Convex, Positive Background Shapes for Rare-Decay Likelihood Fits

When BABAR searches for a rare $B$-meson decay, the final number (a branching fraction or an upper limit) comes from an unbinned maximum-likelihood fit. This is a fit that scores every recorded collision event under a mixture of a "signal" shape and a "background" shape and adjusts how many events belong to each. The background shape is often a polynomial in kinematic variables, or a polynomial efficiency map over a Dalitz plot, the two-dimensional plane of invariant masses that fully describes a three-body decay. These polynomials cause three practical problems. Their coefficients are strongly correlated, so the fit can stall. They can dip below zero, so the likelihood is undefined. And the result depends on how many terms the analyst kept. When the signal is a few dozen events under a large background, all three problems feed straight into the measured yield.

We propose to write the background density as a sum of squares of polynomials, $p(x)=v(x)^\top Q\,v(x)$ with $Q$ a positive semidefinite matrix. The vector $v(x)$ holds polynomials that are orthonormal on the actual fit region, for example the curved Dalitz boundary or the $M_{\rm ES}$–$\Delta E$ window, rather than on a rectangle around it. With this choice the background is positive by construction and its normalization is a linear function of $Q$. The full extended likelihood in the signal yield and the background shape also becomes a concave function, so maximizing it is a convex problem with one global optimum. Choosing a basis orthonormal on the fit region is the step that keeps that convex program numerically solvable.

Core Claim

Write $x$ for an event's fit variables, $\Omega$ for the fit region, $P_s$ for the signal density (fixed from simulation, as is usual in rare searches), and $N_s$ for the signal yield. The extended log-likelihood is

$$\ell(N_s, Q) = \sum_{i=1}^{N} \log\!\big(N_s P_s(x_i) + v(x_i)^\top Q\, v(x_i)\big) - N_s - \operatorname{tr}(G Q), \qquad Q \succeq 0,$$

where $N$ is the number of events and $G=\int_\Omega v\,v^\top d\mu$ is the Gram matrix of the basis over the fit region under measure $\mu$. The expected background yield is $\operatorname{tr}(GQ)$. Each log term takes a function that is linear in $(N_s,Q)$, and the remaining terms are linear, so $\ell$ is jointly concave. The yields and the whole background shape therefore come from a single convex semidefinite program.

The profile likelihood is the curve used to set upper limits: for each fixed $N_s$, maximize over the background shape. Under this formulation every point on that curve is an exact global maximum, not the output of a local minimizer.

The numerical-conditioning literature says this only works with the right basis. Two results set the constraint. Approximate Volume and Integration for Basic Semialgebraic Sets computes integrals over regions defined by polynomial inequalities using moment-based semidefinite programs, and found that monomial bases give ill-conditioned moment matrices that make the solver fail, which Chebyshev bases fix. Mean Squared Error Minimization for Inverse Moment Problems rebuilds densities from moments under positivity constraints, and reports that orthonormal bases plus regularization are stable but that even orthonormal bases degrade at very high degree.

Our claim is that orthonormalizing on the true kinematic region, so that $G=I$, is what makes the convex program solvable and the yield estimate stable at the moderate degrees physics fits need, roughly degree 4 to 12 in two or three variables. Chebyshev polynomials on the bounding box are not orthogonal on a curved Dalitz region, and we expect them to recover part of the gain but not all of it.

Research Program

We would build the method in three parts.

Basis construction. For a given fit region $\Omega$, we would build the orthonormal basis by computing the Gram matrix $G$ of a starting Chebyshev basis. In two or three dimensions this can be done by high-order quadrature mapped to the region, or by the semialgebraic integration hierarchy of Approximate Volume and Integration for Basic Semialgebraic Sets, since Dalitz regions are defined by polynomial inequalities. We would then orthonormalize via a Cholesky factorization of $G$. This is the same inverse-moment-matrix object that underlies the Christoffel function used for streaming outlier detection in Leveraging the Christoffel function for outlier detection in data streams. We reuse its mathematics, not its outlier-detection task.

Fit solver. We would solve the concave program with a conic solver and, for large event counts, with a first-order method on the factor $Q=LL^\top$ that is warm-started from the conic solution. A fixed weight $w(x)$ can multiply the sum of squares to encode known edge behavior, for example the background vanishing at the $M_{\rm ES}$ kinematic endpoint. This keeps polynomial ringing near sharp edges, which the inverse-moment paper flags, away from the region where signal sits.

Analysis of conditioning. We would measure how the condition number of the fit's curvature matrix (the Hessian) and the solver's iteration count grow with degree, comparing monomial, box-Chebyshev, and region-orthonormal bases. We would also measure how this conditioning passes into the variance and bias of $N_s$.

Two applications follow. The first is background modeling in rare searches with event counts and fit variables like those of Search for the rare decay $B \to K \nu \bar{\nu}$ and Search for the Decay $B^+ \to K^+ \tau^\mp \mu^\pm$. The second is positive efficiency and background maps over Dalitz plots of the kind used in Constraints on the CKM angle γ in B0→D¯0K0 and B0→D0K0 from a Dalitz analysis of D0 and D¯0 decays to KSπ+π− and Dalitz plot analysis of the D+ → K- π+ π+ decay.

We are deliberately not proposing to replace the resonant amplitude model with a polynomial. Amplitude fits need complex phases for interference, and a nonnegative density cannot supply them. We only claim the components that are pure densities.

Baselines

The naive reference point is what many analyses do by default. The background is an ordinary polynomial in the monomial basis, and the full likelihood, signal yield included, is maximized with a local quasi-Newton minimizer started from a sideband fit. Negative-density events are handled by clipping. This shows how often current practice fails to converge, hits clipping, or lands on a non-global optimum.

The strongest polynomial competitor is the same nonconvex fit with Chebyshev polynomials on the bounding box. The integration paper above shows this basis already fixes much of the monomial ill-conditioning, and it is easy for practitioners to adopt. We also include the parametric shapes physicists actually use, an endpoint-type shape in $M_{\rm ES}$ and a low-order polynomial in $\Delta E$, as the domain-standard competitor.

The strongest nonparametric competitor is a fixed template, a histogram or kernel density estimate built from background simulation or data sidebands, with only the yields floated. It is positive and easy to use, and it absorbs no fit instability. Its weaknesses are template statistics and simulation mismatch. Search for the rare decay $B \to K \nu \bar{\nu}$ reports background-prediction systematics of 5% for $K^+$, 25% for high-$q^2$ $K^+$, and 10% for $K^0$, and a 14% signal-efficiency systematic. These are the reported levels a background-shape method must not worsen.

There is no reported result found on Lacuna for convergence-failure rates, pull widths, or shape-choice systematics of polynomial background fits in these analyses. We would therefore measure every baseline ourselves.

Fairness rules apply to all methods. Every method sees the same toy event samples and the same fit region. Every flexible method gets the same number of free background parameters. Degree is chosen by one rule, a held-out likelihood on the sidebands, applied identically to all.

Two ablations isolate the key ingredient. The first keeps the convex sum-of-squares formulation but uses box-Chebyshev and then monomial bases; this tests whether region-orthonormality matters. The second keeps the region-orthonormal basis but drops positivity, giving a plain polynomial with a local fit; this tests whether convexity and positivity matter.

The proposal beats the baselines if it reaches the same or higher likelihood in every toy. It must also have zero convergence failures, pulls (the fitted-minus-true yield divided by the reported error) with mean within 0.05 of 0 and width within 0.05 of 1, and correct 90% upper-limit coverage. Finally, the spread in $N_s$ across degree choices must be at most half that of the box-Chebyshev nonconvex fit.

Why not just…?

Why not just use Chebyshev polynomials in the existing fitter? This is the most serious threat. If box-Chebyshev nonconvex fits already converge reliably and give unbiased yields at the degrees analysts use, the gain from conditioning alone is small. The first ablation answers this directly. If the gap appears only for curved Dalitz regions, or only at higher degree, we will narrow the claim to those cases. Positivity and global optimality of the profile likelihood would remain as benefits Chebyshev polynomials alone do not give.

Why not just use a histogram or kernel template from simulation or sidebands? Templates are positive and stable. But they fix the shape to simulation, which is where the data–simulation systematics quoted above come from. They also smooth poorly where sideband events are few. Our toys include mismodeled templates, meaning backgrounds generated from a shape slightly different from the template, to measure the bias this produces against a shape that floats in the fit. If the template wins when there is no mismodeling and loses when there is, that is the expected and useful outcome.

Why not make the density positive with an exponential of a polynomial instead? An exponential gives positivity but makes the normalization a hard multidimensional integral. It also leaves the fit nonconvex in the yields. Mean Squared Error Minimization for Inverse Moment Problems reports 4,108 seconds for maximum-entropy reconstruction, which uses this exponential form, against under one second for the polynomial approach, at degree 8. We include an exponential-polynomial fit as an extra comparison on fit time and failure rate.

Proposed Experiments

Conditioning sweep. For the $D^+\to K^-\pi^+\pi^+$ Dalitz region, the $D^0\to K_S\pi^+\pi^-$ Dalitz region, and a rectangular $M_{\rm ES}$–$\Delta E$ window, we compute Hessian condition numbers and solver iterations as degree goes from 2 to 12 for the three bases. The prediction is that only the region-orthonormal basis keeps conditioning roughly flat over this range.

Toy ensembles. We generate thousands of pseudo-experiments matched to published conditions. One template has a high-purity Dalitz sample of 67,086 events with about 1.1% background, as reported by CLEO-c in Dalitz plot analysis of the D+ → K- π+ π+ decay. The other has rare-search conditions, with small injected signal yields chosen so that the expected limits sit near the reported BABAR $B^+\to K^+\nu\bar\nu$ limit of $1.3\times10^{-5}$ from Search for the rare decay $B \to K \nu \bar{\nu}$. Backgrounds are drawn from non-polynomial truths so that every method is somewhat misspecified. We report bias, pull width, failure rate, 90% CL coverage, and the yield systematic due to degree choice.

Profile-likelihood check. For each toy we compare the convex profile curve with the minimizer's profile curve. Any point where the convex value is higher shows the standard fit missed the global optimum; the count of such points and their effect on the limit is a decisive result in itself.

Real data, where available. Where preserved BABAR data or published distributions permit, we refit an existing background-dominated mass or $M_{\rm ES}$ distribution with both pipelines and check that the central values agree within statistical error while the shape-choice systematic shrinks. If preserved BABAR data or published distributions are not accessible, the claim rests on the toy ensembles, and we will state that.

The proposal fails if the box-Chebyshev nonconvex fit matches it on bias, coverage, and shape systematic across all regions and degrees. In that case the convex form is correct but unnecessary.