coverage for $f_L$ under every acceptance method.

A second check tests the rank claim. We measure the singular-value spectrum of the true acceptance under both forest variants. We then ask whether excess rank beyond one predicts the $f_L$ bias of the uncorrected analysis. If it does, rank becomes a cheap, pre-unblinding warning about any selection.

A third check validates the solver. We reproduce the Dinosaur result of Efficient computation of robust low-rank matrix approximations in the presence of missing data using the L1 norm, the 0.862 against 2.029 RMS comparison, before trusting our implementation.

The idea is falsified in either of two cases. It fails if the $\ell_1$ fit shows no bias advantage over truncated SVD under localized corruption. It also fails if forest reweighting equals it everywhere. In the latter case the finding would be the rank diagnostic alone, and we would say so.

References

  1. Measurement of the branching fraction, polarization, and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:math>asymmetries in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>ρ</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:msup><mml:mi>ρ</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math>decay, and implications for the CKM angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>α</mml:mi></mml:math>
  2. Observation and Polarization Measurements of<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:mi>ϕ</mml:mi><mml:msubsup><mml:mi>K</mml:mi><mml:mn>1</mml:mn><mml:mo>±</mml:mo></mml:msubsup></mml:math>and<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:mi>ϕ</mml:mi><mml:msubsup><mml:mi>K</mml:mi><mml:mn>2</mml:mn><mml:mrow><mml:mo>*</mml:mo><mml:mo>±</mml:mo></mml:mrow></mml:msubsup></mml:math>
  3. Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>meson decays to<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ω</mml:mi><mml:msup><mml:mi>K</mml:mi><mml:mo>*</mml:mo></mml:msup></mml:math>and improved measurements for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ω</mml:mi><mml:mi>ρ</mml:mi></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ω</mml:mi><mml:msub><mml:mi>f</mml:mi><mml:mn>0</mml:mn></mml:msub></mml:math>
  4. Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC
  5. Efficient computation of robust low-rank matrix approximations in the presence of missing data using the L1 norm
  6. Robust Subspace Segmentation by Low-Rank Representation
  7. Robust visual domain adaptation with low-rank reconstruction
  8. Improving the performance of the single-dish Cherenkov telescope MAGIC through the use of signal timing