Do Kinematic Fits Hide a Systematic? Certifying Every Solution of the Photon-Energy Fit

Particle physicists who discover or measure a decay usually pass every candidate event through a kinematic fit. The fit nudges the measured particle energies and momenta, within their known resolutions, until they exactly obey energy–momentum conservation and the known masses of intermediate particles. Every standard fit is iterative: it starts from the measured values and moves downhill. When the constraints are nonlinear, as with a two-photon mass constraint, the fit's objective can have several local minima. No one checks whether the iterative fit ended in the best one. We propose to answer that question exactly. We rewrite the photon-energy kinematic fit as a small system of polynomial equations and solve it with a Gröbner basis solver. A Gröbner basis solver is an algebraic method, standard in computer vision, that finds all solutions of a polynomial system at once by turning it into an eigenvalue problem. With it we can enumerate every stationary point of the fit for every event. This tells us how often, and in which regions of phase space, the conventional fit converges to the wrong answer. It also tells us whether those failures shift reconstructed masses and yields enough to matter against today's residual systematic budgets.

The stakes are concrete. The CLEO-c measurement of the $\eta'$ mass reports $957.793 \pm 0.054\,(\text{stat}) \pm 0.036\,(\text{syst})$ MeV, and its largest systematic is photon energy reconstruction bias (Libby et al.). Photon energies are exactly the quantities a kinematic fit adjusts most. Any bias the fit itself introduces through non-global convergence would be mixed into that "photon bias" term, and today nobody could tell the two apart. The answer may well be that the effect is negligible. A certified negative result would still be useful, because it would remove one candidate from the list of unexplained residual systematics.

Core Claim

For kinematic fits where tracks are well measured and photon energies are the dominant free parameters, every stationary point of the constrained fit can be found by a Gröbner basis solver in well under a millisecond per event. This gives, for the first time, a per-event certificate: either the conventional fit reached the global minimum, or it did not. The contribution is the certificate and what it measures, namely the rate of non-global convergence and the mass and yield bias it induces. We do not claim that algebraic fitting is a better default fitter. The by-product is the number of real, physical stationary points for each photon assignment. We will test whether this count separates correct photon combinations from combinatorial ones, which is a recurring obstacle in multi-photon discovery channels.

Formulation

Fix the charged tracks at their measured momenta. This follows from the resolution hierarchy: tracks are measured far more precisely than photons at these energies. Also fix photon directions in the reduced problem. Let $\mathbf{E}=(E_1,\dots,E_k)$ be the fitted photon energies, $\hat{\mathbf{E}}$ their measured values, and $V$ their diagonal resolution matrix. The fit then solves

$$ \min_{\mathbf{E}} \; (\mathbf{E}-\hat{\mathbf{E}})^\top V^{-1}(\mathbf{E}-\hat{\mathbf{E}}) \quad \text{s.t.} \quad g_j(\mathbf{E}) = 0,\; j=1,\dots,c. $$

Here each $g_j$ is a conservation or intermediate-mass constraint. With fixed directions these constraints are polynomial in $\mathbf{E}$. For example, a two-photon mass constraint $2E_aE_b(1-\cos\phi_{ab}) = m^2$ is bilinear. Momentum conservation is linear in the $E_i$ once track energies are fixed. The first-order optimality conditions of the fit, which set the gradient of the objective plus multipliers times constraint gradients to zero, form a zero-dimensional polynomial system in $\mathbf{E}$ and the multipliers. Its real solutions with positive energies are exactly the candidate fit results. The extended version frees photon angles by writing each direction as a unit vector $\mathbf{u}$ with $\mathbf{u}^\top\mathbf{u}=1$, which keeps the system polynomial. The price is more unknowns.

The method comes from computer vision. Gröbner bases and elimination templates (Kukelova et al.) and the action-matrix eigenvalue route to all roots (Martyushev) supply the core. Automatic generators that emit C++ solvers from an ideal's structure (Ferranti et al.) supply the engineering. Choosing a basis for numerical conditioning (Beyond Grobner Bases) matters directly. Measured energies span a wide dynamic range, and ill-conditioned templates would produce spurious "extra minima" that are really floating-point artifacts. Two further works make root-count management a design principle. The asynchronous-track solvers reduce roots from 120 to 20 by choosing an approximation (Minimal Solvers for Full DoF Motion Estimation from Asynchronous Tracks), and the six-point multi-camera solver removes spurious solution branches through structural constraints (Six-Point Method for Multi-camera Systems with Reduced Solution Space). We borrow that discipline to keep the photon-fit systems small. The closest precedent for using elimination to replace local least-squares in parameter estimation is the global identification of polynomially parametrized dynamical systems by Komatsu et al.. That work treats the equations as noise-free. Our objective is explicitly a resolution-weighted residual, and our question is statistical: what bias does convergence failure induce in the aggregate?

Research Program

The first stage builds solvers for three concrete topologies. The first is $J/\psi \to \gamma\eta'$ with $\eta'\to\pi^+\pi^-\eta$ and $\eta\to\gamma\gamma$, which is mode B of the CLEO-c analysis and has two photons plus the radiative photon. The second is $\eta'\to\rho^0\gamma$, which has one photon and serves as a near-convex control. The third is a $\pi^0$-containing baryonic topology modeled on $B^+\to p\bar p\pi^+\pi^0$ (Study of B → p p̄ π π), where $\pi^0$ reconstruction and continuum suppression dominate the systematic budget. For each topology we derive the stationarity system, generate a template, and report its root count. We validate the solver against high-precision symbolic solutions on random instances.

The second stage is the measurement. We will run large simulated samples with parameterized detector smearing matched to published resolutions. Real CLEO-c or Belle data are not available to us, and we say so plainly. On each event we run the conventional iterative Lagrange-multiplier fit and the algebraic enumerator. We then record whether the two agree, the gap in $\chi^2$, and the shift in the reconstructed parent mass. We aggregate these into three quantities: the non-global convergence rate as a function of photon energy and opening angle, the induced bias on the fitted $\eta'$ mass peak, and the induced change in signal efficiency after a $\chi^2$ cut. We also inject a deliberate photon energy-scale miscalibration of a few per mille, to see whether fit failures grow with miscalibration. If they do, part of what is booked as "photon bias" is actually a fit-algorithm effect.

The third stage asks whether counting stationary points helps discovery. For each photon assignment in an event, the solver returns the number of real, positive-energy stationary points and their $\chi^2$ values. Wrong combinations, where the photons do not come from the hypothesized parent, may more often have no feasible solution near the measurement, or several competing ones. We will test whether adding this count to a standard selection improves separation of correct from combinatorial candidates in the $\pi^0$ topology.

A short transfer check addresses the metasurface direction. Surfaces that simultaneously transmit and reflect radio waves obey a coupled phase constraint, $|\theta_r-\theta_t|\in\{\pi/2,3\pi/2\}$ with $\beta_r^2+\beta_t^2\le1$ (Mu et al.). Mu et al. optimize these phases element by element with alternating optimization, and Zhang et al. use penalty and semidefinite relaxation. Neither enumerates solutions. Substituting unit-circle variables ($c^2+s^2=1$) makes each per-element update a tiny polynomial system. We will use it only as a second test of the claim that certified enumeration can show when a local update misses its global optimum. It is not a headline result.

Baselines

The naive reference is no kinematic fit: the raw invariant mass from measured energies. It sets the resolution floor that any fit must improve, and it cannot suffer convergence failure.

The strongest existing method is the conventional iterative constrained fit of the kind used in the CLEO-c $\eta'$ analysis (Libby et al.), run with standard settings: measured-value initialization, convergence tolerance, and iteration cap. The published measurement's error budget, $\pm0.054$ stat and $\pm0.036$ syst MeV, is the scale against which any induced bias is judged. No reported result was found on Lacuna for the rate at which such fits reach non-global minima, and that gap is exactly what we fill. Both methods receive identical events, identical resolution matrices, and identical constraint sets. "Beating" this baseline in the sense that matters means showing a mass bias or efficiency shift that is resolvable against the 36 keV systematic scale, or certifying that it lies well below it.

The second competitor is multi-start local fitting, where the iterative fit is restarted from many perturbed initializations and the lowest $\chi^2$ is kept. It is run with a start budget chosen to match the algebraic solver's wall-clock time per event. On the solver side, reported Gröbner runtimes of 81–120 μs for comparable-size systems (Minimal Solvers for Full DoF Motion Estimation from Asynchronous Tracks) guide this budget.

The ablation removes basis selection, so that we run a plain Gröbner template without the conditioning improvements of Beyond Grobner Bases. This isolates how many apparent "extra minima" are numerical artifacts rather than physics.

Why not just…?

Why not just use multi-start local fits? This is the real threat. If failures are rare and their basins are wide, a modest number of restarts will find the global minimum just as often. The difference is certification. Multi-start can never report that no better minimum exists, and its miss rate on narrow basins is unknown by construction. The equal-time comparison settles the question directly. If multi-start matches enumeration on every event of every topology, we will report that, and recommend restarts as the cheap fix. The certificate would still be what established that the fix suffices.

Why not just improve photon calibration, since it is already the dominant systematic? Calibration and fit behavior are entangled. A miscalibrated photon moves the measurement toward regions where the constraint surface curves more sharply. The injected-miscalibration experiment tests whether fit failures scale with miscalibration. If they do not, calibration work is the right priority and this project reports a clean null.

Why not just use generic global methods such as homotopy continuation or branch-and-bound? They would also enumerate or bound the minima, and are a reasonable cross-check on small instances. However, they are typically much slower per event than a specialized template. Per-event speed matters when the measurement requires simulated samples of tens of millions of events. We will use a generic global method offline as the ground truth for validating the template.

Decisive Checks

The first test is agreement: on random instances, the template must recover every root found by high-precision symbolic solving. The second is the core measurement: the non-global convergence rate, and the resulting shift of the fitted $\eta'$ peak, for mode B versus the one-photon control mode. A near-convex control that shows no failures while the two-photon mode shows failures would confirm that the effect comes from the constraint geometry. A shift below a few keV would support a certified null. A shift comparable to 36 keV would support a hidden systematic. The third is the miscalibration-scaling test. The fourth is whether the stationary-point count improves correct-combination separation in the $\pi^0$ topology, measured as the change in background rejection at fixed signal efficiency. If all four come back negative, the project still produces a verified bound and a reusable algebraic fitter.