Finding Hidden CP-Violating Regions in Four-Body Decays with Provably Valid Interaction Search

The universe contains far more matter than antimatter. Explaining this requires laws of physics that treat particles and antiparticles differently, a property called CP violation, and the known amount of it is too small to account for what we see. One place physicists look for new CP violation is in decays of charm and beauty mesons into four or more particles. The measurable signal is a difference between how a particle and its antiparticle distribute their decay products in momentum space. The difficulty is that this signal can be confined to a small corner of a five-dimensional space of momenta, and it can have opposite signs in different corners. Current analyses compress the space to one number, such as the triple product of three momenta, or check a few hand-picked subregions. Both cancel or dilute exactly the localized effects they are meant to find.

We propose to treat the search for a CP-violating region as a high-order interaction discovery problem with a provable error guarantee. A candidate region is a conjunction of conditions on several kinematic variables, for example "the $K^+K^-$ mass is in its upper third and the $\pi^+\pi^0$ helicity angle is in its lower half and the triple product is positive." The hypothesis being tested is that particles and antiparticles populate the region equally. The number of such conjunctions grows combinatorially with the number of variables and with depth. A naive scan is therefore either computationally infeasible or, after correcting for the many tests, statistically useless. Statistical interaction-mining methods solve this problem in genomics and tabular data. They prune candidate regions that cannot possibly reach significance and keep the family-wise error rate (the probability of even one false discovery) at the nominal level. Our contribution is to adapt one of them to the special structure of CP tests and to show whether it finds localized asymmetries that the triple-product and binned approaches miss, with a calibrated false-discovery guarantee.

Core Claim

The claim is that a continuous, Tarone-style pruned search over conjunctions of rank-transformed kinematic variables can detect localized CP asymmetries in four-body decays at controlled family-wise error. We expect it to have more power than the one-dimensional triple-product asymmetry and than Bonferroni-corrected fixed-bin scans whenever the asymmetry is confined to a region that requires three or more variables to describe. A further claim is that the method returns that region as an interpretable object. A black-box classifier test does not do this.

The constraint that shapes the design is that the method from the interaction-mining direction cannot be used off the shelf. Finding Statistically Significant Interactions between Continuous Features introduces C-Tarone. C-Tarone scores continuous features by normalized ranks, defines a "copula support" that falls monotonically as conditions are added, and derives a lower bound on the achievable p-value of a likelihood-ratio test. Any candidate whose bound exceeds the corrected threshold is "untestable," and so are all of its extensions, so whole branches of a depth-first search are pruned. The same paper reports a problem that matters a great deal here. At large sample sizes (around $N = 200{,}000$), precision drops because tiny, meaningless dependencies become significant. Flavor datasets are larger than that. The Belle $B\bar B$ sample in Measurement of time-dependent CP violation parameters in $B^0\to K_S^0 K_S^0 K_S^0$ decays at Belle is $772\times10^6$ pairs, so a naive transfer would produce floods of spurious "discoveries." We make the null hypothesis exact rather than approximate, which makes the large-sample problem tractable. CP tests compare an event sample with its own CP-conjugate. If kinematic variables for antiparticle decays are computed on the CP-conjugated final state, and the sign of the triple product is flipped for antiparticles, then any region's particle fraction is exactly one half under CP conservation, apart from production and detection asymmetries. Those asymmetries are not negligible correlations of physics origin. They are instrument effects, and they can be measured in a control channel.

Research Program

The method has three components.

The first is a CP-symmetric feature map. Each event from a $D^+$ or $D^-$ decay is mapped to a vector of continuous kinematic variables: two-body and three-body invariant masses, helicity angles, and the triple product $C_T = \vec p_1\cdot(\vec p_2\times\vec p_3)$ used in Search for CP violation using T-odd correlations in $D_{(s)}^{+}$ decays. The map is applied to the conjugate final state for $D^-$ decays, with the sign of $C_T$ flipped. The label $y\in\{0,1\}$ is the parent flavor. Separate searches are run on events with $C_T>0$ and $C_T<0$. This mirrors Belle's construction $a_{CP}^{T\text{-odd}}=\tfrac12(A_T-\bar A_T)$, which cancels strong final-state phases, while letting the other variables localize the effect.

The second is a pruned search for asymmetric regions. Each variable contributes two monotone "half-space" features per quantile cut: a rank-score above the cut and a rank-score below it. A candidate region is a conjunction of up to $k$ such features. Following C-Tarone, its support is the product of rank scores, which shrinks as conditions are added. The test statistic is a binomial likelihood-ratio test of the flavor fraction inside the region against the reference fraction $\pi_0$:

$$G(R) = 2\,n_R\,\mathrm{KL}\!\left(\hat\pi_R \,\|\, \pi_0\right),$$

Here $n_R$ is the effective event count in region $R$, $\hat\pi_R$ is the observed particle fraction there, and $\pi_0$ is the expected fraction under CP symmetry. For a given $n_R$, $G(R)$ is largest when the region is pure in one flavor. This gives the minimum achievable p-value, and the support monotonicity carries over unchanged, so the same pruning applies. We extend the reference from a constant to a region-dependent $\pi_0(R)$ estimated in a Cabibbo-favored control channel. Belle measured the asymmetry in such a channel, $D^+\to K^-\pi^+\pi^+\pi^0$, to be consistent with zero. The control channel calibrates detection asymmetries, and its estimation uncertainty is propagated into the test.

The third is valid inference after the search. Tarone-style pruning controls family-wise error by counting only testable regions, but the final reported region is still chosen from the data. We therefore borrow from Selective Inference for Sparse High-Order Interaction Models and its faster successor Fast and More Powerful Selective Inference for Sparse High-Order Interaction Model. These works condition the test on the selection event and keep exact error control over search spaces up to $10^{17}$ candidates. They were built for sparse linear models rather than count tests over regions, so the adaptation is a research question in its own right. If it fails, the fallback is sample splitting: search on half the data, test the chosen region on the other half.

The physics direction supplies the application and its exact symmetry. The interaction direction supplies the search engine. The proposal stands or falls on whether this transplant beats the simpler things physicists already do.

Baselines

The naive reference point is the integrated measurement that is standard today: a single global charge asymmetry, plus the one-dimensional T-odd asymmetry. For the latter, the strongest reported number in the file is Belle's $a_{CP}^{T\text{-odd}}(D^+\to K^-K^+\pi^+\pi^0) = (+2.6\pm6.6\pm1.3)\times10^{-3}$, with the largest subregion fluctuation about $2\sigma$, from the Belle T-odd paper. This is the strongest baseline for this question because it is the published method on the same decay. It will be computed on exactly the same simulated samples as our method. Beating it means detecting, at family-wise error 0.05, injected localized asymmetries whose integrated T-odd asymmetry lies inside that measured $\pm6.6\times10^{-3}$ statistical band.

The second baseline is the manual subregion check that Belle performed: a fixed grid of bins on two or three chosen variables, each tested for asymmetry, with a Bonferroni correction. It gets the same events and the same variable set. The bin layout is tuned on independent tuning samples with a search budget equal to the number of hyperparameter configurations we allow ourselves.

The third baseline is median-split binarization followed by standard Tarone itemset mining. Finding Statistically Significant Interactions between Continuous Features reports this approach to be orders of magnitude less precise than C-Tarone. We include it to check whether that finding holds on kinematic data, and it uses the same depth limit $k$.

The fourth baseline is a classifier two-sample test. A gradient-boosted tree ensemble or a small neural network is trained to separate $D^+$ from $D^-$ on the CP-symmetric features, and its held-out accuracy is turned into a p-value by permutation. It gets the same events, the same data split, and a matched tuning budget. No reported result was found on Lacuna for a classifier test on these decays.

Two ablations isolate the key ingredients. In the first, the depth is capped at $k=2$ (pairwise regions only) while everything else stays the same. This tests whether high-order conjunctions matter. In the second, the control-channel reference $\pi_0(R)$ is replaced by a constant one half. This tests whether calibration is what prevents the large-sample collapse.

Why not just…?

Why not just train a classifier to tell particles from antiparticles and test its accuracy? This is the most serious threat. A flexible classifier may match or exceed our power to say that some asymmetry exists. Its weaknesses are that it gives no region, and that its significance comes from a permutation test whose power depends on tuning choices that are hard to fix before looking at the data. Our experiments report power curves for both methods at matched error rates. If the classifier is uniformly more powerful, we will reframe our method as a localization tool: one that, once the classifier flags an asymmetry, returns a region and a post-selection p-value for it. We would not claim superior detection.

Why not just scan a fixed grid of bins, as Belle already did? A grid with Bonferroni correction is transparent, and it is competitive when the asymmetric region aligns with the chosen axes and bin edges. It loses power when the region involves three or more variables, because the number of bins explodes and every bin pays the correction. Pruning avoids paying for regions that could never be significant. The injection experiments vary how many variables define the region, which shows directly where the grid stops being competitive.

Why not use a pairwise model, such as a second-order factorization machine, instead of deep conjunctions? Higher-Order Factorization Machines reports gains from moving beyond pairwise terms (test AUC 0.856 at degree 2 versus 0.875 at degree 3 on NIPS co-authorship). The T-odd triple product is itself a three-way contraction of momenta. Still, a pairwise model on well-chosen variables, $C_T$ included, might capture most realistic signals. The $k=2$ ablation settles this. If it matches the full search, depth is unnecessary and we will say so.

Why not fit an isobar amplitude model, the standard description of a multi-body decay as a sum of two-body resonances? When such a model is correct it is the most powerful analysis. Evidence for direct CP violation from Dalitz-plot analysis of $B^\pm\to K^\pm\pi^\mp\pi^\pm$ shows how fragile it becomes in practice: results depend on lineshape choices, and an unexplained $f_X(1300)$ state had to be added by hand. Measurement of the CP-even fraction of $D^0\to K^+K^-\pi^+\pi^-$ turned to model-independent methods for four-body decays for this reason. Our method is model-independent and complementary. It flags where a model should be scrutinized and does not replace one.

Proposed Experiments

Belle charm data are not publicly accessible, so the core study is simulation. We generate $D^\pm\to K^-K^+\pi^\pm\pi^0$ phase space using a resonance-based amplitude toy model. We inject CP violation as a weak-phase difference confined to a chosen region of $j\in\{1,2,3,4\}$ kinematic variables, at magnitudes spanning roughly 0.1 to 3 times the Belle statistical uncertainty. We add a realistic flavor-dependent detection efficiency that varies across phase space, and we produce a matched control channel with that efficiency and no CP violation. Sample sizes range from $10^5$ to $10^7$ events. This range deliberately crosses the regime where C-Tarone is reported to lose precision.

The decisive checks are five. First, under the null (no injected CP violation, detection asymmetry present), the empirical family-wise error must stay at or below 0.05 across all sample sizes for the calibrated method. We expect the uncalibrated ablation to exceed it as $N$ grows. Second, power at fixed error must be higher than the integrated T-odd asymmetry and the Bonferroni grid for $j\ge3$. Third, localization quality is measured as the overlap (intersection over union) between the returned region and the injected one. Fourth, the comparison with the classifier test at matched error rates decides whether the contribution is detection or localization. Fifth, runtime and the number of testable regions are reported as functions of depth $k$ and $N$.

We also run the method on the Cabibbo-favored channel alone as a pure null. A collaboration with access to Belle or BESIII data could repeat this null test on real data before any search of the signal channels. The result would be a credible, reusable test that searches all of phase space for CP violation with a guarantee against false discoveries. If it fails, the failure will be informative: it will show which simpler tool, the classifier or the grid, already does the job.

References

  1. Finding Statistically Significant Interactions between Continuous Features
  2. Measurement of time-dependent <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> violation parameters 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 stretchy="false">→</mml:mo><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup></mml:math> decays at Belle
  3. Search for $C\!P$ violation using $T$-odd correlations in $D_{(s)}^{+}\to K^{+} K^{-}π^{+}π^{0}$, $D_{(s)}^{+}\to K^{+} π^{-}π^{+}π^{0}$, and $D^{+}\to K^{-}π^{+}π^{+}π^{0}$ decays
  4. Selective Inference for Sparse High-Order Interaction Models
  5. Fast and More Powerful Selective Inference for Sparse High-Order Interaction Model
  6. Higher-Order Factorization Machines
  7. Evidence for direct<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>violation from Dalitz-plot analysis 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:msup><mml:mi>K</mml:mi><mml:mo>±</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>∓</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>±</mml:mo></mml:msup></mml:math>
  8. Measurement of the <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>-even fraction of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>