betaFeedback

WorksMathematics

Stability in the Non-Neutral and Neutral Regimes of the Holonomy-Based Inverse Problem in the Class OC\mathcal{OC}

Abstract

We study anchor reconstruction stability in return-map inverse problems for class OC\mathcal O_C. Local identifiability is governed by the first nonvanishing derivative of the holonomy map H=Tγ−IdH = T_γ- \text{Id}. In the non-neutral case, we prove local Lipschitz stability and multiplier estimates. In the neutral regime, using Faà di Bruno equations and Malgrange preparation, we derive quantitative 1/k1/k-Hölder estimates for order-kk anchor perturbations. We also establish finite jet stability, real-analytic dependence, and global C0C^0-stability of order O(εF1/k)O(\varepsilon_F^{1/k}). Numerical benchmarks on circular-core return geometries validate these theoretical convergence rates and square-root laws.

2020 Mathematics Subject Classification: Primary 37C25; Secondary 34A12, 35R30, 53A05.

Introduction

The return-map construction in the class OC\mathcal{O}_C associates to a known convex core CC and a surrounding admissible normal-graph boundary a discrete map F:∂C→∂CF : \partial C \to\partial C. The forward geometry and regularity of this map were developed in [1, 3]. The inverse problem studied in [2] asks how much of the unknown thickness function dd, and hence of the outer boundary, can be reconstructed when CC and FF are known.

A central result of that inverse theory is that exact non-uniqueness reduces to a scalar anchor value. The inverse system transports a trial value a=d(c0)a = d(c_0) along paths of M:=∂CM := \partial C, and closed paths generate nonlinear holonomy maps TγT_\gamma. The true anchor a∗a_* is a common fixed point. When

H(a):=Tγ(a)−aH(a) := T_\gamma(a) - a

satisfies H′(a∗)≠0H'(a_\ast) \ne0, the anchor is locally isolated by the ordinary transverse-root mechanism. The singular regime begins when H′(a∗)=0H'(a_\ast) = 0: linearized holonomy becomes neutral, but higher vertical variations of the nonlinear transport may still retain identifying information.

The word “holonomy” also appears in several nearby geometric inverse problems, but the data and unknowns are different. Cekić and Lefeuvre recover unitary connections, modulo gauge, from traces of connection holonomies along primitive closed geodesics and prove local injectivity near generic connections [7]; their companion stability paper establishes Hölder-type estimates for that problem [6]. More broadly, transport equations, scattering data, non-Abelian X-ray transforms, boundary rigidity and range questions form a central part of contemporary geometric inverse problems [12, 4, 5, 9]. In the present paper, by contrast, the holonomy is a scalar nonlinear transport in the fibre variable representing the unknown thickness anchor. This distinction is important: the Hölder exponents below arise from finite-order singularity of a scalar anchor fixed point, while their coefficients are generated by the return-map transport geometry.

Contribution and positioning. The 1/k1/k exponent itself should not be confused with the novelty of the paper. Multiple roots are classically ill-conditioned and their numerical treatment requires multiplicity-aware analysis; see, for example, [13]. The purpose here is to close, within the return-map inverse problem, the problem-specific chain

return-data perturbation ⟹\Longrightarrow finite holonomy-jet perturbation ⟹\Longrightarrow isolation order, ⟹\Longrightarrow anchor stability ⟹\Longrightarrow boundary reconstruction stability.

The genuinely new ingredients are the geometric realization of the multiplicity through the exact holonomy jet, its finite-jet stability under noisy return data, the Malgrange-prepared local polynomial unfolding for finite-dimensional smooth families, deterministic multi-loop aggregation, and the propagation of the anchor estimate to the full reconstructed thickness. The scalar multiple-root exponent is classical; its geometric coefficient, observability, and inverse-problem consequences are the contributions developed below.

Define

kγ(a∗):=min⁡{k≥1:H(k)(a∗)≠0},k_\gamma(a_\ast) := \min\{k \ge1 : H^{(k)}(a_\ast) \ne0\},

with kγ(a∗)=+∞k_\gamma(a_\ast) = +\infty if all derivatives vanish. The main results are as follows.

  1. In the non-neutral case we prove a local one-loop stability theorem, an observable multiplier bound, and quantitative transfer of return-map errors to holonomy errors.

  1. We derive the exact second holonomy variation and the explicit third variation, and prove the full finite variational hierarchy

Zn′=∑r=1nGyrBn,r(Z1,…,Zn−r+1),Z'_n = \sum_{r=1}^{n} G_{y^r} B_{n,r}(Z_1,\ldots,Z_{n-r+1}),

where Bn,rB_{n,r} are partial Bell polynomials.

  1. If the first nonzero holonomy derivative has order kk, every sufficiently localized perturbed zero satisfies a quantitative 1/k1/k-Hölder estimate. Odd and even orders have different existence mechanisms.

  1. For every finite-dimensional C∞C^\infty family of admissible return data passing through an order-kk holonomy singularity, the Malgrange preparation theorem gives a local factorization into a

smooth unit and a monic degree-kk polynomial in the anchor displacement. Thus the perturbed anchor set is locally governed by a finite polynomial unfolding, while our stability estimates quantify the conditioning of its roots.

(v) For every fixed finite pp, C0C^0 perturbation of the observable return components controls the holonomy jet (Tγ,Tγ′,…,Tγ(p))(T_\gamma,T_\gamma',\ldots,T_\gamma^{(p)}) on uniformly separated anchor intervals. This makes the higher-order stability coefficients observable from noisy return data.

(vi) We prove a deterministic multi-loop estimate at the minimal isolation order, including approximate common fixed points.

(vii) We prove real-analytic dependence of TγT_\gamma on the initial anchor on uniformly separated intervals. Thus kγ(a∗)=+∞k_\gamma(a_\ast)=+\infty implies that TγT_\gamma is locally the identity near a∗a_\ast.

(viii) Under uniform separation from the singular fibre set, the local anchor estimate propagates to a global estimate

∥d^−d∥C0(M)=O(εF1/k)\lVert\hat{d}-d\rVert_{C^0(M)}=O\left(\varepsilon_F^{1/k}\right)

for every fixed finite order k≥2k\ge2.

(ix) We validate both the non-neutral and neutral regimes on genuine return geometries of the unit circular core. The observed convergence rates are approximately 1.0002 in the non-neutral case and 0.4960 in the tuned neutral case; the corresponding global thickness rates are 1.0002 and 0.4953.

The geometric origin of the higher-order coefficients beyond the explicit vertical variational formulas remains subtler. In particular, invariant small-loop formulas on a general non-umbilic core must retain the Levi-Civita connection terms of the principal frame. We therefore separate the proved holonomy-jet theory from those remaining geometric genericity questions.

The paper is organized as follows. § 2 recalls the nonlinear transport formulation. § 3 treats the non-neutral regime. § 4 develops the variational holonomy jet, its Malgrange-prepared finite-dimensional unfoldings, and the analytic flat regime. § 5 proves finite-order isolation and parity-sensitive existence. § 6 transfers return-data perturbations to finite holonomy jets and develops observable coefficients. § 7 gives deterministic multi-loop aggregation. § 8 propagates anchor errors globally. § 9 records the remaining invariant geometric questions. § 10 contains executed algebraic and geometric numerical experiments, including a TikZ visualization of the tuned neutral geometry.

Framework and nonlinear transport

We use the notation and standing hypotheses of the inverse problem developed in [2], within the class OC\mathcal{O}_C introduced in [1]. Let C⊂RNC\subset\mathbb{R}^N be compact convex with C3C^3 boundary M:=∂CM:=\partial C. Let d:M→(0,∞)d:M\to(0,\infty) be of class C2C^2, let F:M→MF:M\to M be the associated return map, and let SC:=−Dν(c)∣TcM\mathcal{S}_C:=-D\nu(c)|_{T_cM} be the shape operator of MM. We denote by uF,qFu_F,q_F the tangential and normal components of F(c)−cF(c)-c, and by TγT_\gamma the holonomy transport along a closed C1C^1 loop γ\gamma based at c0c_0.

For a fixed closed loop γ:[0,1]→M\gamma:[0,1]\to M based at c0c_0, we write the scalar transport equation as

ya′(t)=Gγ(t,ya(t)),ya(0)=a,(1)y_a'(t)=G_\gamma(t,y_a(t)),\qquad y_a(0)=a, \tag*{(1)}

where

Gγ(t,y)=A(t)−yB(t)y−q(t).(2)G_\gamma(t,y)=\frac{A(t)-yB(t)}{y-q(t)}. \tag*{(2)}

with

A(t)=⟨uF(γ(t)),γ˙(t)⟩,B(t)=⟨Sγ(t)uF(γ(t)),γ˙(t)⟩,q(t)=qF(γ(t)).(3)A(t)=\langle u_F(\gamma(t)),\dot{\gamma}(t)\rangle,\qquad B(t)=\langle S_{\gamma(t)}u_F(\gamma(t)),\dot{\gamma}(t)\rangle,\qquad q(t)=q_F(\gamma(t)). \tag*{(3)}

Set

P(t)=A(t)−q(t)B(t)=⟨(I−q(t)Sγ(t))uF(γ(t)),γ˙(t)⟩.(4)P(t)=A(t)-q(t)B(t)=\langle(I-q(t)S_{\gamma(t)})u_F(\gamma(t)),\dot{\gamma}(t)\rangle. \tag*{(4)}

On the domain y−q(t)>0y-q(t)>0,

∂yGγ(t)(t,y)=−P(t)(y−q(t))2,(5)\partial_yG_{\gamma(t)}(t,y)=-\frac{P(t)}{(y-q(t))^2}, \tag*{(5)}
∂y2Gγ(t)(t,y)=2P(t)(y−q(t))3,(6)\partial_y^2G_{\gamma(t)}(t,y)=2\frac{P(t)}{(y-q(t))^3}, \tag*{(6)}

and more generally, for r≥1r\ge1,

∂yrGγ(t)(t,y)=(−1)rr!P(t)(y−q(t))r+1.(7)\partial_y^rG_{\gamma(t)}(t,y)=(-1)^r r!\frac{P(t)}{(y-q(t))^{r+1}}. \tag*{(7)}

The holonomy displacement is

H(a):=Tγ(a)−a,Tγ(a)=ya(1),(8)H(a):=T_\gamma(a)-a,\qquad T_\gamma(a)=y_a(1), \tag*{(8)}

with H(a∗)=0H(a_*)=0 at the true anchor a∗:=d(c0)a_*:=d(c_0). We set

δγ:=∣H′(a∗)∣.(9)\delta_\gamma:=\lvert H'(a_*)\rvert. \tag*{(9)}

The isolation order is

kγ(a∗):=min⁡{k≥1:H(k)(a∗)≠0},(10)k_\gamma(a_*):=\min\{k\ge1:H^{(k)}(a_*)\ne0\}, \tag*{(10)}

with kγ(a∗)=+∞k_\gamma(a_*)=+\infty if all derivatives vanish. When k<∞k<\infty, we write

mk:=∣H(k)(a∗)∣>0.(11)m_k:=\lvert H^{(k)}(a_*)\rvert>0. \tag*{(11)}

We use two error norms on the holonomy displacement:

ε:=∥H^−H∥C0(I‾),εT:=∥H^−H∥C1(I‾).(12)\varepsilon:=\lVert\widehat{H}-H\rVert_{C^0(\overline{I})},\qquad\varepsilon_T:=\lVert\widehat{H}-H\rVert_{C^1(\overline{I})}. \tag*{(12)}

Evidently ε≤εT\varepsilon\le\varepsilon_T. The non-neutral regime requires C1C^1 control and is stated with εT\varepsilon_T. The neutral regime requires only C0C^0 control plus a Taylor remainder, and is stated with ε\varepsilon; C1C^1 control is invoked only when existence of the perturbed fixed point is established by a contraction.

For j≥1j\ge1, we set

Mj:=∥H(j)∥C0(I‾).(13)M_j:=\lVert H^{(j)}\rVert_{C^0(\overline{I})}. \tag*{(13)}

Uniform separation and persistence under perturbation

All stability statements below require the horizontal lifts to remain at positive distance from the singular fibre set y=qF(γ(t))y=q_F(\gamma(t)). We first record that this quantitative admissibility condition is stable under sufficiently small C0C^0 perturbations of the observed return components.

Definition 2.1 (Uniform separation). Let I0⊂RI_0\subset\mathbb{R} be a compact anchor interval and γ:[0,1]→M\gamma:[0,1]\to M a C1C^1 curve. We say that the exact transport is uniformly separated on I0I_0 with margin η>0\eta>0 if, for every a∈I0a\in I_0, the lift yay_a exists on [0,1][0,1] and

ya(t)−qF(γ(t))≥ηfor all (a,t)∈I0×[0,1].y_a(t)-q_F(\gamma(t))\ge\eta\qquad\text{for all }(a,t)\in I_0\times[0,1].

Lemma 2.2 (Persistence of uniform separation). Assume Definition 2.1 holds with margin η>0\eta>0. Set

εF:=∥u^F−uF∥C0(γ)+∥q^F−qF∥C0(γ).\varepsilon_F := \lVert\hat{u}_F-u_F\rVert_{C^0(\gamma)}+\lVert\hat{q}_F-q_F\rVert_{C^0(\gamma)}.

Then there exists εF∗>0\varepsilon_F^*>0, depending only on η\eta, I0I_0, γ\gamma and uniform bounds for the exact transport on the compact family of exact solution graphs, such that, whenever 0≤εF≤εF∗0\leq\varepsilon_F\leq\varepsilon_F^*, every perturbed lift y^a\hat{y}_a with a∈I0a\in I_0 exists on [0,1][0,1] and

y^a(t)−q^F(γ(t))≥η2for all (a,t)∈I0×[0,1].\hat{y}_a(t)-\hat{q}_F(\gamma(t))\geq\frac{\eta}{2}\qquad\text{for all }(a,t)\in I_0\times[0,1].

Moreover,

sup⁡a∈I0sup⁡0≤t≤1∣y^a(t)−ya(t)∣≤Cγ,I0εF\sup_{a\in I_0}\sup_{0\leq t\leq1}\lvert\hat{y}_a(t)-y_a(t)\rvert\leq C_{\gamma,I_0}\varepsilon_F

for a constant independent of the perturbation once εF≤εF∗\varepsilon_F\leq\varepsilon_F^*.

Proof. Because I0I_0 is compact and the exact lifts depend continuously on (a,t)(a,t), their union is contained in a compact fibre interval. Fix the tube

T:={(a,t,y):a∈I0, 0≤t≤1, ∣y−ya(t)∣≤η/4}.\mathcal{T}:=\{(a,t,y):a\in I_0,\ 0\leq t\leq1,\ \lvert y-y_a(t)\rvert\leq\eta/4\}.

If εF≤η/4\varepsilon_F\leq\eta/4, then throughout T\mathcal{T} one has

y−q^F(γ(t))≥ya(t)−qF(γ(t))−η4−εF≥η2.y-\hat{q}_F(\gamma(t))\geq y_a(t)-q_F(\gamma(t))-\frac{\eta}{4}-\varepsilon_F\geq\frac{\eta}{2}.

Hence the exact and perturbed rational vector fields are uniformly bounded and Lipschitz in yy on this tube, and there is a constant C0C_0 such that

∣G^(t,y)−G(t,y)∣≤C0∣γ˙(t)∣εFon T.\lvert\hat{G}(t,y)-G(t,y)\rvert\leq C_0\lvert\dot{\gamma}(t)\rvert\varepsilon_F\qquad\text{on }\mathcal{T}.

For fixed a∈I0a\in I_0, let τ\tau be the supremum of times for which the perturbed solution exists and remains in the tube, equivalently ∣y^a−ya∣≤η/4\lvert\hat{y}_a-y_a\rvert\leq\eta/4, on [0,τ][0,\tau]. On this interval the standard Gronwall comparison gives

∣y^a(t)−ya(t)∣≤C0ℓ(γ)eL0ℓ(γ)εF,0≤t≤τ,\lvert\hat{y}_a(t)-y_a(t)\rvert\leq C_0\ell(\gamma)e^{L_0\ell(\gamma)}\varepsilon_F,\qquad0\leq t\leq\tau,

with a uniform Lipschitz constant L0L_0 on T\mathcal{T}. Choose εF∗\varepsilon_F^* so that the right-hand side is strictly smaller than η/4\eta/4 and also εF∗≤η/4\varepsilon_F^*\leq\eta/4. A first exit at τ<1\tau<1 is then impossible. Thus τ=1\tau=1, the displayed difference estimate holds on the full interval, and

y^a(t)−q^F(γ(t))≥η−η4−η4=η2.\hat{y}_a(t)-\hat{q}_F(\gamma(t))\geq\eta-\frac{\eta}{4}-\frac{\eta}{4}=\frac{\eta}{2}.

All constants are uniform in a∈I0a\in I_0, which completes the proof. □\square

Remark 2.3 (Conditioning near the singular fibre). The vertical Lipschitz constants below scale like η−2\eta^{-2}. Consequently the Gronwall amplification factors deteriorate rapidly as the exact separation margin tends to zero. This blow-up records genuine ill-conditioning near the non-admissible set y=qFy=q_F, rather than a proof artefact.

Remark 2.4 (Certifying separation near the true anchor). For a concrete geometry it is enough first to certify a positive separation margin along the true horizontal section y∗(t)=d(γ(t))y_*(t)=d(\gamma(t)). Continuity of the transport with respect to the initial anchor then gives a compact interval I0I_0 around a∗a_* on which Definition 2.1 holds, possibly with a smaller margin. In § 10, the interval enclosures of the circular-core geometry provide precisely such quantitative geometric margins; they are used as evidence for a well-separated true section, not as a substitute for Definition 2.1 on an arbitrary anchor interval.

Non-neutral stability

Throughout this section we assume

δγ:=∣H′(a∗)∣>0.\delta_\gamma:= |H'(a_*)| > 0.

This is the non-neutral regime already underlying the first-order holonomy mechanism in [2]. The purpose here is quantitative stability under perturbed return data.

Theorem 3.1 (Local one-loop stability in the non-neutral regime). Let H=Tγ−Id⁡H = T_\gamma-\operatorname{Id} and H^=T^γ−Id⁡\widehat{H} = \widehat{T}_\gamma-\operatorname{Id} on an interval II. Assume

H(a∗)=0,δγ:=∣H′(a∗)∣>0,H(a_*) = 0, \qquad\delta_\gamma:= |H'(a_*)| > 0,
∥H^−H∥C1(I‾)≤εT,∥H′′∥C0(I‾)≤M2,\|\widehat{H}-H\|_{C^1(\overline{I})} \le\varepsilon_T, \qquad\|H''\|_{C^0(\overline{I})} \le M_2,

and

6M2εT<δγ2,4εT<δγ.6M_2\varepsilon_T < \delta_\gamma^2, \qquad4\varepsilon_T < \delta_\gamma.

Set

r:=3εTδγr := \frac{3\varepsilon_T}{\delta_\gamma}

and assume

Br(a∗)‾⊂I.\overline{B_r(a_*)} \subset I.

Then H^\widehat{H} has a unique zero a^\widehat{a} in Br(a∗)‾\overline{B_r(a_*)}; in fact a^∈Br(a∗)\widehat{a}\in B_r(a_*), and

∣a^−a∗∣≤εTδγ−εT−3M2εT/δγ.(14)|\widehat{a}-a_*| \le\frac{\varepsilon_T}{\delta_\gamma-\varepsilon_T-3M_2\varepsilon_T/\delta_\gamma}. \tag*{(14)}

In particular,

∣a^−a∗∣≤4εTδγ.(15)|\widehat{a}-a_*| \le\frac{4\varepsilon_T}{\delta_\gamma}. \tag*{(15)}

If εT/δγ≪1\varepsilon_T/\delta_\gamma\ll1 and M2εT/δγ2≪1M_2\varepsilon_T/\delta_\gamma^2 \ll1, then only the upper asymptotic estimate

∣a^−a∗∣≤εTδγ(1+O(εTδγ+M2εTδγ2))|\widehat{a}-a_*| \le\frac{\varepsilon_T}{\delta_\gamma}\left(1+O\left(\frac{\varepsilon_T}{\delta_\gamma}+\frac{M_2\varepsilon_T}{\delta_\gamma^2}\right)\right)

is asserted. No global uniqueness in II follows from these hypotheses alone.

Proof. Set r=3εT/δγr=3\varepsilon_T/\delta_\gamma. For a∈Br(a∗)‾a\in\overline{B_r(a_*)},

∣H′(a)∣≥δγ−M2r>δγ2,|H'(a)| \ge\delta_\gamma-M_2r > \frac{\delta_\gamma}{2},

so H′H' has constant sign on the closed interval Br(a∗)‾\overline{B_r(a_*)}. Replacing HH and H^\widehat{H} by their negatives if necessary, suppose H′>0H'>0. Moreover,

H(a∗+r)=∫0rH′(a∗+s) ds≥δγr−12M2r2>εT,H(a_*+r)=\int_0^r H'(a_*+s)\,\mathrm{d}s \ge\delta_\gamma r-\frac{1}{2}M_2r^2>\varepsilon_T,

and similarly H(a∗−r)<−εTH(a_*-r)<-\varepsilon_T. Hence [−εT,εT]⊂H(Br(a∗)‾)[-\varepsilon_T,\varepsilon_T]\subset H(\overline{B_r(a_*)}), and the inverse branch H−1H^{-1} is defined there with

∣(H−1)′(e)∣≤2δγ.|(H^{-1})'(e)|\le\frac{2}{\delta_\gamma}.

Define on the complete metric space Br(a∗)‾\overline{B_r(a_*)}

Φ(a):=H−1(−(H^−H)(a)).\Phi(a) := H^{-1}(-(\widehat{H}-H)(a)).

Then

∣Φ(a)−a∗∣≤2εTδγ<r,|\Phi(a)-a_*| \le\frac{2\varepsilon_T}{\delta_\gamma}<r,

so Φ(Br(a∗)‾)⊂Br(a∗)\Phi(\overline{B_r(a_*)})\subset B_r(a_*). Also

∣Φ′(a)∣≤2εTδγ<12.|\Phi'(a)|\le\frac{2\varepsilon_T}{\delta_\gamma}<\frac{1}{2}.

Banach’s fixed-point theorem therefore gives a unique fixed point a^∈Br(a∗)‾\hat{a}\in\overline{B_r(a_*)}, necessarily in the open ball. The identity Φ(a^)=a^\Phi(\hat{a})=\hat{a} is equivalent to H^(a^)=0\widehat{H}(\hat{a})=0, and every zero of H^\widehat{H} in the closed ball is a fixed point of Φ\Phi, proving uniqueness.

Finally,

0=H^(a^)=H^(a∗)+∫a∗a^H^′(s) ds.0=\widehat{H}(\hat{a})=\widehat{H}(a_*)+\int_{a_*}^{\hat{a}}\widehat{H}'(s)\,ds.

On the segment joining a∗a_* to a^\hat{a},

∣H^′(s)∣≥δγ−M2r−εT,|\widehat{H}'(s)|\ge\delta_\gamma-M_2r-\varepsilon_T,

while ∣H^(a∗)∣≤εT|\widehat{H}(a_*)|\le\varepsilon_T. This yields (14); the remaining estimates follow immediately.

Observable first-order coefficient

The exact quantity δγ\delta_\gamma depends on the unknown anchor. Define instead

δ^γ:=∣T^γ′(a^)−1∣=∣H^′(a^)∣.\widehat{\delta}_\gamma:=|\widehat{T}'_\gamma(\hat{a})-1|=|\widehat{H}'(\hat{a})|.

Lemma 3.2 (Proximity of the observed multiplier). Under the hypotheses of Theorem 3.1,

∣δ^γ−δγ∣≤εT+M2∣a^−a∗∣≤εT(1+3M2δγ).|\widehat{\delta}_\gamma-\delta_\gamma|\le\varepsilon_T+M_2|\hat{a}-a_*|\le\varepsilon_T\left(1+\frac{3M_2}{\delta_\gamma}\right).

Proof. Use the reverse triangle inequality and split

H^′(a^)−H′(a∗)=[H^′(a^)−H′(a^)]+[H′(a^)−H′(a∗)].\widehat{H}'(\hat{a})-H'(a_*)=[\widehat{H}'(\hat{a})-H'(\hat{a})]+[H'(\hat{a})-H'(a_*)].

The first term is bounded by εT\varepsilon_T, the second by M2∣a^−a∗∣M_2|\hat{a}-a_*|, and ∣a^−a∗∣<r|\hat{a}-a_*|<r.

Corollary 3.3 (Observable non-neutral bound). Assume, in addition,

εT(1+3M2δγ)≤δγ2.\varepsilon_T\left(1+\frac{3M_2}{\delta_\gamma}\right)\le\frac{\delta_\gamma}{2}.

Then

δγ2≤δ^γ≤3δγ2,\frac{\delta_\gamma}{2}\le\widehat{\delta}_\gamma\le\frac{3\delta_\gamma}{2},

and the anchor error admits the observable-form estimate

∣a^−a∗∣≤9εT2δ^γ.(16)|\hat{a}-a_*|\le\frac{9\varepsilon_T}{2\widehat{\delta}_\gamma}. \tag*{(16)}

Proof. The interval for δ^γ\widehat{\delta}_\gamma follows from Lemma 3.2. Since δ^γ≤3δγ/2\widehat{\delta}_\gamma\le3\delta_\gamma/2, one has δγ≥2δ^γ/3\delta_\gamma\ge2\widehat{\delta}_\gamma/3. Combining this with ∣a^−a∗∣<3εT/δγ|\hat{a}-a_*|<3\varepsilon_T/\delta_\gamma gives (16).

From return-map perturbations to holonomy perturbations

Lemma 3.4 (Continuity of the transport map). Let γ:[0,1]→M\gamma: [0,1] \to M be a C1C^1 curve, and set

εF:=∥u^F−uF∥C0(γ)+∥q^F−qF∥C0(γ).\varepsilon_F := \|\widehat{u}_F-u_F\|_{C^0(\gamma)}+\|\widehat{q}_F-q_F\|_{C^0(\gamma)}.

Fix ε0>0\varepsilon_0>0 and assume 0≤εF≤ε00\leq\varepsilon_F\leq\varepsilon_0. Assume that the exact transport issued from the compact interval I0I_0 is uniformly separated with margin η>0\eta>0 in the sense of Definition 2.1. Shrinking ε0\varepsilon_0 if necessary, Lemma 2.2 gives perturbed separation with margin η/2\eta/2 and a common bounded fibre region. Then there exist constants C0,C1,L0,L1C_0,C_1,L_0,L_1, depending only on that common fibre bound, η\eta, ε0\varepsilon_0, the geometry of MM, and uniform bounds for the exact fields, such that along γ\gamma

∣∂yG(t,y)∣+∣∂yG^(t,y)∣≤L0∣γ˙(t)∣,|\partial_yG(t,y)|+|\partial_y\widehat{G}(t,y)|\leq L_0|\dot{\gamma}(t)|,
∣∂y2G^(t,y)∣≤L1∣γ˙(t)∣,|\partial_y^2\widehat{G}(t,y)|\leq L_1|\dot{\gamma}(t)|,
∣G^(t,y)−G(t,y)∣≤C0∣γ˙(t)∣εF,|\widehat{G}(t,y)-G(t,y)|\leq C_0|\dot{\gamma}(t)|\varepsilon_F,

and

∣∂yG^(t,y)−∂yG(t,y)∣≤C1∣γ˙(t)∣εF.|\partial_y\widehat{G}(t,y)-\partial_yG(t,y)|\leq C_1|\dot{\gamma}(t)|\varepsilon_F.

Consequently, with ℓ(γ)=∫01∣γ˙∣ dt\ell(\gamma)=\int_0^1|\dot{\gamma}|\,dt,

∥T^γ−Tγ∥C0(I0)≤C0ℓ(γ)eL0ℓ(γ)εF,(17)\|\widehat{T}_\gamma-T_\gamma\|_{C^0(I_0)}\leq C_0\ell(\gamma)e^{L_0\ell(\gamma)}\varepsilon_F, \tag*{(17)}
∥T^γ′−Tγ′∥C0(I0)≤Cγ(1)εF,(18)\|\widehat{T}'_\gamma-T'_\gamma\|_{C^0(I_0)}\leq C_\gamma^{(1)}\varepsilon_F, \tag*{(18)}

where Cγ(1)C_\gamma^{(1)} is explicit in C0,C1,L0,L1C_0,C_1,L_0,L_1 and ℓ(γ)\ell(\gamma). In particular, a bound εT≤CγεF\varepsilon_T\leq C_\gamma\varepsilon_F is available on every fixed admissible loop.

Proof. Writing G=(A−yB)/(y−q)G=(A-yB)/(y-q), the identities

∂yG=−A−qB(y−q)2,∂y2G=2A−qB(y−q)3\partial_yG=-\frac{A-qB}{(y-q)^2},\qquad\partial_y^2G=2\frac{A-qB}{(y-q)^3}

give the first two estimates. Let η∗:=η/2\eta_*:=\eta/2 denote the common lower separation margin after applying Lemma 2.2. For the exact–perturbed pair one may take the common Lipschitz constant

L0=1η∗2max⁡{sup⁡M∣uF∣(1+∣qF∣∥S∥op),sup⁡M∣u^F∣(1+∣q^F∣∥S∥op)}.(19)L_0=\frac{1}{\eta_*^2}\max\left\{\sup_M |u_F|(1+|q_F|\|S\|_{\mathrm{op}}),\sup_M|\widehat{u}_F|(1+|\widehat{q}_F|\|S\|_{\mathrm{op}})\right\}. \tag*{(19)}

Because εF≤ε0\varepsilon_F\leq\varepsilon_0, this quantity is bounded uniformly in terms of the exact field bounds, ε0\varepsilon_0, η\eta, and ∥S∥op\|S\|_{\mathrm{op}}. The perturbation estimates follow by differentiating the rational function (A−yB)/(y−q)(A-yB)/(y-q) with respect to A,B,qA,B,q on the common separation region.

If e=y^−ye=\widehat{y}-y denotes the difference of lifts with the same initial value, Gronwall’s inequality gives

∣e(1)∣≤C0ℓ(γ)eL0ℓ(γ)εF.|e(1)|\leq C_0\ell(\gamma)e^{L_0\ell(\gamma)}\varepsilon_F.

For Z=∂ayZ=\partial_a y and Z^=∂ay^\widehat{Z}=\partial_a\widehat{y},

Z′=Gy(t,y)Z,Z^′=G^y(t,y^)Z^.Z'=G_y(t,y)Z,\qquad\widehat{Z}'=\widehat{G}_y(t,\widehat{y})\widehat{Z}.

The exact variation satisfies ∣Z(t)∣≤eL0ℓ(γ)|Z(t)|\leq e^{L_0\ell(\gamma)}. Subtracting the two equations and using

∣G^y(t,y^)−Gy(t,y)∣≤L1∣γ˙∣∣e∣+C1∣γ˙∣εF|\widehat{G}_y(t,\widehat{y})-G_y(t,y)|\leq L_1|\dot{\gamma}||e|+C_1|\dot{\gamma}|\varepsilon_F

followed by a second Gronwall estimate proves (18). ▫

Global propagation in the non-neutral regime

Corollary 3.5 (Global non-neutral reconstruction stability). Assume the hypotheses of Theorem 3.1 and Lemma 3.4, uniformly along paths from c0c_0 to cc. Let d^\widehat{d} be reconstructed by transport with d^(c0)=a^\widehat{d}(c_0)=\widehat{a}. Then

∣d^(c)−d(c)∣≤eL0ℓ(γc)(∣a^−a∗∣+C0ℓ(γc)εF).\lvert\widehat{d}(c)-d(c)\rvert\le e^{L_0\ell(\gamma_c)}\left(\lvert\widehat{a}-a_\ast\rvert+C_0\ell(\gamma_c)\varepsilon_F\right).

with L0L_0 as in (19). If MM is connected and

DM:=diam⁡dM(M)D_M:=\operatorname{diam}_{d_M}(M)

is its intrinsic Riemannian diameter, then

∥d^−d∥C0(M)≤eL0DM(4εTδγ+C0DMεF).\lVert\widehat{d}-d\rVert_{C^0(M)}\le e^{L_0D_M}\left(\frac{4\varepsilon_T}{\delta_\gamma}+C_0D_M\varepsilon_F\right).

The holonomy variational jet

First variation

Define

Z1(t):=∂aya(t)∣a=a∗.(20)Z_1(t):=\left.\partial_a y_a(t)\right|_{a=a_\ast}. \tag*{(20)}

Then

Z1′(t)=Gy(t,y∗(t))Z1(t),Z1(0)=1,(21)Z_1'(t)=G_y(t,y_\ast(t))Z_1(t),\qquad Z_1(0)=1, \tag*{(21)}

so

Z1(t)=exp⁡(∫0tGy(s,y∗(s)) ds).(22)Z_1(t)=\exp\left(\int_0^t G_y(s,y_\ast(s))\,ds\right). \tag*{(22)}

Hence

Tγ′(a∗)=Z1(1),H′(a∗)=Z1(1)−1.(23)T_\gamma'(a_\ast)=Z_1(1),\qquad H'(a_\ast)=Z_1(1)-1. \tag*{(23)}

The neutral condition is therefore

Z1(1)=1.(24)Z_1(1)=1. \tag*{(24)}

Second variation: exact formula

Define

W(t):=∂a2ya(t)∣a=a∗.(25)W(t):=\left.\partial_a^2y_a(t)\right|_{a=a_\ast}. \tag*{(25)}

Differentiating the variational equation gives

W′=Gy(t,y∗)W+Gyy(t,y∗)Z12,W(0)=0.(26)W'=G_y(t,y_\ast)W+G_{yy}(t,y_\ast)Z_1^2,\qquad W(0)=0. \tag*{(26)}

Since Z1′=GyZ1Z_1'=G_yZ_1,

(WZ1)′=Gyy(t,y∗)Z1.(27)\left(\frac{W}{Z_1}\right)'=G_{yy}(t,y_\ast)Z_1. \tag*{(27)}

Therefore

W(1)=Z1(1)∫01Gyy(t,y∗(t))Z1(t) dt.(28)W(1)=Z_1(1)\int_0^1G_{yy}(t,y_\ast(t))Z_1(t)\,dt. \tag*{(28)}

Theorem 4.1 (Exact second holonomy variation). Assume that the horizontal transport is defined on a common fibre interval around a∗a_\ast and is twice differentiable with respect to the initial anchor. Then

H′′(a∗)=Tγ′′(a∗)=Tγ′(a∗)∫01Gyy(t,d(γ(t)))Z1(t) dt.(29)H^{\prime\prime}(a_\ast)=T^{\prime\prime}_{\gamma}(a_\ast)=T^{\prime}_{\gamma}(a_\ast)\int_0^1 G_{yy}(t,d(\gamma(t)))Z_1(t)\,\mathrm{d}t. \tag*{(29)}

Equivalently,

H′′(a∗)=2Tγ′(a∗)∫01P(t)(d(γ(t))−q(t))3Z1(t) dt.(30)H^{\prime\prime}(a_\ast)=2T^{\prime}_{\gamma}(a_\ast)\int_0^1\frac{P(t)}{(d(\gamma(t))-q(t))^3}Z_1(t)\,\mathrm{d}t. \tag*{(30)}

In the neutral regime Tγ′(a∗)=1T^{\prime}_{\gamma}(a_\ast)=1, one has

H′′(a∗)=2∫01P(t)(d(γ(t))−q(t))3Z1(t) dt.(31)H^{\prime\prime}(a_\ast)=2\int_0^1\frac{P(t)}{(d(\gamma(t))-q(t))^3}Z_1(t)\,\mathrm{d}t. \tag*{(31)}

Proof. The formula follows from the variational equation for WW and the identity

(W/Z1)′=GyyZ1.(W/Z_1)^\prime=G_{yy}Z_1.

The explicit expression uses

Gyy(t,y)=2P(t)(y−q(t))3.G_{yy}(t,y)=2\frac{P(t)}{(y-q(t))^3}.

Remark 4.2. Theorem 4.1 replaces the earlier assertion that second-order holonomy is “controlled by dηdd\eta_d”. The correct object is the vertical second variation of the nonlinear transport along the true horizontal section.

Higher-order holonomy jet

For n≥1n\ge1, define

Zn(t):=∂anya(t)∣a=a∗.Z_n(t):=\left.\partial_a^n y_a(t)\right|_{a=a_\ast}.

The first two equations are

Z1′=GyZ1,Z2′=GyZ2+GyyZ12.Z_1^\prime=G_yZ_1,\qquad Z_2^\prime=G_yZ_2+G_{yy}Z_1^2.

The general equation is most compactly written with the partial Bell polynomials Bn,rB_{n,r}. We use the convention

Bn,r(x1,…,xn−r+1)=∑n!j1!⋯jn−r+1!∏m=1n−r+1(xmm!)jm,B_{n,r}(x_1,\ldots,x_{n-r+1})=\sum\frac{n!}{j_1!\cdots j_{n-r+1}!}\prod_{m=1}^{n-r+1}\left(\frac{x_m}{m!}\right)^{j_m},

where the sum is over jm≥0j_m\ge0 satisfying ∑mjm=r\sum_m j_m=r and ∑mmjm=n\sum_m mj_m=n.

Theorem 4.3 (Finite holonomy variational hierarchy). Assume the horizontal transport is defined on a common fibre interval around a∗a_\ast and that Gγ(t,⋅)G_\gamma(t,\cdot) is CpC^p there. Then, for 1≤n≤p1\le n\le p,

Zn′=∑r=1nGy(r)(t,y∗(t))Bn,r(Z1,…,Zn−r+1),(32)Z_n^\prime=\sum_{r=1}^{n}G_y^{(r)}(t,y_\ast(t))B_{n,r}(Z_1,\ldots,Z_{n-r+1}), \tag*{(32)}

with

Z1(0)=1,Zn(0)=0(n≥2).Z_1(0)=1,\qquad Z_n(0)=0\quad(n\ge2).

At the endpoint,

H′(a∗)=Z1(1)−1,H(n)(a∗)=Zn(1)(n≥2).H^\prime(a_\ast)=Z_1(1)-1,\qquad H^{(n)}(a_\ast)=Z_n(1)\quad(n\ge2).

Proof. Differentiate the identity ya′=Gγ(t,ya)y_a' = G_\gamma(t,y_a) nn times with respect to aa. Faà di Bruno’s formula gives (32); the initial conditions follow from ya(0)=ay_a(0)=a. Since Tγ(a)=ya(1)T_\gamma(a)=y_a(1) and H=Tγ−Id⁡H=T_\gamma-\operatorname{Id}, the endpoint identities follow.

For the third derivative the recursion is explicit:

Z3′=GyZ3+3GyyZ1Z2+GyyyZ13,Z3(0)=0.Z_3' = G_y Z_3 + 3G_{yy}Z_1Z_2 + G_{yyy}Z_1^3,\qquad Z_3(0)=0.

Dividing by Z1Z_1 and using Z1′=GyZ1Z_1'=G_yZ_1 gives the following useful formula.

Corollary 4.4 (Exact third holonomy variation). Under the hypotheses of Theorem 4.3,

H′′′(a∗)=Tγ′(a∗)∫01(3Gyy(t,y∗)Z2(t)+Gyyy(t,y∗)Z1(t)2) dt.(33)H'''(a_*)=T_\gamma'(a_*)\int_0^1\left(3G_{yy}(t,y_*)Z_2(t)+G_{yyy}(t,y_*)Z_1(t)^2\right)\,\mathrm{d}t. \tag*{(33)}

where

Gyyy(t,y)=−6P(t)(y−q(t))4.G_{yyy}(t,y)=-6\frac{P(t)}{(y-q(t))^4}.

The holonomy jet is therefore the endpoint jet of the vertical variational hierarchy:

Jγ(d;a∗)=(H′(a∗),H′′(a∗),H′′′(a∗),…).\mathcal{J}_\gamma(d;a_*)=\left(H'(a_*),H''(a_*),H'''(a_*),\ldots\right).

The isolation order is the first nonzero component of this jet.

Finite-dimensional holonomy unfoldings and Malgrange preparation

The finite-order condition is not only a Taylor-order statement. For smooth finite-dimensional families of admissible return data it has an exact local preparation in the sense of Malgrange. We use the standard smooth preparation theorem; see Golubitsky–Guillemin [8 Chap. IV]. The restriction to finite-dimensional parameter families is essential for this direct application: the full space of return-map perturbations is infinite-dimensional and is not covered by the classical theorem without additional functional-analytic structure.

Let Λ⊂Rm\Lambda\subset\mathbb{R}^m be a neighbourhood of 00. Consider a C∞C^\infty family of return components

λ⟼(uλ,qλ),λ∈Λ,\lambda\longmapsto(u_\lambda,q_\lambda),\qquad\lambda\in\Lambda,

along a fixed closed loop γ\gamma, and let Tγ,λT_{\gamma,\lambda} denote the corresponding loop transport. Assume that the horizontal lifts are defined on a common fibre neighbourhood of a∗a_* and remain uniformly separated from the singular set. Standard smooth dependence of ODE solutions on parameters and initial data then gives a jointly smooth displacement

H(λ,a):=Tγ,λ(a)−a.H(\lambda,a):=T_{\gamma,\lambda}(a)-a.

Theorem 4.5 (Malgrange preparation of a finite-order holonomy singularity). Assume that, at λ=0\lambda=0,

H(0,a∗)=0,∂ajH(0,a∗)=0(1≤j<k),∂akH(0,a∗)≠0.H(0,a_*)=0,\qquad\partial_a^jH(0,a_*)=0\quad(1\le j<k),\qquad\partial_a^kH(0,a_*)\ne0.

Set ξ=a−a∗\xi=a-a_*. Then there exist neighbourhoods Λ0∋0\Lambda_0\ni0 and J0∋0J_0\ni0, a smooth function

U∈C∞(Λ0×J0),U(0,0)≠0,U\in C^\infty(\Lambda_0\times J_0),\qquad U(0,0)\ne0,

and smooth coefficient functions c0,…,ck−1∈C∞(Λ0)c_0,\ldots,c_{k-1}\in C^\infty(\Lambda_0) with

cj(0)=0(0≤j<k),c_j(0)=0\qquad(0\le j<k),

such that, after shrinking the neighbourhoods if necessary,

H(λ,a∗+ξ)=U(λ,ξ)(ξk+∑j=0k−1cj(λ)ξj).(34)H(\lambda, a_*+\xi)=U(\lambda,\xi)\left(\xi^k+\sum_{j=0}^{k-1}c_j(\lambda)\xi^j\right). \tag*{(34)}

In particular, UU is nonvanishing on Λ0×J0\Lambda_0 \times J_0, and the local real anchors of the perturbed holonomy equation H(λ,a)=0H(\lambda,a)=0 are exactly the real roots of the prepared polynomial

Pλ(ξ):=ξk+∑j=0k−1cj(λ)ξj.P_\lambda(\xi):=\xi^k+\sum_{j=0}^{k-1}c_j(\lambda)\xi^j.

Proof. The uniform separation assumption removes the pole of the rational transport field from a common neighbourhood of the relevant solution graphs. Since the family (uλ,qλ)(u_\lambda,q_\lambda) is C∞C^\infty, standard parameter dependence for ODEs implies that (λ,a)↦Tγ,λ(a)(\lambda,a)\mapsto T_{\gamma,\lambda}(a), and hence HH, is C∞C^\infty jointly near (0,a∗)(0,a_*). In the coordinate ξ=a−a∗\xi=a-a_*, the displayed derivative conditions say precisely that the smooth germ (λ,ξ)↦H(λ,a∗+ξ)(\lambda,\xi)\mapsto H(\lambda,a_*+\xi) is regular of order kk in the distinguished variable ξ\xi. The Malgrange preparation theorem therefore yields a smooth unit times a monic polynomial of degree kk in ξ\xi, with smooth parameter-dependent coefficients. Because the unperturbed germ has multiplicity exactly kk, the prepared coefficients satisfy cj(0)=0c_j(0)=0. Shrinking the neighbourhoods makes the unit nonvanishing throughout. The last assertion follows immediately from (34).

Corollary 4.6 (Quadratic neutral unfolding and discriminant). If k=2k=2, then locally

H(λ,a∗+ξ)=U(λ,ξ)(ξ2+c1(λ)ξ+c0(λ)).H(\lambda,a_*+\xi)=U(\lambda,\xi)\left(\xi^2+c_1(\lambda)\xi+c_0(\lambda)\right).

Hence the local real-anchor structure is governed by the discriminant

Δ(λ)=c1(λ)2−4c0(λ).\Delta(\lambda)=c_1(\lambda)^2-4c_0(\lambda).

For sufficiently small λ\lambda, Δ>0\Delta>0 gives two nearby simple real anchors, Δ=0\Delta=0 a multiple real anchor, and Δ<0\Delta<0 no nearby real anchor. This is the prepared-polynomial counterpart of the even-order root-splitting alternatives in Proposition 5.2.

Proposition 4.7 (First-order coefficients of a quadratic prepared unfolding). Assume the hypotheses of Corollary 4.6. Write all derivatives below at (λ,a)=(0,a∗)(\lambda,a)=(0,a_*), and set

A:=Haa(0,a∗)≠0,B:=Haaa(0,a∗).A:=H_{aa}(0,a_*)\ne0,\qquad B:=H_{aaa}(0,a_*).

Then, as λ→0\lambda\to0,

c0(λ)=2Aλ⋅∇λH+O(∣λ∣2),(35)c_0(\lambda)=\frac{2}{A}\lambda\cdot\nabla_\lambda H+O(|\lambda|^2), \tag*{(35)}
c1(λ)=2Aλ⋅∇λHa−2B3A2λ⋅∇λH+O(∣λ∣2),(36)c_1(\lambda)=\frac{2}{A}\lambda\cdot\nabla_\lambda H_a-\frac{2B}{3A^2}\lambda\cdot\nabla_\lambda H+O(|\lambda|^2), \tag*{(36)}
Δ(λ)=−8Aλ⋅∇λH+O(∣λ∣2).(37)\Delta(\lambda)=-\frac{8}{A}\lambda\cdot\nabla_\lambda H+O(|\lambda|^2). \tag*{(37)}

In particular, to first order the discriminant is controlled by the parameter derivative of the holonomy residual at the singular anchor, whereas the linear coefficient c1c_1 also feels the cubic anchor jet through BB unless ∇λH(0,a∗)=0\nabla_\lambda H(0,a_*)=0.

Proof. Write ξ=a−a∗\xi= a-a_\ast and

H(λ,a∗+ξ)=U(λ,ξ)(ξ2+c1(λ)ξ+c0(λ)).H(\lambda,a_\ast+\xi)=U(\lambda,\xi)\left(\xi^2+c_1(\lambda)\xi+c_0(\lambda)\right).

At λ=0\lambda=0, comparison with the Taylor expansion of H(0,a∗+ξ)H(0,a_\ast+\xi) gives

U(0,0)=A2,Uξ(0,0)=B6.U(0,0)=\frac{A}{2},\qquad U_\xi(0,0)=\frac{B}{6}.

Setting ξ=0\xi=0 and differentiating with respect to λ\lambda yields

∇λc0(0)=2A∇λH(0,a∗).\nabla_\lambda c_0(0)=\frac{2}{A}\nabla_\lambda H(0,a_\ast).

Differentiating first in ξ\xi and then in λ\lambda, and using c0(0)=c1(0)=0c_0(0)=c_1(0)=0, gives

∇λHa=Uξ(0,0)∇λc0(0)+U(0,0)∇λc1(0),\nabla_\lambda H_a=U_\xi(0,0)\nabla_\lambda c_0(0)+U(0,0)\nabla_\lambda c_1(0),

which is equivalent to (36). Finally c1(λ)2=O(∣λ∣2)c_1(\lambda)^2=O(|\lambda|^2), so

Δ(λ)=c1(λ)2−4c0(λ)=−8Aλ⋅∇λH+O(∣λ∣2).\Delta(\lambda)=c_1(\lambda)^2-4c_0(\lambda)=-\frac{8}{A}\lambda\cdot\nabla_\lambda H+O(|\lambda|^2).

Remark 4.8 (Structural versus quantitative information). The preparation theorem does not replace the explicit Hölder estimates of § 5. It provides an exact local structural reduction for smooth finite-dimensional unfoldings, while Theorem 5.1 quantifies how far a perturbed root can move under norm-bounded perturbations and Theorem 6.1 connects those perturbations to noisy return data. Thus the two viewpoints are complementary:

holonomy jet⟹prepared polynomial⟹discriminant/root geometry,\text{holonomy jet}\Longrightarrow\text{prepared polynomial}\Longrightarrow\text{discriminant/root geometry},

with the stability theory supplying the corresponding conditioning estimates.

Analytic dependence and the flat regime

The rational fibre dependence of GγG_\gamma gives more structure than C∞C^\infty regularity. Analytic dependence of flows on initial conditions for time-varying analytic vector fields is standard; see, for example, [10]. We include the short argument in the present scalar setting.

Theorem 4.9 (Analyticity of the holonomy map). Let J⋐RJ\Subset\mathbb{R} be an interval of initial anchors such that, for every a∈Ja\in J, the horizontal lift ya(t)y_a(t) exists for 0≤t≤10\le t\le1 and

ya(t)−q(t)≥η>0y_a(t)-q(t)\ge\eta>0

uniformly in (a,t)∈J×[0,1](a,t)\in J\times[0,1]. Then a↦Tγ(a)a\mapsto T_\gamma(a) is real analytic on JJ. Consequently, if a∗∈Ja_\ast\in J, H(a∗)=0H(a_\ast)=0, and

H(n)(a∗)=0(n≥1),H^{(n)}(a_\ast)=0\qquad(n\ge1),

then there exists a neighbourhood U⊂JU\subset J of a∗a_\ast on which

Tγ(a)=a.T_\gamma(a)=a.

Thus infinite isolation order for one loop is equivalent locally to identity holonomy for that loop.

Proof. Fix a0∈Ja_0 \in J. By the uniform separation and continuous dependence of solutions, the exact solution graph through a0a_0 is contained in a compact real strip that has positive distance from the pole y=q(t)y=q(t). The rational function Gγ(t,z)G_\gamma(t,z) therefore extends holomorphically in zz to a complex neighbourhood of that strip, uniformly for t∈[0,1]t \in[0,1]. On a sufficiently short time interval, Picard iteration for the complexified ODE maps a common complex neighbourhood of a0a_0 into itself; every iterate is holomorphic in the initial value and the iterates converge uniformly. Hence the local flow is holomorphic in the initial value. A finite composition of such short-time flow maps covers [0,1][0,1], so TγT_\gamma is holomorphic near a0a_0 and therefore real analytic on JJ.

Since H=Tγ−Id⁡H=T_\gamma-\operatorname{Id} is real analytic, vanishing of H(a∗)H(a_*) and all derivatives at a∗a_* forces its Taylor series to vanish identically in a neighbourhood of a∗a_*. The identity theorem gives H≡0H\equiv0 there.

Finite-order anchor isolation and parity-sensitive existence

The scalar exponent 1/k1/k follows from the Taylor structure of a zero of multiplicity kk. In the present inverse problem, the substantive point is that the multiplicity is the first nonzero component of the holonomy jet derived in § 4.

Theorem 5.1 (Order-kk isolation). Let k≥2k \ge2, let H∈Ck+1(Iˉ)H \in C^{k+1}(\bar{I}), and assume

H(a∗)=0,H(j)(a∗)=0(1≤j<k),mk:=∣H(k)(a∗)∣>0.H(a_*)=0,\qquad H^{(j)}(a_*)=0\quad(1\le j<k),\qquad m_k:=\lvert H^{(k)}(a_*)\rvert>0.

Set

Mk+1:=∥H(k+1)∥C0(Iˉ).M_{k+1}:=\lVert H^{(k+1)}\rVert_{C^0(\bar{I})}.

If Mk+1>0M_{k+1}>0, set

rk:=mk4Mk+1r_k:=\frac{m_k}{4M_{k+1}}

and assume Bˉrk(a∗)⊂I\bar{B}_{r_k}(a_*)\subset I. Let H^=H+E\widehat{H}=H+E, with ∥E∥C0(Iˉ)≤ε\lVert E\rVert_{C^0(\bar{I})}\le\varepsilon, and suppose H^(a^)=0\widehat{H}(\widehat{a})=0. Assume either

(a) Mk+1>0M_{k+1}>0 and ∣a^−a∗∣≤rk\lvert\widehat{a}-a_*\rvert\le r_k, or

(b) Mk+1=0M_{k+1}=0 and a^∈I\widehat{a}\in I.

Then

∣a^−a∗∣≤rk1/k(εmk)1/k,κk:=4(k+1)!4k+3.(38)\lvert\widehat{a}-a_*\rvert\le r_k^{1/k}\left(\frac{\varepsilon}{m_k}\right)^{1/k},\qquad\kappa_k:=\frac{4(k+1)!}{4k+3}. \tag*{(38)}

Proof. Taylor’s formula gives

H(a^)=H(k)(a∗)k!(a^−a∗)k+Rk+1.H(\widehat{a})=\frac{H^{(k)}(a_*)}{k!}(\widehat{a}-a_*)^k+R_{k+1}.

If Mk+1>0M_{k+1}>0, the localization hypothesis implies

∣Rk+1∣≤Mk+1(k+1)!∣a^−a∗∣k+1≤mk4(k+1)!∣a^−a∗∣k.\lvert R_{k+1}\rvert\le\frac{M_{k+1}}{(k+1)!}\lvert\widehat{a}-a_*\rvert^{k+1}\le\frac{m_k}{4(k+1)!}\lvert\widehat{a}-a_*\rvert^k.

Since H^(a^)=0\widehat{H}(\widehat{a})=0, one has ∣H(a^)∣≤ε\lvert H(\widehat{a})\rvert\le\varepsilon, and therefore

mk(4k+3)4(k+1)!∣a^−a∗∣k≤ε.\frac{m_k(4k+3)}{4(k+1)!}\lvert\widehat{a}-a_*\rvert^k\le\varepsilon.

This is (38). If Mk+1=0M_{k+1}=0, the Taylor remainder vanishes identically on II; the same bound remains valid (and is slightly weaker than the exact remainder-free estimate).

:::

Proposition 5.2 (Parity-sensitive local existence). Let k≥2k \ge2, let H∈Ck+1(Iˉ)H \in C^{k+1}(\bar{I}), and assume

H(a∗)=0,H(j)(a∗)=0(1≤j<k),mk:=∣H(k)(a∗)∣>0.H(a_*) = 0, \qquad H^{(j)}(a_*) = 0 \quad(1 \le j < k), \qquad m_k := \lvert H^{(k)}(a_*)\rvert> 0.

Set Mk+1:=∥H(k+1)∥C0(Iˉ)M_{k+1} := \lVert H^{(k+1)}\rVert_{C^0(\bar{I})} and let H^\hat{H} satisfy

∥H^−H∥C0(I)≤ε.\lVert\hat{H} - H\rVert_{C^0(I)} \le\varepsilon.

Choose ρ>0\rho> 0 such that

[a∗−ρ,a∗+ρ]⊂I[a_*-\rho,a_*+\rho] \subset I

and, when Mk+1>0M_{k+1} > 0,

ρ≤(k+1)mk2Mk+1.\rho\le\frac{(k+1)m_k}{2M_{k+1}}.

If

ε<mk4k!ρk,\varepsilon< \frac{m_k}{4k!}\rho^k,

then:

  • (i) if kk is odd, H^\hat{H} has at least one zero in (a∗−ρ,a∗+ρ)(a_*-\rho,a_*+\rho);

  • (ii) if kk is even and H^(a∗)H(k)(a∗)<0\hat{H}(a_*)H^{(k)}(a_*) < 0, then H^\hat{H} has at least one zero in each of (a∗−ρ,a∗)(a_*-\rho,a_*) and (a∗,a∗+ρ)(a_*,a_*+\rho);

  • (iii) if kk is even and H^(a∗)=0\hat{H}(a_*) = 0, then a∗a_* itself remains a zero.

Proof. Taylor’s formula and the choice of ρ\rho give

∣H(a∗±ρ)∣≥mk2k!ρk.\lvert H(a_* \pm\rho)\rvert\ge\frac{m_k}{2k!}\rho^k.

The perturbation bound preserves the endpoint signs. For odd kk, these signs are opposite, so the intermediate value theorem gives a zero. For even kk, the endpoint signs agree with H(k)(a∗)H^{(k)}(a_*); the stated sign condition at a∗a_* gives one zero on each side. The last case is immediate. □\square

Corollary 5.3 (Finite-order stability hierarchy). Let kγ(a∗)<∞k_\gamma(a_*) < \infty.

  • (i) If kγ(a∗)=1k_\gamma(a_*) = 1, Theorem 3.1 gives first-order stability ∣a^−a∗∣=O(εT)\lvert\hat{a}-a_*\rvert= O(\varepsilon_T).

  • (ii) If kγ(a∗)=k≥2k_\gamma(a_*) = k \ge2, every perturbed zero satisfying the localization hypothesis of Theorem 5.1 obeys

∣a^−a∗∣≤κk1/k(εmk)1/k.\lvert\hat{a}-a_*\rvert\le\kappa_k^{1/k}\left(\frac{\varepsilon}{m_k}\right)^{1/k}.

Existence is automatic for sufficiently small perturbations when kk is odd, and for even kk follows under the sign alternatives of Proposition 5.2.

If kγ(a∗)=+∞k_\gamma(a_*) = +\infty, no finite-order location estimate follows from that loop alone.

Remark 5.4 (Size of the quantitative constants). For the pure local model

H(a)=mkk!(a−a∗)k,H^=H−ε,H(a) = \frac{m_k}{k!}(a-a_*)^k, \qquad\hat{H} = H-\varepsilon,

the exact root displacement is

∣a^−a∗∣=(k!)1/k(εmk)1/k.|\hat{a}-a_*|=(k!)^{1/k}\left(\frac{\varepsilon}{m_k}\right)^{1/k}.

Theorem 5.1 gives instead the coefficient κk1/k\kappa_k^{1/k}. For k=2,3,4k=2,3,4 the exact pure-root coefficients are approximately 1.4142, 1.8171, 2.2134, while the theorem gives 1.4771, 1.8566, 2.2419. Thus the explicit Taylor-remainder constants are quantitative rather than optimal, but are already close to the pure-root asymptotic constants at low order.

Stability and observable finite-order coefficients

The previous section assumes control of the holonomy displacement. We now close the inverse-problem chain by showing that perturbations of the observable return components control every fixed finite vertical holonomy jet.

Theorem 6.1 (Finite-jet stability under return-data perturbations). Fix an integer p≥0p \ge0, a closed C1C^1 loop γ\gamma, a compact anchor interval I0I_0, and a perturbation ceiling ε0>0\varepsilon_0 > 0. Let (uF,qF)(u_F,q_F) and (u^F,q^F)(\hat{u}_F,\hat{q}_F) be exact and perturbed return components satisfying

∥u^F−uF∥C0(γ)+∥q^F−qF∥C0(γ)≤εF≤ε0.\|\hat{u}_F-u_F\|_{C^0(\gamma)}+\|\hat{q}_F-q_F\|_{C^0(\gamma)}\le\varepsilon_F\le\varepsilon_0.

Assume that the exact horizontal lifts are uniformly separated on I0I_0 with margin η>0\eta>0 in the sense of Definition 2.1. Shrinking ε0\varepsilon_0 if necessary, Lemma 2.2 supplies existence, a common bounded fibre region, and perturbed separation with margin η/2\eta/2. Then, for every 0≤j≤p0\le j\le p, there is a constant Cp,γ,jC_{p,\gamma,j} depending only on pp, η\eta, ε0\varepsilon_0, the exact geometric bounds, I0I_0 and γ\gamma such that

∥T^γ(j)−Tγ(j)∥C0(I0)≤Cp,γ,jεF.(39)\|\hat{T}^{(j)}_\gamma-T^{(j)}_\gamma\|_{C^0(I_0)}\le C_{p,\gamma,j}\varepsilon_F. \tag*{(39)}

In particular,

∥H^(j)−H(j)∥C0(I0)≤Cp,γ,jεF(j≥1),\|\hat{H}^{(j)}-H^{(j)}\|_{C^0(I_0)}\le C_{p,\gamma,j}\varepsilon_F \qquad(j\ge1),

while the same conclusion for j=0j=0 follows from H^−H=T^γ−Tγ\hat{H}-H=\hat{T}_\gamma-T_\gamma.

Proof. For r≥1r\ge1, the exact fibre derivatives are

∂yrGγ(t,y)=(−1)rr!P(t)(y−q(t))r+1.\partial_y^rG_\gamma(t,y)=(-1)^r r!\frac{P(t)}{(y-q(t))^{r+1}}.

and the perturbed derivatives have the same form with hatted data. Uniform separation and boundedness imply, for 0≤r≤p+10\le r\le p+1, uniform bounds

∣∂yrGγ∣+∣∂yrG^γ∣≤Cr|\partial_y^rG_\gamma|+|\partial_y^r\hat{G}_\gamma|\le C_r

and, by direct subtraction of the rational expressions,

∣∂yrG^γ−∂yrGγ∣≤Cr′εF.|\partial_y^r\hat{G}_\gamma-\partial_y^rG_\gamma|\le C'_r\varepsilon_F.

The derivative at order p+1p+1 is used when comparing ∂ypG^γ(t,y^)\partial_y^p\hat{G}_\gamma(t,\hat{y}) and ∂ypGγ(t,y)\partial_y^pG_\gamma(t,y) by the mean-value estimate in the fibre variable. The case j=0j=0 is the standard Gronwall comparison of the two transport equations. For j≥1j\ge1, let Zj=∂ajyaZ_j=\partial_a^j y_a and Z^j=∂ajy^a\hat{Z}_j=\partial_a^j\hat{y}_a. Theorem 4.3 gives triangular Bell-polynomial systems for ZjZ_j and Z^j\hat{Z}_j. Uniform bounds for the lower-order variations follow inductively from Gronwall. Subtracting the two equations at order jj yields a linear term in Z^j−Zj\widehat{Z}_j-Z_j plus a forcing bounded by CεFC\varepsilon_F and lower-order differences, already controlled by the induction hypothesis. A further Gronwall estimate gives

∥Z^j−Zj∥C0(I0×[0,1])≤Cp,γ,jεF.\|\widehat{Z}_j-Z_j\|_{C^0(I_0\times[0,1])}\le C_{p,\gamma,j}\varepsilon_F.

Evaluating at t=1t=1 proves (39). □

Remark 6.2. Theorem 6.1 is specific to the vertical anchor jet. It requires only C0C^0 perturbation of the observable fields along the loop because all anchor derivatives are derivatives with respect to the scalar fibre variable; no spatial differentiation of the noisy return data is introduced.

The theoretical coefficient mk=∣H(k)(a∗)∣m_k=|H^{(k)}(a_*)| is evaluated at the unknown true anchor. It is therefore useful to quantify how accurately it can be replaced by a coefficient computed at the reconstructed anchor.

Proposition 6.3 (Observable holonomy coefficient). Let k≥2k\ge2. Assume H∈Ck+1(I‾)H\in C^{k+1}(\overline{I}), H^∈Ck(I‾)\widehat{H}\in C^k(\overline{I}), and

∥H^(k)−H(k)∥C0(I‾)≤εk.\|\widehat{H}^{(k)}-H^{(k)}\|_{C^0(\overline{I})}\le\varepsilon_k.

Set

m^k:=∣H^(k)(a^)∣,mk:=∣H(k)(a∗)∣,Mk+1:=∥H(k+1)∥C0(I‾).\widehat{m}_k:=|\widehat{H}^{(k)}(\widehat{a})|,\qquad m_k:=|H^{(k)}(a_*)|,\qquad M_{k+1}:=\|H^{(k+1)}\|_{C^0(\overline{I})}.

Then

∣m^k−mk∣≤εk+Mk+1∣a^−a∗∣.(40)|\widehat{m}_k-m_k|\le\varepsilon_k+M_{k+1}|\widehat{a}-a_*|. \tag*{(40)}

If the right-hand side is at most θmk\theta m_k for some 0<θ<10<\theta<1, then

11+θm^k≤mk≤11−θm^k.\frac{1}{1+\theta}\widehat{m}_k\le m_k\le\frac{1}{1-\theta}\widehat{m}_k.

Consequently, whenever Theorem 5.1 applies,

∣a^−a∗∣≤((1+θ)κk)1/k(εm^k)1/k.(41)|\widehat{a}-a_*|\le\left((1+\theta)\kappa_k\right)^{1/k}\left(\frac{\varepsilon}{\widehat{m}_k}\right)^{1/k}. \tag*{(41)}

Proof. The reverse triangle inequality gives

∣m^k−mk∣≤∣H^(k)(a^)−H(k)(a^)∣+∣H(k)(a^)−H(k)(a∗)∣,|\widehat{m}_k-m_k|\le|\widehat{H}^{(k)}(\widehat{a})-H^{(k)}(\widehat{a})|+|H^{(k)}(\widehat{a})-H^{(k)}(a_*)|,

which is (40). The two-sided comparison of mkm_k and m^k\widehat{m}_k is immediate, and substituting mk≥m^k/(1+θ)m_k\ge\widehat{m}_k/(1+\theta) into Theorem 5.1 gives (41). □

For k=2k=2 and θ=1/2\theta=1/2, the observable estimate becomes

∣a^−a∗∣≤(36ε11m^2)1/2,(42)|\widehat{a}-a_*|\le\left(\frac{36\varepsilon}{11\widehat{m}_2}\right)^{1/2}, \tag*{(42)}

whereas the theorem constant expressed with the unknown exact coefficient m2m_2 is 24/11\sqrt{24/11}. The leading scalar asymptotic coefficient for a pure quadratic root is 2\sqrt{2}, so the displayed theorem constant is quantitative rather than optimal.

Corollary 6.4 (Observable coefficients from return data). Under the hypotheses of Theorem 6.1, for every fixed k≥2k\ge2,

∣m^k−mk∣≤Ck,γεF+Mk+1∣a^−a∗∣.|\widehat{m}_k-m_k|\le C_{k,\gamma}\varepsilon_F+M_{k+1}|\widehat{a}-a_*|.

Hence whenever the right-hand side is at most θmk\theta m_k with 0<θ<10<\theta<1, the observable estimate (41) applies with a coefficient computable from the perturbed holonomy jet.

Multi-loop finite-order aggregation

A single loop may have a weak or high-order signal. Several loops sharing the same true anchor can be combined deterministically without assuming statistical independence.

Let γ1,…,γr\gamma_1,\ldots,\gamma_r be closed loops based at the same c0c_0, with holonomy displacements Hj=Tγj−Id⁡H_j=T_{\gamma_j}-\operatorname{Id}. Set

k∗:=min⁡jkγj(a∗).k_* := \min_j k_{\gamma_j}(a_*).

Assume k∗<∞k_*<\infty and set

J∗:={j:kγj(a∗)=k∗}.J_* := \{j:k_{\gamma_j}(a_*)=k_*\}.

For j∈J∗j\in J_* define

vj:=Hj(k∗)(a∗),m∗:=(∑j∈J∗∣vj∣2)1/2>0,v_j := H_j^{(k_*)}(a_*),\qquad m_* := \left(\sum_{j\in J_*}|v_j|^2\right)^{1/2}>0,

and

M∗:=(∑j∈J∗∥Hj(k∗+1)∥C0(I‾)2)1/2.M_* := \left(\sum_{j\in J_*}\|H_j^{(k_*+1)}\|_{C_0(\overline{I})}^2\right)^{1/2}.

Theorem 7.1 (Deterministic multi-loop finite-order bound). Assume Hj(a∗)=0H_j(a_*)=0 and Hj(ℓ)(a∗)=0H_j^{(\ell)}(a_*)=0 for 1≤ℓ<k∗1\leq\ell<k_* and every j∈J∗j\in J_*. Let

H=(Hj)j∈J∗,H^=(H^j)j∈J∗,\mathbf{H}=(H_j)_{j\in J_*},\qquad\widehat{\mathbf{H}}=(\widehat{H}_j)_{j\in J_*},

and suppose

sup⁡a∈I∥H^(a)−H(a)∥2≤ε∗.\sup_{a\in I}\|\widehat{\mathbf{H}}(a)-\mathbf{H}(a)\|_2\leq\varepsilon_*.

Let a^∈I\widehat{a}\in I be an approximate common anchor satisfying

∥H^(a^)∥2≤τ.\|\widehat{\mathbf{H}}(\widehat{a})\|_2\leq\tau.

If M∗>0M_*>0, set r∗=m∗/(4M∗)r_* = m_*/(4M_*), assume B‾r∗(a∗)⊂I\overline{B}_{r_*}(a_*)\subset I, and suppose ∣a^−a∗∣≤r∗|\widehat{a}-a_*|\leq r_*. If M∗=0M_*=0, assume only a^∈I\widehat{a}\in I. Then

∣a^−a∗∣≤κk∗(ε∗+τm∗)1/k∗,κk=4(k+1)!4k+3.(43)|\widehat{a}-a_*|\leq\kappa_{k_*}\left(\frac{\varepsilon_*+\tau}{m_*}\right)^{1/k_*},\qquad \kappa_k=\frac{4(k+1)!}{4k+3}. \tag*{(43)}

Proof. Write s=a^−a∗s=\widehat{a}-a_*. Vector Taylor expansion gives

H(a∗+s)=sk∗k∗!v+Rk∗+1(s),v=(vj)j∈J∗,\mathbf{H}(a_*+s)=\frac{s^{k_*}}{k_*!}v+\mathbf{R}_{k_*+1}(s),\qquad v=(v_j)_{j\in J_*},

with

∥Rk∗+1(s)∥2≤M∗(k∗+1)!∣s∣k∗+1.\|\mathbf{R}_{k_*+1}(s)\|_2\leq\frac{M_*}{(k_*+1)!}|s|^{k_*+1}.

Under the localization hypothesis,

∥H(a∗+s)∥2≥m∗(4k∗+3)4(k∗+1)!∣s∣k∗.\|\mathbf{H}(a_*+s)\|_2\geq\frac{m_*(4k_*+3)}{4(k_*+1)!}|s|^{k_*}.

On the other hand,

∥H(a^)∥2≤∥H(a^)−H^(a^)∥2+∥H^(a^)∥2≤ε∗+τ.\|\mathbf{H}(\widehat{a})\|_2\leq\|\mathbf{H}(\widehat{a})-\widehat{\mathbf{H}}(\widehat{a})\|_2+\|\widehat{\mathbf{H}}(\widehat{a})\|_2\leq\varepsilon_*+\tau.

Combining the two estimates yields (43). If M∗=0M_*=0, the remainder vanishes and the same bound follows directly.

Remark 7.2 (Loop-design consequence). The theorem gives a deterministic ordering principle: first minimize the isolation order kγk_\gamma; among loops of the minimal order, increase the collective coefficient m∗m_\ast while controlling loop length and transport conditioning. The residual τ\tau allows numerically computed loops that do not possess an exactly common perturbed fixed point.

Global reconstruction stability at finite holonomy order

The local anchor estimate propagates through the same scalar transport equation used for exact reconstruction.

Theorem 8.1 (Global finite-order reconstruction stability). Let k=kγ(a∗)≥2k=k_\gamma(a_\ast)\ge2 be finite. Assume:

(i) the exact and perturbed transports satisfy the uniform separation and boundedness hypotheses of Lemma 3.4 along paths connecting c0c_0 to every point of MM;

(ii) H^\hat H has a zero a^\hat a satisfying the localization hypotheses of Theorem 5.1 (for example, under the odd/even sufficient conditions of Proposition 5.2);

(iii) the holonomy perturbation obeys ε≤CγεF\varepsilon\le C_\gamma\varepsilon_F for the chosen identifying loop.

Let

DM:=diam⁡dM(M).D_M:=\operatorname{diam}_{d_M}(M).

Then

∥d^−d∥C0(M)≤Cglob(εFmk)1/k+CpropεF,(44)\lVert\hat d-d\rVert_{C^0(M)} \le C_{\mathrm{glob}}\left(\frac{\varepsilon_F}{m_k}\right)^{1/k} +C_{\mathrm{prop}}\varepsilon_F, \tag*{(44)}

where one may take

Cglob=eL0DMκk1/kCγ1/k,Cprop=C0DMeL0DM,C_{\mathrm{glob}}=e^{L_0D_M}\kappa_k^{1/k}C_\gamma^{1/k}, \qquad C_{\mathrm{prop}}=C_0D_Me^{L_0D_M},

with L0L_0 from (19). Hence, for fixed mk>0m_k>0,

∥d^−d∥C0(M)=O(εF1/k).\lVert\hat d-d\rVert_{C^0(M)}=O\left(\varepsilon_F^{1/k}\right).

Proof. For a path γc\gamma_c from c0c_0 to cc, the standard Gronwall comparison gives

∣d^(c)−d(c)∣≤eL0ℓ(γc)(∣a^−a∗∣+C0ℓ(γc)εF).|\hat d(c)-d(c)|\le e^{L_0\ell(\gamma_c)} \left(|\hat a-a_\ast|+C_0\ell(\gamma_c)\varepsilon_F\right).

Choose paths with ℓ(γc)≤DM\ell(\gamma_c)\le D_M. By Theorem 5.1,

∣a^−a∗∣≤κk1/k(εmk)1/k≤κk1/kCγ1/k(εFmk)1/k.|\hat a-a_\ast| \le\kappa_k^{1/k}\left(\frac{\varepsilon}{m_k}\right)^{1/k} \le\kappa_k^{1/k}C_\gamma^{1/k} \left(\frac{\varepsilon_F}{m_k}\right)^{1/k}.

Substitution yields (44).

Geometric interpretation and remaining geometric questions

The exact second- and third-variation formulas give rigorous geometric signals beyond the linearized multiplier: H′′(a∗)H''(a_\ast) and H′′′(a∗)H'''(a_\ast) are integrals of vertical derivatives of the nonlinear return transport along the true horizontal section. Theorem 4.3 extends this to every finite order, while Theorem 4.9 settles the local analytic meaning of the flat regime for a fixed loop. The material below is deliberately stated as geometric perspective and open problems; no unproved small-loop curvature expansion or genericity assertion is used in the theorem chain.

Connection-curvature perspective

The inverse transport can be written as the horizontal distribution of the one-form

ΘF=dz−ωF(c,z)\Theta_F = dz - \omega_F(c,z)

on the scalar bundle EF={(c,z):z>qF(c)}E_F = \{(c,z): z > q_F(c)\}, as in the nonlinear-holonomy formulation of [2]; the connection language is standard, see [11]. If

Xi=∂i+ωi(c,z)∂z,X_i = \partial_i + \omega_i(c,z)\partial_z,

then

[Xi,Xj]vert=(∂iωj−∂jωi+ωi∂zωj−ωj∂zωi)∂z.[X_i,X_j]_{\mathrm{vert}} = (\partial_i\omega_j - \partial_j\omega_i + \omega_i\partial_z\omega_j - \omega_j\partial_z\omega_i)\partial_z.

This suggests the scalar curvature two-form

KF=dMωF+ωF∧∂zωF.\mathcal{K}_F = d_M\omega_F + \omega_F \wedge\partial_z\omega_F.

A remaining invariant problem is to express small-loop holonomy-jet coefficients in terms of the vertical curvature jet ∂zKF\partial_z\mathcal{K}_F. Such a formula would connect the finite-order scalar singularity theory proved here to intrinsic local geometry of the inverse connection.

Spherical and non-umbilic perturbations

For a perturbation dλ=d0+λϕd_\lambda= d_0 + \lambda\phi, the true anchor is

a∗(λ)=dλ(c0)=d0+λϕ(c0),a_*(\lambda) = d_\lambda(c_0) = d_0 + \lambda\phi(c_0),

and the true horizontal section is y∗(t;λ)=dλ(γ(t))y_*(t;\lambda) = d_\lambda(\gamma(t)). Expansions in λ\lambda must therefore be performed along this moving section.

Likewise, a compact convex surface in R3\mathbb{R}^3 should not be modelled globally by two distinct constant principal curvatures. For a genuine non-umbilic core, such as an ellipsoid, an invariant leading coefficient must include the Levi-Civita connection terms of the principal frame. Deriving such a coefficient and proving generic non-vanishing remain open geometric problems. These questions are distinct from the vertical anchor-jet hierarchy, which is exact once the observed transport is fixed.

Neutral tuning

If a one-parameter family satisfies

H′(a∗;λ)=D1λ+O(λ2),D1≠0,H'(a_*;\lambda) = D_1\lambda+ O(\lambda^2), \qquad D_1 \ne0,

then the sign is fixed for sufficiently small positive λ\lambda; a nonzero linear coefficient prevents nearby neutral tuning. A generic search for neutral configurations should therefore use either a two-parameter family dλ,μ=d0+λϕ+μψd_{\lambda,\mu} = d_0 + \lambda\phi+ \mu\psi or a specially designed one-parameter family. Section § 10 shows that exact neutral tuning can nevertheless be achieved explicitly in a finite Fourier family on the circular core.

Numerical experiments

The numerical section has two purposes. First, it calibrates the finite-order scalar exponent without geometric complications. Second, it tests the actual return-map transport on genuine admissible geometries. All computations reported below were performed in double precision with Python 3.13.5, NumPy 2.3.5 and SciPy 1.17.0. The transport equations and their first two variational equations were integrated by DOP853 with relative tolerance 2×10−112 \times10^{-11} (and 2×10−122 \times10^{-12} for the exact baseline derivatives), absolute tolerance at most 10−1310^{-13}, and maximum step 0.02. Roots were refined by Brent’s method to relative and absolute tolerances of order 10−1310^{-13}. Geometric margins were first sampled on a grid of 20,000 angles and then enclosed independently by outward-rounded interval arithmetic (mpmath.iv) on 2,000 angular subintervals. A reference implementation accompanies the manuscript.

Algebraic finite-order calibration

For H(a)=akH(a) = a^k at a∗=0a_\ast= 0 and the perturbation H^ε(a)=ak−ε\widehat{H}_{\varepsilon}(a) = a^k - \varepsilon, the relevant real root has exact magnitude ε1/k\varepsilon^{1/k} (for even kk we take the positive branch). Thus the exponents 1/21/2, 1/31/3 and 1/41/4 are exact for k=2,3,4k = 2, 3, 4.

order kkexact displacementpredicted exponentobserved log–log slope
2ε1/2\varepsilon^{1/2}1/21/20.500000
3ε1/3\varepsilon^{1/3}1/31/30.333333
4ε1/4\varepsilon^{1/4}1/41/40.250000

Table 1. Exact scalar calibration of the finite-order exponent.

This benchmark only tests scalar singular-root geometry. The next experiments use return data generated by actual normal-incidence geometry.

Circular-core return geometry

Let CC be the unit disk and write

c(θ)=(cos⁡θ,sin⁡θ),τ(θ)=(−sin⁡θ,cos⁡θ),S=−I.c(\theta) = (\cos\theta, \sin\theta), \qquad\tau(\theta) = (-\sin\theta, \cos\theta), \qquad S = -I.

For a positive thickness d(θ)d(\theta) set

w(θ)=d′(θ)1+d(θ).w(\theta) = \frac{d'(\theta)}{1 + d(\theta)}.

The outer point is Φ(θ)=(1+d(θ))c(θ)\Phi(\theta) = (1 + d(\theta))c(\theta) and its inward normal is

n(θ)=−c(θ)+w(θ)τ(θ)1+w(θ)2.n(\theta) = \frac{-c(\theta) + w(\theta)\tau(\theta)}{\sqrt{1 + w(\theta)^2}}.

Writing the return decomposition as

F(θ)−c(θ)=u(θ)τ(θ)+q(θ)c(θ),(45)F(\theta) - c(\theta) = u(\theta)\tau(\theta) + q(\theta)c(\theta), \tag*{(45)}

one obtains

D(θ)=1−w(θ)2((1+d(θ))2−1),(46)D(\theta)=1-w(\theta)^2\left((1+d(\theta))^2-1\right), \tag*{(46)}
σ(θ)=1+d(θ)−D(θ)1+w(θ)2,(47)\sigma(\theta)=\frac{1+d(\theta)-\sqrt{D(\theta)}}{1+w(\theta)^2}, \tag*{(47)}
q(θ)=d(θ)−σ(θ),uF(θ):=u(θ)=σ(θ)w(θ).(48)q(\theta)=d(\theta)-\sigma(\theta),\qquad u_F(\theta):=u(\theta)=\sigma(\theta)w(\theta). \tag*{(48)}

The inverse transport around the positively oriented circle is therefore

y′(θ)=uF(θ)(1+y(θ))y(θ)−q(θ),0≤θ≤2π.(49)y'(\theta)=\frac{u_F(\theta)(1+y(\theta))}{y(\theta)-q(\theta)},\qquad0\leq\theta\leq2\pi. \tag*{(49)}

For the perturbation experiments we keep uFu_F fixed and add smooth measurement error to the observed normal component,

q^ρ(θ)=q(θ)−ρcos⁡4θ,εF=∥q^ρ−q∥∞=ρ.(50)\widehat{q}_{\rho}(\theta)=q(\theta)-\rho\cos4\theta,\qquad\varepsilon_F=\lVert\widehat{q}_{\rho}-q\rVert_{\infty}=\rho. \tag*{(50)}

The perturbed pair need not itself be an exact return map; it represents smooth observational error in the return components, which is precisely the setting of the stability theory.

A genuine non-neutral example

Take

dnn(θ)=0.8+0.2(cos⁡θ+sin⁡θ+cos⁡2θ+sin⁡2θ).(51)d_{\mathrm{nn}}(\theta)=0.8+0.2(\cos\theta+\sin\theta+\cos2\theta+\sin2\theta). \tag*{(51)}

The true anchor at θ=0\theta=0 is a∗=1.2a_*=1.2. Direct integration of (49) and its variational equations gives

H(a∗)=−5.6×10−15,H′(a∗)=2.94769678×10−2,H′′(a∗)=−1.39481188×10−1.H(a_*)=-5.6\times10^{-15},\qquad H'(a_*)=2.94769678\times10^{-2},\qquad H''(a_*)=-1.39481188\times10^{-1}.

The geometry is well separated from degeneracy. A 20,000-point grid gives

min⁡dnn=0.38140,min⁡D=0.50373,min⁡σ=0.38140,max⁡⟨n(Φ),ν(F)⟩=−0.65104.\min d_{\mathrm{nn}}=0.38140,\qquad\min D=0.50373,\qquad\min\sigma=0.38140,\qquad\max\langle n(\Phi),\nu(F)\rangle=-0.65104.

The independent interval enclosure on 2,000 angular cells yields the outward-rounded bounds

dnn>0.38022,D>0.49950,σ>0.38022,⟨n(Φ),ν(F)⟩<−0.64323.d_{\mathrm{nn}}>0.38022,\qquad D>0.49950,\qquad\sigma>0.38022,\qquad\langle n(\Phi),\nu(F)\rangle<-0.64323.

Thus the normal graph stays embedded and the first-hit discriminant is uniformly positive. The transversality is uniformly negative throughout the circle.

The perturbation response at the true anchor satisfies numerically

∂ρH^ρ(a∗)∣ρ=0=0.13825171.\left.\partial_{\rho}\widehat{H}_{\rho}(a_*)\right|_{\rho=0}=0.13825171.

First-order root perturbation therefore predicts

∣a^−a∗∣∼0.138251710.0294769678ρ=4.69016 ρ.|\widehat{a}-a_*|\sim\frac{0.13825171}{0.0294769678}\rho=4.69016\,\rho.

At ρ=10−8\rho=10^{-8} the observed coefficient is 4.69017, agreeing with the predicted leading constant to the displayed precision. For the perturbation (50), Table 2 shows the reconstructed anchor and global thickness errors. A least-squares fit over all nine noise levels 10−8≤ρ≤10−410^{-8}\leq\rho\leq10^{-4} gives slopes 1.00018 and 1.00023, respectively.

ρ=εF\rho= \varepsilon_F∣a^−a∗∣\lvert\hat{a}-a_*\rvert∥d^−d∥∞\lVert\hat{d}-d\rVert_\infty
10−810^{-8}4.69017×10−84.69017 \times10^{-8}9.02152×10−89.02152 \times10^{-8}
10−710^{-7}4.69017×10−74.69017 \times10^{-7}9.02153×10−79.02153 \times10^{-7}
10−610^{-6}4.69027×10−64.69027 \times10^{-6}9.02178×10−69.02178 \times10^{-6}
10−510^{-5}4.69130×10−54.69130 \times10^{-5}9.02430×10−59.02430 \times10^{-5}
10−410^{-4}4.70162×10−44.70162 \times10^{-4}9.04963×10−49.04963 \times10^{-4}

Table 2. Non-neutral geometric experiment.

Log-log plot of local anchor error and global thickness error versus return-data perturbation

Figure 1. Non-neutral return geometry: both local anchor error and global thickness error are first order in the return-data perturbation.

A numerically tuned neutral example

Consider the one-parameter Fourier family

dα(θ)=0.8+αcos⁡θ+0.12866554sin⁡θ(52)d_\alpha(\theta) = 0.8 + \alpha\cos\theta+ 0.12866554\sin\theta \tag*{(52)}
dα(θ)−0.16790919cos⁡2θ+0.08267596sin⁡2θ−0.11676398cos⁡3θ.(53)\phantom{d_\alpha(\theta)}{} - 0.16790919\cos2\theta+ 0.08267596\sin2\theta- 0.11676398\cos3\theta. \tag*{(53)}

We tune the first coefficient by solving the scalar neutrality equation

Hα′(a∗(α))=0H'_\alpha(a_*(\alpha)) = 0

with Brent’s method on [0.14476636,0.14476637][0.14476636, 0.14476637]. This gives

α∗=0.14476636480382166.\alpha_* = 0.14476636480382166.

We then set

dneu:=dα∗.(54)d_{\mathrm{neu}} := d_{\alpha_*}. \tag*{(54)}

At the base point c0=(1,0)c_0 = (1,0),

a∗=dneu(0)=0.6600931948038217.a_* = d_{\mathrm{neu}}(0) = 0.6600931948038217.

The tuned loop transport is neutral to machine precision,

H(a∗)=−2.2×10−16,H′(a∗)=−8.9×10−16,H(a_*) = -2.2 \times10^{-16}, \qquad H'(a_*) = -8.9 \times10^{-16},

while the exact second-variation formula gives

H′′(a∗)=3.13933613×10−3≠0.(55)H^{\prime\prime}(a_*) = 3.13933613 \times10^{-3} \ne0. \tag*{(55)}

The geometry again has substantial admissibility margins. The 20,000-point grid gives

min⁡dneu=0.58204,min⁡D=0.56673,min⁡σ=0.58204,max⁡⟨n(Φ),ν(F)⟩=−0.70167.\min d_{\mathrm{neu}} = 0.58204,\quad\min D = 0.56673,\quad\min\sigma= 0.58204,\quad\max\langle n(\Phi),\nu(F)\rangle= -0.70167.

The independent interval enclosure on 2,000 angular cells gives

dneu>0.58064,D>0.56375,σ>0.58064,⟨n(Φ),ν(F)⟩<−0.69535.d_{\mathrm{neu}} > 0.58064,\quad D > 0.56375,\quad\sigma> 0.58064,\quad\langle n(\Phi),\nu(F)\rangle< -0.69535.

Hence the tuned normal graph is uniformly admissible for the numerical return construction. This is a numerically tuned second-order neutral holonomy generated by the return geometry, not an imposed scalar model.

For the perturbation (50), the residual at the true anchor has the linear response

∂ρH^ρ(a∗)∣ρ=0=−0.16119036.\left.\partial_{\rho}\widehat{H}_{\rho}(a_*)\right|_{\rho=0} = -0.16119036.

Combining this with (55) gives the quadratic normal-form prediction

∣a^−a∗∣∼2(0.16119036)0.00313933613ρ=10.13364ρ.|\hat{a}-a_*| \sim\sqrt{\frac{2(0.16119036)}{0.00313933613}}\sqrt{\rho} = 10.13364\sqrt{\rho}.

At ρ=10−8\rho= 10^{-8} the observed upper-branch coefficient is 10.1268, again matching the leading constant as well as the exponent.

Figure 2 displays the core, the Fourier outer boundary and several actual inward-normal return segments. The segment from c0c_0 to Φ(c0)\Phi(c_0) has length a∗a_* and the arrow on the core represents the closed loop used in the scalar transport.

Figure 2. TikZ rendering of the numerically tuned neutral circular-core geometry (54). The arrows are actual inward-normal return segments for five base points.

Under the same perturbation (50), the even-order neutral root splits into two branches. Table 3 reports both anchor displacements and the global thickness error for the upper branch. The fitted slopes over the complete range 10−8≤ρ≤3×10−510^{-8} \le\rho\le3 \times10^{-5} are

ρ=εF\rho= \varepsilon_Flower root errorupper root error∣d^−d∣∞|\hat{d}-d|_{\infty}
10−810^{-8}1.01404×10−31.01404 \times10^{-3}1.01268×10−31.01268 \times10^{-3}1.09443×10−31.09443 \times10^{-3}
10−710^{-7}3.21135×10−33.21135 \times10^{-3}3.19773×10−33.19773 \times10^{-3}3.45493×10−33.45493 \times10^{-3}
10−610^{-6}1.02017×10−21.02017 \times10^{-2}1.00656×10−21.00656 \times10^{-2}1.08659×10−21.08659 \times10^{-2}
10−510^{-5}3.27262×10−23.27262 \times10^{-2}3.13661×10−23.13661 \times10^{-2}3.37753×10−23.37753 \times10^{-2}
3×10−53 \times10^{-5}5.75431×10−25.75431 \times10^{-2}5.34719×10−25.34719 \times10^{-2}5.74398×10−25.74398 \times10^{-2}

Table 3. Tuned neutral geometric experiment.

0.49597for ∣a^−a∗∣,0.49528for ∥d^−d∥∞,0.49597\quad\text{for }|\hat{a}-a_*|,\qquad0.49528\quad\text{for }\|\hat{d}-d\|_{\infty},

in agreement with the theoretical exponent 1/21/2.

The observed separation min⁡θ(d^(θ)−q^ρ(θ))\min_{\theta}(\hat{d}(\theta)-\hat{q}_{\rho}(\theta)) remains above 0.58 at the smallest noise levels and above 0.63 at ρ=3×10−5\rho= 3 \times10^{-5}, so the experiment stays well inside the uniformly separated regime of the transport estimates. The numerics therefore validate not only the scalar root exponent but the full chain from perturbed return components to anchor reconstruction and global thickness propagation.

Conclusion and open problems

The paper develops a finite-order singular stability theory for the anchor parameter in the nonlinear holonomy formulation of the return-map inverse problem. The proved chain is

return-map perturbation⟹finite holonomy-jet perturbation,⟹finite isolation order⟹anchor stability,⟹multi-loop aggregation⟹global thickness stability.\begin{aligned} \text{return-map perturbation} &\Longrightarrow\text{finite holonomy-jet perturbation},\\ &\Longrightarrow\text{finite isolation order} \Longrightarrow\text{anchor stability},\\ &\Longrightarrow\text{multi-loop aggregation} \Longrightarrow\text{global thickness stability}. \end{aligned}

At first order, the non-neutral regime is locally Lipschitz stable. At finite neutral order k≥2k \ge2, the anchor obeys a 1/k1/k-Hölder location estimate with parity-sensitive existence. Exact uniform separation is stable under sufficiently small return-data perturbations, so the singular fibre set enters the stability theory through an explicit quantitative conditioning margin. The exact second and third holonomy variations and the Bell-polynomial hierarchy identify the relevant coefficients as vertical derivatives of the geometric transport. For smooth finite-dimensional families, Malgrange preparation further shows that an order-kk holonomy singularity is locally equivalent, up to multiplication by a nonvanishing smooth factor, to a monic degree-kk polynomial in the anchor displacement; in the quadratic case Proposition 4.7 identifies the first-order prepared coefficients and discriminant correctly, including the contribution of the cubic anchor jet to c1c_1. The finite-jet perturbation theorem makes the holonomy coefficients quantitatively observable from perturbed return data, while the multi-loop theorem combines several closed-loop signals without a statistical-independence assumption. Finally, the analytic-dependence theorem shows that infinite-order flatness for a fixed uniformly separated loop means local identity holonomy, rather than an arbitrary C∞C^\infty flat germ. The numerical experiments support the theorem chain on genuine return geometries. A non-neutral circular example gives first-order anchor and global reconstruction rates, while a numerically tuned neutral Fourier geometry has H′(a∗)H'(a_*) at machine zero, H′′(a∗)≠0H''(a_*) \ne0 and yields observed rates close to 1/21/2 for both the anchor and the full reconstructed thickness. In both regimes the observed leading anchor constant agrees with the corresponding local normal-form prediction.

Log-log plot of anchor error $|\hat{a}-a_*|$ and globally reconstructed thickness error $\|\hat{d}-d\|_{\infty}$ versus $\varepsilon_F$, with reference slope $1/2$

Figure 3. Tuned neutral return geometry: the anchor and globally reconstructed thickness both exhibit the square-root stability law.

The principal remaining questions are geometric rather than scalar:

  1. derive invariant formulas relating small-loop holonomy jets to the curvature and vertical curvature jet of the nonlinear inverse connection;

  1. determine generic non-vanishing of those coefficients on genuine non-umbilic compact cores, including ellipsoids, without suppressing Levi-Civita connection terms;

  1. characterize configurations for which all relevant loop holonomies are locally the identity and decide whether global admissibility or topology still isolates the anchor;

  1. optimize loop families by combining isolation order, collective jet amplitude, loop length and transport conditioning;

  1. extend the reconstruction and stability theory to lower regularity, for example d∈C1,1(M)d \in C^{1,1}(M);

  1. develop a range characterization for noisy candidate return maps that couples the transport residuals with geometric admissibility of the reconstructed normal graph.

References

  1. [1]M. Barkatou, Some geometric properties for a class of non-Lipschitz domains, New York J. Math. 8 (2002), 189–213.
  2. [2]M. Barkatou and M. El Morsalani, Inverse Problems for the Return Map in the Class O_C: Reconstruction and Identifiability, preprint (2026), arXiv:2604.24915 [math.DS].arxiv.org/abs/2604.24915
  3. [3]M. Barkatou and M. El Morsalani, The Return Map in the Class O_C: Geometry, Dynamics, and Thickness Regularity, Journal of Nonlinear Dynamics and Applications 2 (2026), no. 3, 192–216.
  4. [4]J. Bohr, Stability of the non-abelian X-ray transform in dimension ≥ 3, J. Geom. Anal. 31 (2021), 11226–11269. https://doi.org/10.1007/s12220-021-00679-0.
  5. [5]M. Cekić, C. Guillarmou, and T. Lefeuvre, Local lens rigidity for manifolds of Anosov type, Anal. PDE 17 (2024), no. 8, 2737–2795. https://doi.org/10.2140/apde.2024.17.2737.
  6. [6]M. Cekić and T. Lefeuvre, Stability estimates for the holonomy inverse problem, Comm. Partial Differential Equations 49 (2024), no. 5–6, 428–469. https://doi.org/10.1080/03605302.2024.2346942.
  7. [7]M. Cekić and T. Lefeuvre, The holonomy inverse problem, J. Eur. Math. Soc. 27 (2025), no. 6, 2187–2250. https://doi.org/10.4171/JEMS/1409.
  8. [8]M. Golubitsky and V. Guillemin, Stable Mappings and Their Singularities, Graduate Texts in Mathematics, vol. 14, Springer-Verlag, New York, 1973; Chapter IV, “The Malgrange Preparation Theorem,” pp. 91–110. https://doi.org/10.1007/978-1-4615-7904-5.
  9. [9]H. Grebnev, The non-abelian X-ray transform on asymptotically hyperbolic spaces, J. Geom. Anal. 36 (2026), Article 73. https://doi.org/10.1007/s12220-025-02296-7.
  10. [10]S. Jafarpour and A. D. Lewis, Time-Varying Vector Fields and Their Flows, SpringerBriefs in Mathematics, Springer, Cham, 2014. https://doi.org/10.1007/978-3-319-10139-2.
  11. [11]S. Kobayashi and K. Nomizu, Foundations of Differential Geometry, Vol. I, Interscience Publishers, 1963.
  12. [12]G. P. Paternain, M. Salo, and G. Uhlmann, Geometric Inverse Problems: With Emphasis on Two Dimensions, Cambridge Studies in Advanced Mathematics 204, Cambridge University Press, 2023.
  13. [13]Z. Zeng, Computing multiple roots of inexact polynomials, Math. Comp. 74 (2005), no. 250, 869–903. https://doi.org/10.1090/S0025-5718-04-01692-8.

Paper details

vlnc:W_kzndc963
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents