Root-Constrained Background Models for Low-Count Signal Searches

When physicists search for a rare signal, such as a faint bump in an energy spectrum from a decay that may not exist, they fit the data with two parts: a signal shape and a smooth "background" shape. With only tens to hundreds of events, a flexible background can quietly bend to absorb a real signal or to fake one. If it is made rigid instead, it fits poorly and biases the result. This note proposes controlling that flexibility through where the background polynomial's complex roots are allowed to sit. If every root is kept at least a distance $\delta$ from the fit window, two things follow. The density is positive on the window, and its log-curvature is capped by a quantity that can be compared directly with the width of the signal being sought. The flexibility knob then becomes a geometric quantity with a provable meaning: "this background cannot produce a bump narrower than X."

The idea transfers a tool from algebraic combinatorics, where confining a polynomial's roots to a region forces its coefficients to be positive, log-concave, or unimodal, into the setting where small event counts limit what an experiment can claim. The mathematics side supplies the mechanism and a positivity certificate. The physics side supplies the problem: false-signal rates and lost sensitivity in unbinned fits with few events.

Core Claim

Write the background density on a fit window $I$ as $p(x) = c\prod_{j=1}^{n}(x - z_j)$, where the $z_j$ are complex roots in conjugate pairs or real roots outside $I$. Differentiating $\log|p|$ twice gives

$$\left|\frac{d^2}{dx^2}\log p(x)\right| = \Big|\sum_{j=1}^{n}\mathrm{Re}\,\frac{1}{(x-z_j)^2}\Big| \le \frac{n}{\delta^2}, \qquad \delta = \min_j \operatorname{dist}(z_j, I).$$

A Gaussian signal of width $\sigma$ that adds a fraction $a$ to the local density changes the log-curvature at its peak by roughly $a/\sigma^2$, which follows from expanding $\log(b + a b\, e^{-x^2/2\sigma^2})$ to second order. The background therefore cannot imitate a signal of relative amplitude $a$ unless $a \lesssim n\sigma^2/\delta^2$. That makes $n\sigma^2/\delta^2$ an explicit, a priori bound on the "spurious signal," meaning the signal yield a fit reports when the data contain only background.

Real roots outside $I$ contribute only concave terms to $\log p$. All the wiggle comes from complex roots, and only from their distance to the window. This is the continuous analogue of the univariate fact in the research file that real-rootedness forces log-concavity of coefficients. The claim is that root-distance constraints give a better trade between bias and sensitivity than the two standard choices: positive Bernstein coefficients, and smoothness penalties.

Research Program

The work has three parts.

A positivity certificate. We will prove a quantitative "root location implies coefficient positivity" statement for the Bernstein basis on $I$. Bernstein polynomials are the standard positive basis for background polynomials on an interval. The statement to prove: if all roots lie outside an ellipse of size $\rho$ around $I$, then the degree-$m$ Bernstein coefficients are positive once $m \ge m(n,\rho)$. This mirrors the result in Coefficients and Roots of Peak Polynomials, which bounds the complex roots of peak polynomials in a region to prove positivity of their binomial-basis expansion. That paper supplies the proof template; we carry it from a combinatorial basis to a statistical one. The certificate places the root-constrained family inside a known Bernstein cone. That nesting lets us compare the two families at matched size and gives a cheap validity check during fitting.

A fitting procedure. We will fit directly in root coordinates (conjugate pairs written as $a_j \pm i b_j$), with $\delta$ either fixed from the signal width or profiled. The fits are extended unbinned likelihoods of the kind reviewed in Unbinned Maximum Likelihood Estimation for Signal Extraction. For falling spectra on $[0,\infty)$, the excluded region is a sector around the positive axis rather than an ellipse. This borrows the geometry of Fractionally log-concave and sector-stable polynomials, which uses root-free sectors to obtain correlation bounds for sampling. We use the same geometry for curvature bounds on a density. The product-of-roots form multiplies an exponential tail where needed. Exponential varieties constrains this choice: it shows that when a power of a polynomial serves as a partition function, the polynomial must be hyperbolic. We therefore stay with explicit normalization rather than claiming an exponential-family structure.

An empirical test. We will check whether the bound $n\sigma^2/\delta^2$ predicts the measured spurious signal, and whether, at a fixed spurious-signal tolerance, root-constrained backgrounds give tighter expected limits than the alternatives.

The motivating evidence is the observation in the research file that unbinned fits with few events show oscillation and edge artifacts, as in a study on meson decays. It also matters that systematic uncertainty in the background shape, rather than event count alone, limited the complete EXO-200 search.

Baselines

The naive reference point is an unconstrained polynomial background of the same degree fit by unbinned maximum likelihood. It is the "do nothing" option. It shows how often fits go negative or produce bumps, and every other method must beat it on spurious signal without losing interval coverage.

The strongest existing competitor is a Bernstein polynomial background with non-negative coefficients. This is common practice in collider searches, but the research file contains no citing paper for it and no reported spurious-signal figure on these benchmarks. It is the main competitor because it already guarantees positivity, so any advantage of our method must come from curvature control. We will run it at every degree from 2 to 8 on identical toy datasets, with the same signal model and the same optimizer restarts, and select degree by the same information criterion we use for our own model. We beat it if, at equal spurious signal (at most 20% of the statistical uncertainty on the signal yield), the root-constrained model gives a smaller median expected upper limit or a higher median discovery significance while keeping 68% and 90% coverage.

A third baseline is a binned Poisson fit with an equal number of parameters, which is the standard alternative to unbinned fitting.

The fourth baseline is a smoothness-penalized background with a penalty on the squared second derivative of $\log p$. Its penalty strength is tuned on held-out toy sets using the same number of tuning runs we spend choosing $\delta$.

Two ablations isolate the key ingredient. The first keeps the root parametrization but requires only that roots avoid $I$, with no distance floor. This tests whether positivity by roots is enough, or whether $\delta$ is doing the work. The second restricts the model to real roots, which forces a log-concave background. This tests whether complex roots are needed at all.

For the physical endpoint, the strongest reported number is the EXO-200 median 90% CL half-life sensitivity of $5.0 \times 10^{25}$ yr, from Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset. We cannot reproduce that analysis, because the data are not public. We use it only to set the realistic event-count regime of our toys. No result was found on Lacuna for spurious-signal rates of any background family on these datasets.

Why not just…?

Why not just use non-negative Bernstein coefficients, since they already guarantee positivity? A positive-coefficient polynomial can still have conjugate roots very close to the window. Such a background has a sharp local bump that can mimic a signal. The usual fix is to lower the degree, which makes the background too stiff. Positivity of coefficients says nothing about curvature at the signal's scale. Our certificate shows the two families are nested, so the comparison at matched degree in the experiments decides directly whether the root constraint helps. If Bernstein-positive fits already achieve the same spurious signal at equal limits, the proposal reduces to a reparametrization, and we will report that.

Why not just add a curvature penalty? This is the real threat. A penalty on $(\log p)''$ targets the same quantity we bound. The differences are that a penalty is soft, needs per-analysis tuning, and gives no a priori statement about which signal widths are protected. A root constraint gives a hard bound set by $\sigma$ before looking at the data. The penalty baseline gets the same tuning budget we use. If it matches us on both spurious signal and limits, our remaining contribution is only the analytic bound and the positivity certificate. We will state that plainly.

Why not just learn the background with a neural network or take it from simulation templates? Neural networks improved signal-versus-background separation in EXO-200 by about 25% per that analysis. That is a gain in classification, not a remedy for shape uncertainty, and background-model mismatch remained a dominant limitation in the same analysis. A learned or simulated template that is wrong in shape is exactly the failure a flexible fitted background exists to absorb. Our method sits downstream of any classifier and can be combined with one.

Proposed Experiments

Toy studies. All toys are synthetic, because the research file contains no public event-level physics data. The first toy family mimics the EXO-200 energy window: a Gaussian signal at 2457.83 keV with resolution σ/E = 1.15%, the value reported in the EXO-200 paper. The background is a smoothly falling continuum plus a nearby gamma line, with total event counts from 20 to 2000. The second family mimics low-count invariant-mass searches at the 4.5 fb⁻¹ scale of the BESIII semileptonic search, with backgrounds drawn from shapes outside every fitted family, so that every method is misspecified.

Metrics. For each method we measure:

  • the spurious signal under background-only toys, in units of statistical uncertainty;
  • coverage of the signal-yield intervals;
  • the fraction of fits that produce a negative or invalid density;
  • the median expected 90% CL upper limit and the median discovery significance under signal injection;
  • the local-to-global significance ratio from a scan over signal positions, measured with toys in the way the Telescope Array hotspot analysis measured its 5.1σ-to-3.4σ drop.

Decisive checks. The first decisive check is whether the measured spurious signal falls below the predicted $n\sigma^2/\delta^2$ across the grid of degree, $\delta$, $\sigma$, and event count. If it does not, the bound is wrong or loose in practice, and the method has no guarantee to offer. The second is whether, at matched spurious signal, the root-constrained model beats Bernstein-positive and curvature-penalized fits on expected limits. Neither result is assumed. We also expect a narrower look-elsewhere penalty, on the reasoning that a background unable to bend at scale $\sigma$ should shrink the effective number of independent trials in a position scan. That is a secondary prediction, and the same toys test it.