At B-factories, many searches for rare decays with invisible particles depend on one technique. The physicist fully reconstructs one of the two B mesons in the event (the "tag") from its hadronic decay products. Because the beam energy is known, subtracting the tag from it gives the momentum of the other meson, and that in turn reveals any missing energy. Modern tag reconstruction, such as the Full Event Interpretation (FEI), tries hundreds of decay chains at once. Each analysis then keeps some of these chains, usually by applying one cut on a classifier score. That choice is tuned for statistical power. At Belle II luminosities, however, statistical error is no longer the main limit. The limit is the residual systematic uncertainty: places where the simulation disagrees with the detector, and those errors are correlated across tag channels.
We propose to treat tag-channel selection as an explicit discrete optimization. Each "cell" (a tag decay mode crossed with a classifier-score bin) is either kept or dropped. The objective is the full expected variance of the signal measurement: statistical error plus a covariance of correlated systematics plus the cost of calibrating each kept cell. Writing the keep/drop variables as $\pm 1$ turns this into a problem with a constraint that every entry has modulus one. Such problems have well-developed solvers in radar and antenna design. We borrow one of them, Extreme Point Pursuit, which relaxes the hard constraint to its convex hull and adds a concave penalty. A second step, a sweep over signal yield in the relaxed problem, produces a certified lower bound on the best achievable uncertainty. The deliverable is therefore a selection together with a number that says how close any selection could get to optimal.
Core Claim
Two things are missing from current practice. First, selections should be scored by the total uncertainty they will carry, including systematics shared between channels. Second, the selection should come with a certificate of near-optimality, which no cut on a classifier score can supply. The constant-modulus machinery from signal processing does more than relabel a binary problem. It gives a solver with exact-penalty guarantees, and its convex relaxation yields the bound. We expect gains to concentrate in exactly the regime the literature flags as systematics-limited. That regime includes the extra-calorimeter-energy mismodeling that limited the BABAR $B^+\to\tau^+\nu$ search, and the particle-identification mismatch between simulation and data seen in Belle II taggers.
Background and Positioning
Hadronic tagging supplies the application and the data structure. BABAR used hadronic tags to search for $B^+\to\tau^+\nu$ (Search for B+ -> tau+ nu decays with hadronic B tags). It measured $(1.8^{+0.9}_{-0.8}\pm0.4\pm0.2)\times10^{-4}$, and the background systematic of $\pm0.4$ was traced to finite simulation samples and mismodeled extra calorimeter energy.
The BABAR $B\to K\nu\bar\nu$ search (Search for the rare decay B -> K nu nu) reports a different set of systematics. Background-prediction uncertainties are 5% for $K^+$, 25% for high-momentum-transfer $K^+$, and 10% for $K^0$, and the signal-efficiency uncertainty is 14%. The 90% CL limits are $\mathcal B(B^+\to K^+\nu\bar\nu)<1.3\times10^{-5}$ and $\mathcal B(B^0\to K^0\nu\bar\nu)<5.6\times10^{-5}$.
Belle's $\Lambda$-plus-missing-energy search (Search for B0 meson decays into Lambda and missing energy with a hadronic tagging method at Belle) shows FEI in use for exotic searches. Its 90% CL limits range from $2.1$ to $3.8\times10^{-5}$.
The Belle II flavor-tagging study (B-flavor tagging at Belle II) found that a deep-network tagger lost its simulation advantage on data, reaching $28.8\%$ effective efficiency against $30.0\%$ for a boosted-tree tagger, because particle-identification likelihoods were mismodeled.
These papers play the motivating role. Each shows that channel-dependent, correlated mismodeling sets the precision floor. None chooses its tag selection with that floor in the objective. They also constrain the design. Systematics in these analyses are per-source and shared across modes, so a per-channel figure of merit that treats channels independently cannot be right.
The idea of cancelling systematics by construction is supplied by ratio measurements. Normalizing $B\to D\ell\nu$ to inclusive $B\to X\ell\nu$ (Measurement of |V_cb| and the Form-Factor Slope...) and normalizing $\tau$ to light-lepton modes (Observation of the Semileptonic Decays B -> D* tau- nu_tau and Evidence for B -> D tau- nu_tau) both do this. Our objective generalizes the idea: the optimizer rewards selections whose systematics are common to signal and normalization and therefore cancel.
The solver comes from constant-modulus optimization. Extreme Point Pursuit - Part I and Part II replace the unit-modulus set by its convex hull and subtract $\lambda\|x\|^2$. Large enough $\lambda$ forces solutions to extreme points, and error bounds give exact penalization at finite $\lambda$. The method is solved by projected gradient with a $\lambda$ homotopy. Alternating-minimization and deep-unfolding methods for hybrid beamforming (Alternating Minimization Algorithms for Hybrid Precoding in Millimeter Wave MIMO Systems; Joint Communications and Sensing Hybrid Beamforming Design via Deep Unfolding) solve related problems. They exploit factorized structure that our problem lacks, so we use EXPP and do not use them.
To our knowledge from the file, no prior work links tag selection, systematics-aware objectives and constant-modulus solvers.
Formulation
Index the cells by $i=1,\dots,N$, where $N$ is in the hundreds to low thousands. For each cell $i$, let $s_i$ and $b_i$ be the expected signal and background yields after signal-side selection. Let $\Sigma$ be the $N\times N$ covariance of absolute yield errors. We build $\Sigma$ as $\Sigma=\sum_k v_k v_k^\top + D$:
- each vector $v_k$ is the yield shift in every cell induced by one shared systematic source $k$, for example particle-identification efficiency, $\pi^0$ efficiency, extra-energy shape, or tag-calibration scale;
- $D$ is diagonal and holds each cell's own calibration uncertainty, which comes from the statistics of its control sample.
With keep variables $x\in\{0,1\}^N$, the approximate variance of the fitted signal rate is
$$\sigma^2(x)=\frac{\sum_i (s_i+b_i)x_i + x^\top \Sigma x}{\left(\sum_i s_i x_i\right)^2}.$$
Substituting $x=(1+u)/2$ gives $u\in\{-1,+1\}^N$, a real constant-modulus set. EXPP then solves
$$\min_{u\in[-1,1]^N}\ \sigma^2\!\left(\tfrac{1+u}{2}\right)-\lambda\|u\|_2^2,$$
increasing $\lambda$ by homotopy until $u$ is binary. The certificate uses a separate observation. For a fixed signal yield $t=\sum_i s_i x_i$, minimizing the numerator over the box is a convex quadratic program. Sweeping $t$ on a fine grid therefore gives a global lower bound on $\sigma^2$ over all binary selections, because the box contains them all. The bound is exact up to grid resolution, which we control by bracketing each grid interval. The reported gap is the distance between the EXPP selection and this bound.
Research Program
The work has four parts.
- Benchmark. We build a tag-selection benchmark at two levels. The first uses Belle II simulated FEI output for two signal channels, $B^+\to K^+\nu\bar\nu$ and $B^+\to\tau^+\nu$, which requires collaboration access. The second is a public surrogate whose cell yields and systematic vectors are parameterized to match the per-source magnitudes quoted in the BABAR papers above. This lets the method be studied outside the collaboration.
- Systematic vectors. We estimate $\Sigma$ from the standard variations used in these analyses: shifting particle-identification tables, smearing extra energy, and reweighting by tag-calibration factors.
- Solver and certificate. We implement EXPP and the yield-sweep bound, and report the optimality gap for every selection method, including the baselines.
- Validation under mismodeling. We test whether the selected cells actually behave better. Pseudo-data is built with injected mismodeling that the optimizer never sees. This includes a particle-identification shift of the kind reported in the Belle II flavor-tagging study and extra calorimeter noise clusters. On this pseudo-data we measure bias, coverage and expected upper limits.
Baselines
Do-nothing reference. The naive reference is the standard practice: one cut on the FEI signal probability, tuned to maximize $s/\sqrt{s+b}$ on simulation. Systematics are evaluated afterward. This is what the cited Belle and Belle II analyses do, and it defines the selection the community trusts. All methods use the same simulation, the same signal-side selection and the same $\Sigma$.
Strongest existing method. The strongest competitor is the published analysis selection, a tuned FEI cut followed by a binned profile-likelihood fit with nuisance parameters. We rerun it on our benchmark. The reported reference numbers available in the file are the BABAR 90% CL limits $\mathcal B(B^+\to K^+\nu\bar\nu)<1.3\times10^{-5}$ and $\mathcal B(B^+\to\tau^+\nu)=(1.8^{+0.9}_{-0.8}\pm0.4\pm0.2)\times10^{-4}$. The file reports no FEI-cut expected limit on Belle II simulation, so we regenerate that number ourselves under identical conditions.
Strong heuristic. A greedy forward selection minimizes the same $\sigma^2(x)$, adding the cell that most reduces it at each step. It gets the same objective, the same $\Sigma$, and unlimited passes.
General-purpose solver. A mixed-integer solver works on the quadratic-over-linear form under a matched wall-clock budget.
Ablation. Running EXPP with $\Sigma$ restricted to its diagonal isolates whether the correlated-systematics term is what matters.
What counts as winning. EXPP beats the baselines if it reduces expected total uncertainty, or the expected upper limit, by more than the spread across reruns with resampled simulation. It must do this while also keeping bias and coverage within nominal limits on mismodeled pseudo-data. Separately, the certificate is valuable if it shows that the FEI cut sits measurably far from the bound.
Why not just…?
Why not just keep the classifier cut and let the profile likelihood absorb systematics? The profile fit handles nuisance parameters on whatever cells are kept. It cannot undo the decision to keep cells whose systematics are large and shared. The BABAR $\tau\nu$ result shows this: the background systematic is comparable to the statistical error on a selection made for statistical power. Our first experiment compares the two directly, with the profile fit applied after both selections.
Why not just run greedy selection on the same objective? This is the real threat. With the right $\Sigma$, greedy may capture most of the gain. Greedy's weakness is that it cannot drop a cell once added. It will tend to add many mutually correlated high-yield cells early, cells whose combined systematic only appears later. We will report the greedy and EXPP gaps to the certified bound side by side.
If greedy closes the gap, we will say so. The contribution then shrinks to the systematics-aware objective plus the certificate, which tells analysts when greedy is already good enough. That remains useful, because no current selection comes with any optimality statement.
Why not just use continuous per-cell weights instead of keep/drop? Fractional weights are statistically fine, but each kept cell needs its own data-driven efficiency calibration. Weighting complicates both the calibration and the reweighting of the control sample. The $\lambda=0$ relaxation is exactly the continuous-weight solution, so we report it as an upper reference. If the binary selection matches it, the hard constraint costs nothing.
Proposed Experiments
Uncertainty decomposition. On the $K\nu\bar\nu$ and $\tau\nu$ benchmarks, we compare expected total uncertainty and expected 90% CL limit across all methods. Each result is decomposed into statistical, correlated-systematic and calibration parts.
Decisive robustness check. On pseudo-data with mismodeling the optimizer never saw, we compare bias and confidence-interval coverage. A systematics-aware selection should degrade less than the FEI cut when particle-identification and extra-energy modeling are perturbed. If it does not, the claim fails.
Certificate check. We report the gap to the yield-sweep bound for every method, and check whether the gap predicts out-of-sample uncertainty.
Scaling check. We study how EXPP runtime and gap behave as the number of cells grows from tens to thousands, with and without the homotopy schedule. This tests whether the radar-derived solver's guarantees carry over to this objective.
References
- Search for<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>τ</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mi>ν</mml:mi></mml:math>decays with hadronic<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>tags
- 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>
- Search for <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:math> meson decays into <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="normal">Λ</mml:mi></mml:math> and missing energy with a hadronic tagging method at Belle
- B-flavor tagging at Belle II
- Measurement of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mo stretchy="false">|</mml:mo><mml:msub><mml:mi>V</mml:mi><mml:mi>cb</mml:mi></mml:msub><mml:mo stretchy="false">|</mml:mo></mml:math>and the Form-Factor Slope in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mover accent="true"><mml:mi>B</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mo>→</mml:mo><mml:mi>D</mml:mi><mml:msup><mml:mi>ℓ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>ℓ</mml:mi></mml:msub></mml:math>Decays in Events Tagged by a Fully Reconstructed<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi></mml:math>Meson
- Observation of the Semileptonic Decays<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:msup><mml:mi>D</mml:mi><mml:mo>*</mml:mo></mml:msup><mml:msup><mml:mi>τ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>τ</mml:mi></mml:msub></mml:math>and Evidence for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:mi>D</mml:mi><mml:msup><mml:mi>τ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>τ</mml:mi></mml:msub></mml:math>
- Extreme Point Pursuit - Part I: A Framework for Constant Modulus Optimization
- Extreme Point Pursuit - Part II: Further Error Bound Analysis and Applications
- Alternating Minimization Algorithms for Hybrid Precoding in Millimeter Wave MIMO Systems
- Joint Communications and Sensing Hybrid Beamforming Design via Deep Unfolding