How Stable Is the Abel Center Condition? Hyers–Ulam Constants for Polynomial Moments

Some differential equations have a "center": every nearby solution returns exactly to where it started, so all of them are periodic. If you perturb such an equation slightly, most of those periodic solutions disappear, but a few survive as isolated periodic orbits, called limit cycles. Batenkov and Binyamini showed that, for the polynomial Abel equation, how many survive is controlled by a finite list of integrals called moments. They also showed that the number of moments you need to check depends only on the polynomial degrees, not on the coefficients. That theorem assumes each moment is either exactly zero or not. In practice, whether a moment is computed by validated numerics or measured from a model, it is only known to be small. This note asks the Hyers–Ulam question, the classical "if something approximately satisfies an equation, is it close to something that satisfies it exactly?", about the center condition itself. Suppose the first $N$ moments are all at most $\varepsilon$. Is the perturbation then within $C\varepsilon$ of one that produces an exact center? How does the stability constant $C$ depend on the polynomials?

Our central claim is that $C$ cannot be bounded uniformly in terms of the degrees. It blows up exactly as the integrated drift polynomial $P$ approaches a polynomial that factors as a composition. We conjecture that the blow-up is at a predictable, generically linear, rate in the distance to that set. This would make the "definite polynomial" hypothesis in the existing cyclicity bound a quantitative quantity rather than a yes/no assumption. It would also give a recipe for turning "approximately zero moments" into a rigorous count of limit cycles outside a small radius that depends on $\varepsilon$.

Core Claim

Write the Abel equation as $dy/dx = p(x)y^2 + \epsilon\, q(x)y^3$. Let $P = \int p$, and define the moments $\tilde m_k(q) = \int_a^b P(z)^k q(z)\,dz$. Uniform upper bounds for the cyclicity of the zero solution of the Abel differential equation proves that the first nonzero moment has index at most $N(P,q) = 2 + \deg q + 3(\deg P-1)^2$. Consequently, the first $N$ moments vanish only if all of them vanish. For definite $p$, the paper also bounds the number of periodic solutions that branch off the zero solution by $5+\deg q+3\deg^2 p$. Our key observation is that, once $P$ is fixed, each moment depends linearly on $q$. The "moment-center" set $Z(P)$, meaning all $q$ of degree at most $m$ whose moments all vanish, is therefore the kernel of a finite matrix $M_P$ that sends the coefficients of $q$ to $(\tilde m_0,\dots,\tilde m_{N-1})$. The Hyers–Ulam statement then takes the form

$$\operatorname{dist}(q, Z(P)) \le C(P)\,\max_{k<N} |\tilde m_k(q)|, \qquad C(P) = 1/\sigma_{+}(M_P),$$

where $\sigma_+$ is the smallest nonzero singular value, measured in a fixed coefficient norm.

For a single fixed $P$, the existence of such a constant is elementary linear algebra. What is new is how $C(P)$ behaves uniformly as $P$ ranges over normalized polynomials of degree $d$. The rank of $M_P$ can drop on the set of composite $P = \tilde P\circ W$, where extra $q$ satisfy the composition condition. Where the rank drops, a singular value that is nonzero nearby must go to zero. Weyl's perturbation bound for singular values says this happens at least linearly in the distance from $P$ to the rank-drop set, and generically exactly linearly. We conjecture

$$C(P) \asymp \delta(P)^{-1}, \qquad \delta(P) = \operatorname{dist}(P, \text{composite locus of degree } d),$$

with constants depending only on $d$ and $m$. We also conjecture that the exponent rises above 1 on non-generic strata, where $P$ factors in more than one way.

The cyclicity consequence is the second half of the claim. The first-order displacement function of the perturbed Abel equation is a power series in the initial value $y_0$. Its coefficients are built from these moments. If $q$ lies within $\eta = C(P)\varepsilon$ of an exact center, the displacement function should be uniformly $O(\eta)$-close to one that vanishes identically. We conjecture that zeros in an annulus $r(\eta) \le |y_0| \le R$ are then still bounded by the degree bound. Any extra near-zero crossings would be confined to the inner disk, whose radius $r(\eta)$ shrinks as a power of $\eta$. That power comes from comparing the size of the $N$-th coefficient with the leading $O(\eta)$ part of the series. Under this conjecture, "approximately vanishing moments" yields a rigorous limit-cycle count, with an explicit dead zone whose size depends on $\varepsilon$ and $\delta(P)$.

Relation to Prior Work

Batenkov and Binyamini supply the method and the exact-case anchor. Their Fuchsian-equation analysis of the generating function $H(t)=\int q/(t-P)$ fixes $N$, and their definiteness hypothesis is the condition we want to quantify. They prove nothing about moments that are only small. Ulam Stability of Second-Order Periodic Linear Differential Equations with Periodic Damping gives the closest template from the Hyers–Ulam side. It computes the minimal Ulam constant for periodic linear oscillators, for instance $1/|pq|$ in the constant-coefficient case. It shows this constant diverges when a Riccati mean vanishes, which motivates our expectation of blow-up at a degeneracy. It works only with linear equations and analytic integrals, though. It never counts cycles and never ties the constant to degree data. Restricted stability and shadowing constrains the design. It recasts Hyers–Ulam stability as shadowing of pseudo-orbits, with constant $1/(M-M_\Psi)$, and shows in its Example 3.2 that stability fails when the expansion factor is $M=1$. A center has return-map multiplier exactly 1, so we deliberately put the stability question in the space of moments, where the relevant operator is a finite matrix, and not on trajectories. On the Computational Complexity of Limit Cycles in Dynamical Systems motivates the problem from the other side. It proves that approximate cycles always exist, and notes that they need not lie near any true limit cycle. Our dead-zone statement is a structured, Abel-specific answer to that gap. Efficient and Validated Numerical Evaluation of Abelian Integrals supplies tooling and a warning. Its rigorous approximation-plus-error-bound method corrected a published count for Johnson's quartic system from 26 to at least 24 limit cycles, which shows that unvalidated near-zero quantities produce wrong cycle counts.

Research Program

The work has three stages. In the first, we build $M_P$ exactly, in rational arithmetic, for $\deg P \le 6$ and $\deg q \le 8$. We also compute $\sigma_+(M_P)$ rigorously with interval arithmetic. Endpoints $a,b$ are treated as parameters subject to $P(a)=P(b)$, the condition under which this moment problem is posed. In the second stage, we map the composite locus for each degree, for example $P = W^2+\text{lower terms}$ when $d=4$ and $P=W^3$ or $W^2\circ$cubic when $d=6$. We then measure $C(P)$ along curves $P_\delta = \tilde P\circ W + \delta\, h$ that approach it from several directions $h$. The goal is to estimate the blow-up exponent on each stratum, and then to prove the linear rate on generic strata by perturbing the kernel of $M_P$. The likely tool is the explicit description of the kernel at composite $P$ given by the composition condition. In the third stage, we integrate the perturbed Abel equation with validated numerics, using interval Taylor methods in the spirit of the error-enclosed approximations of Bréhard et al. We do this for $q$ placed at controlled distances $\eta$ from $Z(P)$, and rigorously count zeros of the displacement function on annuli. This tests both the degree bound outside $r(\eta)$ and the predicted power law for $r(\eta)$.

Baselines

The naive reference point treats every moment with $|\tilde m_k| \le \varepsilon$ as exactly zero and applies the exact theory. It declares a center, or quotes $5+\deg q+3\deg^2 p$. We run it on the same $(P,q,\varepsilon)$ grid as our method and score it by how often its verdict disagrees with the validated cycle count. Our method beats it if it predicts the disagreement region, meaning large $C(P)\varepsilon$, with no false certificates.

The strongest existing method is the exact cyclicity bound of Uniform upper bounds for the cyclicity of the zero solution of the Abel differential equation. As a bound on small-amplitude cycles, that is the published number, $5+\deg q+3\deg^2 p$, for definite $p$. The file contains no empirical counts on test families, so we compute validated counts ourselves for the same families. That bound is silent whenever $p$ is near-composite or the moments are only approximately zero. "Beating" it means a certified count over an explicit amplitude range in exactly those cases, together with agreement with it in the exact case.

The instance-specific competitor is direct validated zero counting of the displacement function, the AbIntValid-style approach of Efficient and Validated Numerical Evaluation of Abelian Integrals. It gets the same interval-arithmetic budget per instance. On individual instances it will likely match or exceed our counts. We claim an advantage only in covering whole neighborhoods of $P$ with one certificate. We test this by counting how many instance runs are needed to cover a $\delta$-neighborhood versus one evaluation of $C(P)$.

The key ablation replaces $\delta(P)$, the distance to the composite locus, with a generic conditioning measure of $M_P$: the ratio of largest to smallest nonzero singular value, which could reflect Vandermonde-type ill-conditioning unrelated to composition. If that generic quantity predicts the observed blow-up as well as $\delta(P)$ does, the structural claim fails. For the stability constant itself, no reported result was found on Lacuna for any Hyers–Ulam constant of polynomial moment problems.

Why not just…?

Why not just compute the condition number of $M_P$ and stop? This is the real threat, because for fixed $P$ the existence of $C(P)$ really is linear algebra. What linear algebra does not give is where and how fast $C(P)$ diverges uniformly over a degree class, or how the divergence turns into a cycle count. The ablation above decides the matter. If the generic conditioning measure does as well as $\delta(P)$, we report the stability statement as a corollary and drop the structural claim.

Why not use the standard hyperbolic shadowing lemma on approximate periodic solutions? At a center the return-map multiplier is exactly 1. Restricted stability and shadowing shows by example that shadowing-type stability fails completely at expansion factor 1. Applying shadowing on trajectories therefore has nothing to work with. Moving to moment space is a response to this obstacle, not a choice of style.

Why not just run validated numerics on each system? It works one instance at a time. Efficient and Validated Numerical Evaluation of Abelian Integrals certifies single fixed systems and gives no statement over parameter neighborhoods. We use it as the ground truth in stage three and as a baseline, not as a substitute.

Why not expand the displacement function and work with the Bautin ideal directly? Batenkov and Binyamini report that the Bautin-ideal approach stalled because its generators grow uncontrollably with degree. That is why they moved to generating functions. Our matrix $M_P$ inherits their degree-only size $N$ and avoids the ideal entirely.

Proposed Experiments

The decisive check is a log–log plot of rigorously computed $C(P_\delta)$ against $\delta$ along approach curves, for degrees 3 through 6. It should show slope $-1$ on generic composite strata and steeper slopes on strata where $P$ factors in more than one way. If the slope is not $-1$, the linear-rate conjecture fails. If $C(P)$ shows no blow-up at all, so that it is uniformly bounded, the claim fails and uniform Hyers–Ulam stability would hold instead. Either outcome is informative. The second check uses validated integration of the Abel equation on a grid of $(\delta,\eta)$. For each point we record the certified number of periodic solutions on $[r, R]$ and find the smallest $r$ at which the count stays within $5+\deg q+3\deg^2 p$. We then fit $r(\eta)$ against the predicted power law. The third check is a sanity anchor. In the exact case $\eta=0$ with definite $p$, our counts must never exceed the published bound, and they should come close to it on the extremal families we construct. Any excess would point to an error in our pipeline before it says anything about the theory. All computations use exact or interval arithmetic, so the reported numbers are themselves certified. This is the standard that the 26-versus-24 correction in Johnson's system shows is necessary.