Recovering Lost Phases to Tell Resonances from Kinematic Cusps

Some particle-physics "discoveries" are stuck in limbo. The charged structure $Z_c(3900)$ shows up clearly as a bump in collision data. Yet four physically different explanations fit that bump equally well. It could be a genuine bound state, a "virtual" state, or a kinematic artifact produced by particles rescattering in a loop, known as a triangle singularity. The analysis in Amplitude analysis and the nature of the Z(3900) finds that all four scenarios reach the same fit quality, and that no two can be separated at better than 3σ. Our premise is that this is largely a phase retrieval problem. Detectors count events, so they measure only the squared magnitude of a complex quantum amplitude. The competing explanations may differ mostly in how the amplitude's phase changes with energy, which is exactly the part the measurement throws away.

Signal processing has a well-developed theory of when lost phases can be recovered from intensity-only measurements. Its recent extension to quaternions treats several coupled channels as one algebraic object. Quaternions are four-component numbers whose multiplication does not commute. Phase Retrieval of Quaternion Signal via Wirtinger Flow shows that such signals are recoverable from on the order of $d$ measurements, up to a global "right phase" factor. We propose to recast pole-versus-cusp discrimination as structured quaternion phase retrieval of a two-channel amplitude. From that formulation we would derive which additional observables would break the degeneracy, then test those predictions against the standard approach in particle physics: brute-force simulated pseudo-experiments.

Core Claim

The coupled-channel amplitude for $Z_c(3900)$ has two complex components, $A_J(E)$ for the $J/\psi\pi$ final state and $A_D(E)$ for $D\bar D^*$, as functions of energy $E$. We pack them into a single quaternion

$$ q(E) = A_J(E) + \overline{A_D(E)}\, j , $$

where $j$ is a quaternion unit and the bar denotes complex conjugation. The conjugate is not cosmetic. Right-multiplying $q$ by $e^{i\theta}$ then rotates both channels by the same phase, which is the one ambiguity physics actually allows. The naive encoding $A_J + A_D j$ would instead rotate the two channels in opposite directions. That would impose a symmetry the data do not have.

The remaining right-multiplications, by unit quaternions with a $j$ component, mix the two channels. The data must break these directions. Separate per-channel intensities break them, and so do interference regions of the Dalitz plot, the two-dimensional map of how momentum is shared among the three final-state particles. This mirrors how a zero-real-part constraint collapses the continuous phase ambiguity to a sign in the pure-quaternion variant of Phase Retrieval of Quaternion Signal via Wirtinger Flow.

The claim has three parts:

  1. Write each binned observable, whether a channel yield, an interference band, or a different collision energy, as a quadratic measurement of $q$ sampled on an energy grid.
  2. Use the rank and conditioning of the resulting Hessian to predict which measurement sets make the four scenarios distinguishable, and at roughly what event count. The Hessian here is the augmented quaternion Hessian of Convex Quaternion Optimization for Signal Processing: Theory and Applications, computed with Generalized Hamilton-Real (GHR) calculus. GHR calculus gives chain and product rules for differentiating real-valued losses directly with respect to quaternion variables, as formalized in A new proof of the generalized Hamiltonian–Real calculus.
  3. Validate the predictions with pseudo-experiments in a conventional K-matrix fitter. A K-matrix model is the standard way to build multi-channel amplitudes that respect probability conservation.

Research Program

Step 1: check the premise. We take the four fitted K-matrix scenarios from Amplitude analysis and the nature of the Z(3900), refitting them if published parameters are incomplete. We then measure how far apart the scenarios are in amplitude modulus versus phase across the threshold region. If they differ substantially in modulus, the phase-retrieval framing is wrong and we stop here.

Step 2: build the measurement model. Each observable becomes a measurement vector acting on the sampled $q$. The per-channel yields at $\sqrt s = 4.23$ and $4.26$ GeV from Determination of the Spin and Parity of the $Z_c(3900)$ give diagonal measurements. The Dalitz region where the $Z_c^+$ and $Z_c^-$ bands cross gives interference measurements. The neutral $D\bar D^*$ channel from Observation of a Neutral Structure near the D(D)over-bar* Mass Threshold adds a second energy pair and isospin partner. Its measured isospin decay ratio of $0.96 \pm 0.18 \pm 0.12$ is a usable constraint.

Step 3: identifiability analysis. We compute the Hessian of the intensity loss around each scenario's amplitude. We identify the directions that are flat modulo the physical $U(1)$ phase, and ask which candidate added measurement removes them.

Step 4: recovery algorithm. We run quaternion Wirtinger flow with spectral initialization, which starts the optimization from the top eigenvector of a data matrix built from the measurements. We compare it to multi-start fitting in the same parameterization.

Step 5: out-of-sample test. As a second case, we apply the same analysis to the $p\bar p$-threshold kink in $J/\psi\to\gamma\eta'\pi^+\pi^-$. There, Observation of an anomalous line shape reports that a Flatté fit and a coherent two-resonance model are about equally good. A Flatté fit is a resonance shape that accounts for a decay channel opening at threshold.

What prior work does and does not provide. The phase retrieval papers supply the algorithm and the sample-complexity intuition. Hypercomplex Phase Retrieval also supplies the octonion extension, which we do not need here. Both prove their guarantees for Gaussian random measurements. Our measurements are structured by kinematics. So we use the theory to organize the identifiability calculation, not as a transferable guarantee.

What constrains the design. Simultaneous diagonalisation of the covariance and complementary covariance matrices proves that no single unitary transform diagonalizes non-commuting quaternion second-order statistics. We therefore do not expect a clean decoupled basis for the channel covariances, and we keep the full augmented Hessian.

Baselines

Naive reference point. This is the status quo of Amplitude analysis and the nature of the Z(3900), whose four K-matrix scenarios all reach χ²/DOF ≈ 1.3 with model discrimination below 3σ on the combined BESIII and Belle projections. The simpler single-resonance description used in Determination of the Spin and Parity of the $Z_c(3900)$ sits alongside it. This is the number to beat. Any measurement design we recommend must yield above 3σ expected discrimination where the current design gives below 3σ, at matched integrated luminosity.

Strongest competitor: pseudo-experiment design search. This is the method particle physicists would actually use. Generate toy datasets from each scenario and fit them with a conventional complex K-matrix and many random restarts. Then search over candidate added observables for the one that maximizes expected discrimination. It needs no new mathematics. We run it on the same candidate observable list, the same toy statistics, and the same total CPU budget as our pipeline. Our method "beats" it under either of two conditions. The first is that it finds an equally discriminating design at a fraction of the compute. The second is that it predicts a design the budget-limited search misses.

Ablation 1: real-flattened parameterization. Here the amplitude is split into real and imaginary parts, with the identical physical constraints (common phase, isospin ratio) imposed explicitly. This tests whether quaternion algebra contributes anything beyond bookkeeping.

Ablation 2: wrong encoding. We use $A_J + A_D j$, without the conjugate, to test whether matching the ambiguity group to physics is the ingredient that matters.

Why not just…?

Why not just run pseudo-experiments, as is standard? This is the real threat. For a K-matrix fit with few parameters, brute-force toy studies may answer every question we pose. Our expected advantage is narrower than "better answers." It is a direct statement of which flat directions exist and which observable removes each one, which turns a blind search into a targeted one. If the matched-budget toy search finds the same design, we will report the identifiability calculation as an interpretive tool rather than a practical improvement.

Why not just fit complex amplitudes with more random restarts? For low-dimensional fits, multi-start probably reaches the global minimum reliably. We therefore do not claim an optimization advantage. Spectral initialization is compared to multi-start only as a secondary result. The proposal does not rest on it.

Why not just use Flatté or Breit-Wigner forms and wait for more data? Observation of an anomalous line shape shows that a highly significant threshold feature can stay ambiguous under standard shapes. More data in the same observables may not help if the ambiguity lives in the phase. Our experiments will show directly whether expected discrimination grows with event count under the current design, or only under designs that add interference measurements.

Why not just use quaternion-free real flattening? Ablation 1 settles this. The two are information-equivalent. If they perform identically, the quaternion encoding is a convenient language for deriving the ambiguity group, and we will say so.

Proposed Experiments

Sanity check. We first reproduce the synthetic result that quaternion Wirtinger flow recovers $d=100$ Gaussian quaternion signals exactly at $m/d \approx 9$ to $12$. Real concatenation needs about 16, per Phase Retrieval of Quaternion Signal via Wirtinger Flow. We then repeat the test with kinematically structured rather than Gaussian measurement vectors, to measure how much of that advantage survives.

Main experiment. We generate toy datasets from each of the four $Z_c(3900)$ scenarios. Statistics are scaled from the 1.92 fb⁻¹ BESIII samples up to ten times that. Each toy is fit under three measurement designs:

  1. per-channel projections only, which is the current practice;
  2. projections plus the Dalitz crossing region;
  3. projections plus the neutral-channel data at 4.226 and 4.257 GeV.

The metrics are expected pairwise discrimination significance, events needed to reach 3σ and 5σ, pole-position bias, and the fraction of fits that reach the global minimum.

Decisive check. The decisive check is whether the Hessian analysis correctly ranks the three designs before any toys are run, and whether its predicted event counts agree with the toy results within their statistical spread. Agreement plus a design that crosses 3σ where the status quo does not would establish the method. Disagreement would falsify the phase-retrieval framing.

Out-of-sample test. The $\eta'\pi^+\pi^-$ kink provides a second case with a different threshold ($p\bar p$) and a reported coupling ratio $g^2_{p\bar p}/g^2_0 = 2.31 \pm 0.37$ to reproduce.