Precision measurements of inclusive B-meson decays at BABAR and CLEO report a handful of averages of the decay spectrum, called moments, rather than the full spectrum. Examples are the mean and spread of the lepton energy, or of the squared mass of the hadronic system, computed above a series of lepton-energy cuts. Theory predicts these moments directly, and fits to them determine quark masses and the quark-mixing parameter $|V_{cb}|$. The limiting error is not statistical. It comes from the decay model used to correct for the detector. In CLEO's neutrino-reconstruction analysis, the first hadronic-mass moment carries a statistical error of 0.014 GeV²/c⁴ and a model error of 0.109 GeV²/c⁴ (Moments of the B meson inclusive semileptonic decay rate using neutrino reconstruction). That model error is estimated by swapping in a few alternative decay models and recording how much the answer moves. It is a sensitivity study, not a bound.
We propose to replace part of that procedure with mathematics that holds for any decay model. A list of numbers can be the moments of a real, non-negative spectrum only if certain small matrices built from them are positive semidefinite. A Hankel matrix has entries $m_{i+j}$ that are constant along anti-diagonals. A positive semidefinite matrix has no negative eigenvalues, so it never assigns negative probability. Moments measured above ten nested cuts must all come from one spectrum, which links them through a chain of such matrix constraints. We use these constraints in two ways. First, we test whether published moment sets are consistent with any physical spectrum at all. Second, we compute the narrowest interval for each true moment that is compatible with the data and the detector response for every non-negative spectrum. That interval is a model-independent counterpart to the quoted model error. A second stage adds one piece of physics, the known masses of the narrow charm states $D$ and $D^*$. This gives the moment matrix a "low-rank-plus-continuum" structure, and we measure how far that structure tightens the interval.
Core Claim
Moments measured across a cut scan should be treated as one structured object, not as independent scalars. Write $\mu$ for the true spectrum in a kinematic variable $E$ (lepton energy, or $M_X^2$). The partial moments above a cut $c$ are $M_k(c)=\int_c^{E_{\max}} E^k\,d\mu(E)$. For two adjacent cuts $c_j<c_{j+1}$, the slab moments $s_k^{(j)}=M_k(c_j)-M_k(c_{j+1})$ are the moments of $\mu$ restricted to the interval $[c_j,c_{j+1}]$. The classical Hausdorff conditions then require
$$ H\big(s^{(j)}\big)\succeq 0,\qquad L_j\big(s^{(j)}\big)\succeq 0 \quad \text{for every slab } j. $$
Here $H(s)$ is the Hankel matrix $[s_{i+l}]$. $L_j$ is the corresponding "localizing" Hankel matrix built from the shifted sequence $(E-c_j)(c_{j+1}-E)E^k$, which enforces support on the slab.
The estimator minimizes $(m-\hat m)^\top \Sigma^{-1}(m-\hat m)$ subject to these cone constraints and to a linear moment-space detector map $\hat m_{\text{reco}} = A\, m_{\text{true}}$. Here $\hat m$ is the measured moment vector and $\Sigma$ is its full covariance across cuts and orders. The minimum value is a goodness-of-fit statistic against all non-negative spectra. Minimizing and maximizing a single true moment over the feasible set gives the certified interval.
What would be new is threefold: the cross-cut coupling, the use of positivity as a replacement for model variation rather than as a cosmetic fix, and the explicit test of whether positivity alone is informative enough to matter.
We are explicit about what does not carry over from the low-rank literature. These matrices are at most about 4×4 per slab. The sublinear-time Hankel algorithms of Sublinear Time Low-Rank Approximation of Hankel Matrices are therefore irrelevant for speed. We borrow from that work only the structural fact that a positive semidefinite Hankel matrix factors as $V D V^\top$, with $V$ a Vandermonde matrix over support points and $D\ge 0$ their weights. That factorization is what makes the second stage possible. Likewise, the NP-hardness of weighted low-rank approximation (Low-Rank Matrix Approximation with Weights or Missing Data Is NP-Hard) does not bite at this size, and the first stage is a small convex semidefinite program.
Research Program
Stage one: the convex positivity layer.
- Detector map. We build the moment-space detector map $A$ from simulated events conditioned on the true value of $E$. If the reconstructed variable's conditional moments are polynomial in the true value, which we will check, then reconstructed moments are a fixed linear combination of true moments.
- Where model dependence enters. The decay model then affects $A$ only through correlations with other phase-space variables, not through the spectrum shape itself. Measuring that residual dependence is part of the program.
- Outputs. We solve the constrained fit and the min/max interval problems with a standard semidefinite solver. Because the constraint set has boundaries, we calibrate the goodness-of-fit statistic by parametric bootstrap rather than assuming a chi-squared distribution.
Stage two: structured low-rank-plus-continuum decomposition.
- Why structure helps. A spectrum made of a few atoms (point masses) has a rank-deficient Hankel matrix whose rank equals the number of atoms. Relative to detector resolution, $D$ and $D^*$ are nearly atoms at known masses.
- Decomposition. We write the true moment vector as $V_{\text{known}}w + m_{\text{cont}}$. Here $V_{\text{known}}$ is the Vandermonde block at the $D$ and $D^*$ masses, $w\ge0$ holds their weights, and $m_{\text{cont}}$ is a continuum moment vector constrained only to the positivity cone. With fixed nodes this problem remains convex.
- Free nodes. We then let one extra node float, a broad $D^{**}$-like state, and fit it by the variable-projection method of Structured Low-Rank Approximation with Missing Data. That method solves for the linear weights in closed form inside an outer nonlinear search over the node. The fit asks whether the data themselves prefer an additional state.
- Joint rank constraints. Where rank constraints are imposed jointly on several slab matrices, we use the hybrid penalty method of A Hybrid Penalty Method for a Class of Optimization Problems with Multiple Rank Constraints. That method was built for exactly such coupled Hankel rank constraints, and its authors note that local minima make initialization matter. We therefore start from the stage-one solution and report results across many restarts.
Dimensional scope. We stay in one variable per fit, and at most two variables at low degree. The size of multivariate moment matrices grows combinatorially (The Combinatorial Scaling of Moment Matrices), and a full three-dimensional treatment like CLEO's fit over $q^2$, $M_X^2$ and $\cos\theta_{W\ell}$ is out of scope.
What positivity cannot do. Positivity alone cannot extrapolate below a cut. Adding non-negative mass below the cut only raises the partial rate, so the photon-energy extrapolation from 1.9 to 1.6 GeV in Measurement of the B → X_s γ branching fraction and photon energy spectrum using the recoil method receives only a lower bound from stage one. Any upper bound must come from stage-two structure. We will report this limitation openly rather than claim a general cure for truncated spectra.
Baselines
Naive reference: the moments as published. This is per-cut moments with Monte Carlo calibration factors, as in the BABAR semileptonic analysis (Measurement and interpretation of moments in inclusive semileptonic decays B̄ → X_c ℓ⁻ ν̄) and the CLEO lepton-energy analysis (Measurement of the B-meson inclusive semileptonic branching fraction and electron-energy moments). The model error is assigned by varying the decay model.
- Reference number. The concrete value to compare against is CLEO's first hadronic-mass moment at $E_\ell>1.0$ GeV: 0.456 ± 0.014 (stat) ± 0.045 (syst) ± 0.109 (model) GeV²/c⁴. At $E_\ell>1.5$ GeV the model error is 0.048 (Moments of the B meson inclusive semileptonic decay rate using neutrino reconstruction).
- BABAR moments. No reported result was found on Lacuna for individual BABAR moment values or their model errors.
- Fair comparison. On toy data, the naive baseline gets the same events and the same set of alternative decay models used to assign its model error.
- What counts as beating it. Equal-or-narrower intervals with correct coverage when the true generator lies outside the variation set. Matching the width while covering the truth where the baseline does not is already a win, because the baseline's error has no coverage guarantee.
Strongest unbinned method: Moment Unfolding (Moment Unfolding).
- What it is. It fits an exponential-polynomial reweighting with an adversarial classifier, and its learned coefficients recover true moments directly from smeared data without binning. It was demonstrated on simulated proton-proton jet observables and is reported to be about 100 times faster than iterative neural unfolding.
- Why it is the strongest competitor. It also skips binning and targets moments directly.
- Gap we exploit. It enforces no positivity of the moment matrix and no coupling across cuts.
- Fair comparison. We run it on identical toys, with the same simulated response sample and a matched hyperparameter search over the number of moments $n$, which the authors identify as their bias-variance knob.
- Reported precision. No reported result was found on Lacuna for its moment precision on $e^+e^-$ data.
- What counts as beating it. Lower mean-squared error on higher-order moments at BABAR-like sample sizes, together with valid intervals, which it does not produce natively.
Classical binned unfolding. We also run regularized binned matrix-inversion unfolding followed by moment computation, with the regularization strength tuned on held-out toys.
Ablations isolating each ingredient.
- Covariance-weighted projection versus eigenvalue clipping. This compares our projection against the cheap eigenvalue-clipping repair of a moment matrix in the style of Practical Sketching Algorithms for Low-Rank Matrix Approximation, which projects onto the semidefinite cone by discarding negative eigenvalues. It isolates whether the measurement covariance matters.
- Per-cut versus cross-cut constraints. This isolates the coupling between cuts.
- Positivity only versus positivity plus known nodes versus nodes without positivity. This isolates the low-rank structure.
Why not just…?
Why not just keep varying decay models, more aggressively? This is the real threat. Our certified interval is exactly the limit of varying over all non-negative spectra, so it may simply be far wider than any sensible model scan. If stage one gives an interval on the first $M_X^2$ moment much larger than 0.109 GeV²/c⁴ at CLEO-like statistics, positivity alone is uninformative for that observable. The contribution then rests entirely on stage two, where known resonance masses carry the information. The first toy experiment is designed to settle this before any data reanalysis.
Why not just clip negative eigenvalues after an ordinary estimate? The clip ignores the strong correlations between moments at neighboring cuts. It can move a well-measured moment in order to fix a poorly measured one. It also enforces nothing across cuts. The second ablation measures the bias this causes. If the clip matches the full projection in bias and coverage, we will say so and recommend it.
Why not just use Moment Unfolding or another neural unfolding method? Those methods estimate moments but give no guarantee that the moment set is physically realizable, and no worst-case interval. The authors also report that precision degrades as more moments are added (Moment Unfolding). Positivity is precisely the constraint that should stabilize the high-order end. The head-to-head toy comparison tests this. A positivity layer could also be stacked on top of Moment Unfolding's outputs, and we will test that combination.
Why not just low-rank everything with a truncated singular value decomposition? A truncated SVD of a moment matrix generally destroys the anti-diagonal (Hankel) structure, and averaging the structure back raises the rank again. Markovsky and Usevich show that structured approximation stays stable where singular value thresholding degrades with missing entries (Structured Low-Rank Approximation with Missing Data). At our tiny matrix sizes this is a correctness argument rather than a speed argument, and we include the SVD-plus-averaging route among the stage-two alternatives.
Proposed Experiments and Decisive Checks
Experiment 1: toy closure study.
- Generator. $D$ and $D^*$ lines, broad $D^{**}$ resonances and a non-resonant continuum, with fractions drawn from a wide family, including compositions outside the baseline's variation set.
- Detector. Resolution and acceptance fitted to the published descriptions.
- Sample sizes. Matched to the CLEO 9.4 fb⁻¹ sample and to the BABAR samples of 232 million $B\bar B$ pairs and roughly 680,000 recoil-tagged events.
- Metrics. Bias and root-mean-squared error of the first three moments, interval coverage and width, and the false-alarm rate of the consistency test.
Decisive check (Experiment 1). The stage-one interval width is compared against 0.109 and 0.048 GeV²/c⁴ at the two lepton cuts. A ratio near one means positivity alone can replace the model error. A ratio well above one shifts the claim to stage two. Coverage when the truth lies outside the variation set is the second decisive check.
Experiment 2: reanalysis of published moment scans. We apply the consistency test to the CLEO electron-energy moments at ten thresholds between 0.6 and 1.5 GeV, and to the BABAR hadronic-mass moment scans, wherever full covariances are published. The prediction is not that violations exist. A clean pass at all cuts is a useful negative result for those analyses. A failure localized to the lowest cuts would flag exactly the region the CLEO authors already identify as contaminated by cascade charm decays.
Experiment 3: stage two on the same toys and scans. We ask two questions. Does fixing the $D$ and $D^*$ nodes shrink the interval enough to fall below the conventional model error? Does a floating extra node land near the known $D^{**}$ masses, or wander, which would indicate the data cannot resolve it?
Access caveat. Event-level BABAR reanalysis would require collaboration access and is not assumed. Every claim above can be tested on toys and published tables alone.
References
- Moments of theBmeson inclusive semileptonic decay rate using neutrino reconstruction
- Sublinear Time Low-Rank Approximation of Hankel Matrices
- Low-Rank Matrix Approximation with Weights or Missing Data Is NP-Hard
- Structured Low-Rank Approximation with Missing Data
- A Hybrid Penalty Method for a Class of Optimization Problems with Multiple Rank Constraints
- Measurement of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:msub><mml:mi>X</mml:mi><mml:mi>s</mml:mi></mml:msub><mml:mi>γ</mml:mi></mml:math>branching fraction and photon energy spectrum using the recoil method
- Measurement and interpretation of moments in inclusive semileptonic decays<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:msub><mml:mi>X</mml:mi><mml:mi>c</mml:mi></mml:msub><mml:msup><mml:mi>ℓ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover></mml:math>
- Measurement of theB-meson inclusive semileptonic branching fraction and electron-energy moments
- Moment Unfolding
- Practical Sketching Algorithms for Low-Rank Matrix Approximation