betaFeedback

WorksMathematics

Perimeter-Regularized Critical Shapes in the Class OC\mathcal{O}_C: Core-Normal Transfer and First-Order Return-Map Response

Abstract

Let C⊂RNC \subset \mathbb{R}^N be a compact strictly convex core and OC\mathcal{O}_C denote the class of domains whose outer boundary is a global outward normal graph over ∂C\partial C and whose inward boundary normals return to the interior of CC. We study volume-constrained critical domains of the Dirichlet-perimeter functional: Eμ(Ω)=−12∫Ω∣∇uΩ∣2 dx+μP(Ω),−ΔuΩ=f,uΩ∣∂Ω=0, \mathcal{E}_μ(Ω) = -\frac{1}{2}\int_Ω|\nabla u_Ω|^{2}\,\mathrm{d}x + μP(Ω), \qquad -Δu_Ω = f, \quad u_Ω|_{\partialΩ}=0, in the large-perimeter regime μ→∞μ\to\infty. We give an adapted Lyapunov–Schmidt construction near translated spheres to extract the first physical shape coefficient of scale μ−1μ^{-1}. We then prove that OC\mathcal{O}_C is open under small C1C^1 thickness perturbations and establish a quantitative coordinate-transfer theorem to a unique core-normal thickness. Finally, we prove that the forward return-map operator d↦Fdd \mapsto F_d is Fréchet C1C^1 between Hölder spaces and provide its explicit derivative. This yields a shape-to-dynamics sensitivity theorem for OC\mathcal{O}_C: a perimeter-regularized PDE-constrained critical deformation produces an explicitly controlled first-order deformation of the return dynamics. A nonradial example confirms that the response coefficient is generically nonzero.

Introduction

Problem, positioning, and guiding obstruction

Let C⊂RNC \subset\mathbb{R}^{N}, N≥2N \ge2, be a compact strictly convex body. The class OC\mathcal{O}_{C}, originating in the geometric normal-property framework of Barkatou, consists here of domains surrounding CC whose outer boundary is a global normal graph over ∂C\partial C and whose inward outer-boundary normals return to int⁡C\operatorname{int} C; see [3, 5, 7]. The geometry carries a return map

Fd:∂C⟶∂C,F_{d}:\partial C \longrightarrow\partial C,

obtained by moving first along the outward core normal to the outer boundary and then back to the core along the inward normal of that outer boundary. We study the interaction of this return geometry with the volume-constrained Dirichlet-perimeter functional

Eμ(Ω)=−12∫Ω∣∇uΩ∣2 dx+μP(Ω),μ>0,(1)\mathcal{E}_{\mu}(\Omega)=-\frac{1}{2}\int_{\Omega}|\nabla u_{\Omega}|^{2}\,dx+\mu P(\Omega), \qquad\mu>0, \tag*{(1)}

where uΩu_{\Omega} solves

−ΔuΩ=fin Ω,uΩ=0on ∂Ω,(2)-\Delta u_{\Omega}=f \quad\text{in }\Omega,\qquad u_{\Omega}=0 \quad\text{on }\partial\Omega, \tag*{(2)}

and νΩ\nu_{\Omega} denotes the outward unit normal. We impose

∣Ω∣=V0.(3)|\Omega|=V_{0}. \tag*{(3)}

The source ff is fixed in physical space.

The large-μ\mu regime is dominated by perimeter. A necessary-limit argument shows that any connected smooth C2C^{2} limit of critical domains with uniformly bounded boundary flux must have constant mean curvature and is therefore spherical. Thus the correct reference object is a selected ball of volume V0V_{0}, not an arbitrary nonspherical core. The convex core remains essential in a different role: it provides the global normal coordinates and the return geometry through which the critical branch is represented and observed.

What is inherited from shape optimization and what is new here

The translated-ball existence mechanism itself is not new. Hayouni and Novruzi [13] used the implicit-function theorem to construct regular critical points for a two-dimensional Dirichlet-perimeter free-boundary problem for large perimeter weight. Novruzi [16] subsequently developed C2,αC^{2,\alpha} existence theory for a broader class of shape functionals, explicitly separating translations from genuine shape deformations and using nondegeneracy of the reduced translated-ball energy. Related free-boundary work by Barkatou, Seck and Ly [4] treated an interior electromagnetic-casting problem through shape optimization and curvature boundary conditions.

Accordingly, the Lyapunov–Schmidt construction in § 3–§ 5 is included as a self-contained preparatory layer adapted to the present notation and to the first-order coefficient required later. The principal new results begin when this classical branch is connected to OC\mathcal{O}_{C}:

  • (N1) a C1C^{1}-openness theorem for the geometric normal-return class OC\mathcal{O}_{C} (Theorem 6.1);

  • (N2) a quantitative coordinate transfer from a translated spherical graph to a unique core-normal thickness, including the joint contribution of center drift and higher spherical modes (Theorem 7.2);

  • (N3) Fréchet C1C^{1} differentiability of the forward return operator R:d↦Fd\mathcal{R}:d\mapsto F_{d}, with an explicit derivative obtained from the first-hit equation (Theorem 8.2).

The paper therefore does not claim a new general existence theory for perimeter-regularized critical shapes. Its novelty is the shape-to-return-dynamics sensitivity theory in the class OC\mathcal{O}_{C}.

Relation with the preceding OCO_C papers

The present paper is the next step in a sequence with a progressively changing object of study. The first return-map paper [5] established the exact normal geometry of d↦Fdd \mapsto F_d, the fixed-point/critical-point relation, and the spatial linearization DcFdD_cF_d at fixed points. The inverse paper [6] studies the opposite direction, reconstructing thickness information from a given return map through an exact nonlinear transport system and its linearized compatibility structure. The elliptic superlevel-set paper [7] keeps the PDE domain fixed and studies one distinguished deformation path

dt=d0−tb0+oC1(t),Ft=F0+tA0+o(t),d_t = d_0 - tb_0 + o_{C^1}(t), \qquad F_t = F_0 + tA_0 + o(t),

generated by small Dirichlet superlevels.

The present problem is complementary to all three. The domain itself is varied by an Euler–Lagrange condition, and the main forward object is now the operator

R:d⟼Fd.\mathcal{R}: d \longmapsto F_d.

Theorem Theorem 8.2 differentiates this operator with respect to an arbitrary thickness perturbation. Hence the earlier PDE coefficient A0A_0 becomes one special direction of a general Fréchet shape derivative, while the inverse-paper compatibility system becomes a natural target for a future range characterization of R⟦⟦′\mathcal{R}_{\llbracket\llbracket'}.

Comparison with recent work

Recent shape-optimization literature emphasizes several complementary questions. Buttazzo–Maiale–Mazzoleni–Tortone–Velichkov [8] develop free-boundary regularity for broad PDE-dependent integral shape functionals. Celentano–Nitsch–Trombetti [10] use overdetermined elliptic systems to characterize critical shapes under volume constraints. Caubet–Dambrine–Darde [9] show that perimeter penalization can improve stability in a Dirichlet inverse-obstacle shape problem. Plotnikov–Sokolowski [17] study geometric gradient-flow frameworks for regularized inverse and free-boundary shape problems, and Amato–Nitsch–Trombetti–Villone [2] give recent local stability results for functionals combining Dirichlet spectral/torsional quantities with perimeter constraints.

These works concern existence, regularity, stability, symmetry, or optimization dynamics of the shape itself. The question addressed here is different: once a smooth critical-shape branch is available, how does an additional geometric observable—the normal return map attached to a convex core—change with the shape? The new object DR[d0]D\mathcal{R}[d_0] is therefore a response operator from boundary deformation to discrete return dynamics rather than a shape Hessian or a gradient-flow generator.

Two adjacent theories should be distinguished from this claim. In classical Poincaré/first-impact theory, transversality of a flow with a section yields differentiability of the return time and hence of the first-return map by an implicit-function argument; see, for example, Immler–Traut [15]. Thus the projection that removes the component generated by variation of the hitting time is not, by itself, new. In shape calculus, Feppon–Allaire–Dapogny [12] study shape derivatives of geometric constraints involving signed distance and integration along normal rays. Their normal-ray computations are closely related to our equations (8)–(53), but they do not address return maps.

Theorem hierarchy and main result

To make the novelty hierarchy explicit, we distinguish one preparatory theorem from the principal geometric results.

WorkMain issueRelation to the present paper
Hayouni–Novruzi (2002); Novruzi [16]Large-perimeter critical/optimal domains; translation versus pure deformationClassical existence mechanism used as preparatory background; no OC\mathcal{O}_C return response
Barkatou–Seck–Ly (2006)Curvature free-boundary condition and electromagnetic castingMotivates the PDE-perimeter model and applications; no normal-return sensitivity
Buttazzo et al. [8]Regularity of PDE-dependent optimal setsGlobal/free-boundary regularity, complementary to our local smooth branch
Caubet–Dambrine–Darde (2026)Perimeter regularization in inverse obstacle optimizationShows current relevance of perimeter stabilization; different observable and inverse problem
Earlier OC\mathcal{O}_C papersGeometry, inverse reconstruction, and one PDE-generated deformation of FdF_dSupply the exact return geometry; the present work differentiates the full forward operator d↦Fdd \mapsto F_d
Present paperCore-normal transfer and first-order return-map responseFréchet response operator and combined critical-shape-to-dynamics theorem

Table 1. Position of the present results relative to closely related work.

  1. Preparatory selected-ball branch. We adapt the classical translated-ball implicit-function/Lyapunov–Schmidt framework and extract the physical μ−1\mu^{-1} coefficient

JRψ1=12PX(g02).J_R\psi_1 = \frac{1}{2}P_X(g_0^2).
  1. OC\mathcal{O}_C openness and transfer. Small C1C^1 thickness perturbations preserve OC\mathcal{O}_C, and a nearby translated spherical graph has a unique core-normal representation.

  1. Principal response theorem. The forward return operator is Fréchet C1C^1, with an explicit derivative DR[d0]\mathcal{D}R[d_0].

  1. Combined shape-to-dynamics theorem. Along the critical branch,

dμ=d0+μ−1η1+oC2,α(μ−1),Fμ=F0+μ−1DR[d0]η1+oC1,α(μ−1).d_\mu= d_0 + \mu^{-1}\eta_1 + o_{C^{2,\alpha}}(\mu^{-1}), \qquad F_\mu= F_0 + \mu^{-1}\mathcal{D}R[d_0]\eta_1 + o_{C^{1,\alpha}}(\mu^{-1}).

No theorem below uses numerical evidence. Exact geometric identities are established before asymptotic expansions; every asymptotic statement is labelled by its topology and small parameter.

Potential applications

The applications are secondary to the geometric analysis, but the theorem chain has several natural interpretations. First, Dirichlet-perimeter free-boundary conditions arise historically in electromagnetic casting and shaping, where curvature regularization competes with a PDE-generated boundary pressure; see [13, 4]. Second, perimeter penalties remain active in contemporary inverse-shape problems because they can improve stability and numerical regularity [9]. Third, the new operator DR[d0]\mathcal{D}R[d_0] provides a sensitivity map from boundary design variables to a geometric observable. This can be relevant whenever an interface is probed or controlled through normal incidence/return information, for example in inverse geometry, geometric metrology, or PDE-constrained interface design. We do not claim an application-specific sensing model here; the contribution is the mathematical response operator that such models could use.

Methodological conventions and reproducibility

Several conventions are fixed globally to avoid ambiguities that are particularly damaging in normal-return problems. We use

S=−Dν∣TMS=-D\nu|_{TM}

so that a convex core has S≤0S\le0 and I−dS≥II-dS\ge I for every outward thickness d≥0d\ge0. Outward and inward normals are denoted by NdN_d and nd=−Ndn_d=-N_d, respectively. Exact identities, first-order expansions, and heuristic interpretations are kept separate. The analytical benchmark in § 11 is exact. A computational protocol is stated there for a nonradial test, but no numerical result is used as evidence for any theorem.

Scope and nonclaims

The results are local. We construct constrained critical domains; we do not claim global minimality. The existence of a selected nondegenerate center and of a containing reference sphere of the prescribed volume is assumed (see Remark 9.2). We do not claim that Ran⁡DR[d0]\operatorname{Ran} DR[d_0] equals the full compatibility space of the linearized inverse problem; a precise conjecture is stated in § 12. Finally, the preparatory large-perimeter branch is presented as an adaptation of established shape-optimization machinery, not as a new general existence theorem.

Geometric and variational setup

The class OC\mathcal{O}_C

Let

M:=∂C∈C4,α,0<α<1,M:=\partial C\in C^{4,\alpha},\qquad0<\alpha<1,

and let ν:M→SN−1\nu:M\to\mathbb{S}^{N-1} be the outward unit normal. We adopt the convention

S=−Dν∣TM(4)S=-D\nu|_{TM} \tag*{(4)}

For a convex hypersurface the eigenvalues of SS are −κ1,…,−κN−1-\kappa_1,\ldots,-\kappa_{N-1}, hence S≤0S\le0.

For d:M→(0,∞)d:M\to(0,\infty) define

Ψd(c)=c+d(c)ν(c)(5)\Psi_d(c)=c+d(c)\nu(c) \tag*{(5)}

and

Ω[d]=C∪{c+rν(c):c∈M, 0<r<d(c)}.(6)\Omega[d]=C\cup\{c+r\nu(c):c\in M,\ 0<r<d(c)\}. \tag*{(6)}

Definition 2.1 (The class OC\mathcal{O}_C). A domain Ω[d]\Omega[d] with d∈C2,α(M)d\in C^{2,\alpha}(M) and d>0d>0 belongs to OC\mathcal{O}_C if every inward normal ray issued from ∂Ω[d]\partial\Omega[d] intersects int⁡C\operatorname{int} C.

The global normal-graph property is already built into (6). Set

wd=Ld−1∇Md,χd=1+∣wd∣2.(7)w_d=L_d^{-1}\nabla_M d,\qquad\chi_d=\sqrt{1+|w_d|^2}. \tag*{(7)}

The exact outward and inward unit normals are

Nd=ν−wdχd,nd=−Nd=−ν+wdχd.(8)N_d=\frac{\nu-w_d}{\chi_d},\qquad n_d=-N_d=\frac{-\nu+w_d}{\chi_d}. \tag*{(8)}

These formulas are exact and are the starting point for all later differentiations; compare [7].

Energy and Euler–Lagrange equation

Fix

f∈Cc4,α(RN),f≥0,f≠0.(9)f \in C_c^{4,\alpha}(\mathbb{R}^{N}), \qquad f \ge0, \qquad f \ne0. \tag*{(9)}

For each smooth domain containing supp⁡f\operatorname{supp} f, let uΩu_{\Omega} solve (2).

For a smooth deformation with normal velocity VNV_N, the Hadamard formula (see, e.g., [11, 14, 18]) gives

D(−12∫Ω∣∇uΩ∣2)[VN]=−12∫∂Ω(∂NuΩ)2VN dS.(10)D\left(-\frac{1}{2}\int_{\Omega} \lvert\nabla u_{\Omega}\rvert^2\right)[V_N] = -\frac{1}{2}\int_{\partial\Omega}(\partial_N u_{\Omega})^2 V_N\,\mathrm{d}S. \tag*{(10)}

With the convention H=div⁡∂ΩNH=\operatorname{div}_{\partial\Omega}N, the perimeter and volume derivatives are

DP(Ω)[VN]=∫∂ΩHVN dS,D∣Ω∣[VN]=∫∂ΩVN dS.DP(\Omega)[V_N]=\int_{\partial\Omega}HV_N\,\mathrm{d}S,\qquad D\lvert\Omega\rvert[V_N]=\int_{\partial\Omega}V_N\,\mathrm{d}S.

Thus a smooth volume-constrained critical domain satisfies

−12g2+μH+Λμ=0,g:=−∂NuΩ.(11)-\frac{1}{2}g^2+\mu H+\Lambda_{\mu}=0,\qquad g:=-\partial_Nu_{\Omega}. \tag*{(11)}

Set

ε=μ−1,λε=Λμ/μ.(12)\varepsilon=\mu^{-1},\qquad\lambda_{\varepsilon}=\Lambda_{\mu}/\mu. \tag*{(12)}

Then

H+λε−ε2g2=0.(13)H+\lambda_{\varepsilon}-\frac{\varepsilon}{2}g^2=0. \tag*{(13)}

Rigidity of a smooth large-μ\mu limit

Proposition 2.2 (CMC obstruction). Let Ωμj\Omega_{\mu_j} be smooth volume-constrained critical domains for (1), with μj→∞\mu_j\to\infty. Assume

∂Ωμj→Σ0in C2\partial\Omega_{\mu_j}\to\Sigma_0\quad\text{in }C^2

and

sup⁡j∥∂Njuμj∥L∞(∂Ωμj)<∞.\sup_j\left\lVert\partial_{N_j}u_{\mu_j}\right\rVert_{L^\infty(\partial\Omega_{\mu_j})}<\infty.

Then Σ0\Sigma_0 has constant mean curvature.

Proof. Subtract the surface average from (11):

μj(Hj−Hj‾)=12(gj2−gj2‾).\mu_j(H_j-\overline{H_j})=\frac{1}{2}(g_j^2-\overline{g_j^2}).

The flux bound implies

∥Hj−Hj‾∥L∞≤Cμj→0.\left\lVert H_j-\overline{H_j}\right\rVert_{L^\infty}\le\frac{C}{\mu_j}\to0.

The C2C^2 convergence gives uniform convergence of the mean curvatures, so HΣ0H_{\Sigma_0} is constant.

Remark 2.3. In the compact connected embedded setting relevant here, a closed constant-mean-curvature hypersurface is spherical by Alexandrov’s theorem [1]. Proposition 2.2 is only a necessary-limit statement; it does not by itself provide compactness or existence of a large-μ\mu branch.

Selected spherical reference states: preparatory reduction

Fix R>0R>0 by

∣BR∣=V0.(14)|B_R|=V_0. \tag*{(14)}

Let U⊂RNU\subset\mathbb{R}^N be a neighborhood of a point x0x_0 such that

supp⁡f⋐BR−δ(x)for every x∈U(15)\operatorname{supp} f\Subset B_{R-\delta}(x)\qquad\text{for every }x\in U \tag*{(15)}

for some δ>0\delta>0.

For x∈Ux\in U, let uxu_x solve

−Δux=fin BR(x),ux=0on ∂BR(x).-\Delta u_x=f\quad\text{in }B_R(x),\qquad u_x=0\quad\text{on }\partial B_R(x).

Define the translated-ball reduced energy

h(x)=−12∫BR(x)∣∇ux∣2 dy.(16)h(x)=-\frac{1}{2}\int_{B_R(x)}|\nabla u_x|^2\,\mathrm{d}y. \tag*{(16)}

Lemma 3.1 (Gradient of the reduced energy). Let

gx(ω)=−∂Nux(x+Rω).g_x(\omega)=-\partial_Nu_x(x+R\omega).

Then

∇h(x)=−RN−12∫SN−1gx(ω)2ω dω.(17)\nabla h(x)=-\frac{R^{N-1}}{2}\int_{S^{N-1}}g_x(\omega)^2\omega\,\mathrm{d}\omega. \tag*{(17)}

Proof. Translate the ball in the direction a∈RNa\in\mathbb{R}^N. The normal velocity is a⋅ωa\cdot\omega. The Hadamard formula (10) gives

Dh(x)[a]=−12∫∂BR(x)gx2(a⋅N) dS,Dh(x)[a]=-\frac{1}{2}\int_{\partial B_R(x)}g_x^2(a\cdot N)\,\mathrm{d}S,

which is equivalent to (17). □

Assumption 3.2 (Selected nondegenerate center). There exists x0∈Ux_0\in U such that

∇h(x0)=0,det⁡D2h(x0)≠0.(18)\nabla h(x_0)=0,\qquad\det D^2h(x_0)\ne0. \tag*{(18)}

The nondegeneracy is the finite-dimensional hypothesis needed to resolve the translation kernel of the spherical Jacobi operator.

Exact-volume chart and smooth PDE dependence

Let σN−1=∣SN−1∣\sigma_{N-1}=|S^{N-1}|. Define

Xk,α={ψ∈Ck,α(SN−1):∫SN−1ψ dω=0, ∫SN−1ψω dω=0}.(19)X^{k,\alpha}=\left\{\psi\in C^{k,\alpha}(S^{N-1}):\int_{S^{N-1}}\psi\,\mathrm{d}\omega=0,\ \int_{S^{N-1}}\psi\omega\,\mathrm{d}\omega=0\right\}. \tag*{(19)}

Exact volume preservation

For x∈Ux\in U and ψ∈X3,α\psi\in X^{3,\alpha}, consider

Σx,ψ={x+[R+a(ψ)+ψ(ω)]ω:ω∈SN−1}.(20)\Sigma_{x,\psi}=\{x+[R+a(\psi)+\psi(\omega)]\omega:\omega\in S^{N-1}\}. \tag*{(20)}

Lemma 4.1 (Exact volume chart). There is a neighborhood of 00 in X3,αX^{3,\alpha} and a unique smooth scalar function a(ψ)a(\psi) such that the domain bounded by (20) has volume V0V_0. Moreover

a(0)=0,Da(0)=0,(21)a(0)=0,\qquad Da(0)=0, \tag*{(21)}

and

a(ψ)=−N−12RσN−1∫SN−1ψ2 dω+O(∥ψ∥C3,α3).(22)a(\psi)=-\frac{N-1}{2R\sigma_{N-1}}\int_{S^{N-1}}\psi^2\,\mathrm{d}\omega+O(\|\psi\|_{C^{3,\alpha}}^3). \tag*{(22)}

Proof. The enclosed volume is

V(a,ψ)=1N∫SN−1(R+a+ψ)N dω.V(a,\psi)=\frac{1}{N}\int_{S^{N-1}}(R+a+\psi)^N\,d\omega.

At (0,0)(0,0),

∂aV(0,0)=RN−1σN−1>0.\partial_a V(0,0)=R^{N-1}\sigma_{N-1}>0.

The implicit-function theorem gives a(ψ)a(\psi). Since ψ\psi has zero mean,

DψV(0,0)[ϕ]=RN−1∫SN−1ϕ=0,D_\psi V(0,0)[\phi]=R^{N-1}\int_{S^{N-1}}\phi=0,

so Da(0)=0Da(0)=0. Differentiating twice gives

D2a(0)[ϕ,η]=−N−1RσN−1∫SN−1ϕη,(23)D^2a(0)[\phi,\eta]=-\frac{N-1}{R\sigma_{N-1}}\int_{S^{N-1}}\phi\eta, \tag*{(23)}

which yields (22).

Smooth dependence of the state and flux

Lemma 4.2 (Smooth state and flux map). Under (9)–[15], after shrinking UU and the neighborhood of 00 in X3,αX^{3,\alpha} if necessary, the solution ux,ψu_{x,\psi} of

−Δux,ψ=fin Ωx,ψ,ux,ψ=0on ∂Ωx,ψ,-\Delta u_{x,\psi}=f \quad\text{in }\Omega_{x,\psi},\qquad u_{x,\psi}=0 \quad\text{on }\partial\Omega_{x,\psi},

depends C2C^2-smoothly on (x,ψ)(x,\psi) after pullback to BRB_R. The pulled-back boundary flux

g(x,ψ)(ω)=−∂Nux,ψ(x+[R+a(ψ)+ψ(ω)]ω)(24)g(x,\psi)(\omega)=-\partial_N u_{x,\psi}\left(x+[R+a(\psi)+\psi(\omega)]\omega\right) \tag*{(24)}

is a C2C^2 map into C1,α(SN−1)C^{1,\alpha}(S^{N-1}).

Proof. Choose a smooth radial cutoff χ\chi equal to 00 on [0,R−2δ][0,R-2\delta] and 11 near RR. Define Tx,ψ(y)=x+yT_{x,\psi}(y)=x+y on a fixed neighborhood of the origin and, writing y=rωy=r\omega outside that neighborhood,

Tx,ψ(rω)=x+[r+χ(r)(a(ψ)+ψ(ω))]ω.T_{x,\psi}(r\omega)=x+[r+\chi(r)(a(\psi)+\psi(\omega))]\omega.

Because the two formulas agree where χ=0\chi=0, this gives a globally C3,αC^{3,\alpha} map on BRB_R. For ψ\psi small it is a C3,αC^{3,\alpha} diffeomorphism from BRB_R onto Ωx,ψ\Omega_{x,\psi}.

Put

XD=C02,α(BR),Y=C0,α(BR),X_D=C^{2,\alpha}_0(B_R),\qquad Y=C^{0,\alpha}(B_R),

and define

F(x,ψ,v)=−div⁡(Aψ∇v)−Jψf(Tx,ψ),\mathcal{F}(x,\psi,v)=-\operatorname{div}(A_\psi\nabla v)-J_\psi f(T_{x,\psi}),

with

Aψ=JψDTx,ψ−1DTx,ψ−T,Jψ=det⁡DTx,ψ.A_\psi=J_\psi DT_{x,\psi}^{-1}DT_{x,\psi}^{-T},\qquad J_\psi=\det DT_{x,\psi}.

Then

F:U×X3,α×XD⟶Y\mathcal{F}:U\times X^{3,\alpha}\times X_D\longrightarrow Y

is C2C^2: determinant and matrix inversion are smooth on the uniformly invertible coefficient set, the composition (x,ψ)↦f∘Tx,ψ(x,\psi)\mapsto f\circ T_{x,\psi} is C2C^2 because f∈C4,αf\in C^{4,\alpha}, and multiplication/divergence have the required Hölder mapping properties. Moreover,

DvF(x,ψ,v)z=−div⁡(Aψ∇z)D_v\mathcal{F}(x,\psi,v)z=-\operatorname{div}(A_\psi\nabla z)

is a uniformly elliptic Dirichlet isomorphism XD→YX_D\to Y by Schauder theory. The Banach-space implicit-function theorem therefore yields a C2C^2 map (x,ψ)↦vx,ψ∈C02,α(BR)(x,\psi)\mapsto v_{x,\psi}\in C^{2,\alpha}_0(B_R).

Finally,

∇ux,ψ=DTx,ψ−T∇vx,ψ,\nabla u_{x,\psi}=DT_{x,\psi}^{-T}\nabla v_{x,\psi},

and the outer normal of Σx,ψ\Sigma_{x,\psi} depends C2C^2-smoothly on ψ\psi. The trace map C2,α(BR)→C1,α(SN−1)C^{2,\alpha}(B_R)\to C^{1,\alpha}(S^{N-1}) is bounded. Hence the pulled-back flux (24) is C2C^2 into C1,α(SN−1)C^{1,\alpha}(S^{N-1}).

Adapted Lyapunov–Schmidt construction

Let H(ψ)H(\psi) denote the mean curvature of Σx,ψ\Sigma_{x,\psi} pulled back to SN−1\mathbb{S}^{N-1}. It is independent of xx by translation invariance. Define

M(x,ψ,ε)=H(ψ)−ε2g(x,ψ)2.(25)\mathcal{M}(x,\psi,\varepsilon)=H(\psi)-\frac{\varepsilon}{2}g(x,\psi)^2. \tag*{(25)}

Equation (13) holds for some scalar λε\lambda_\varepsilon exactly when the nonconstant part of M\mathcal{M} vanishes. Let

T=span⁡{ω1,…,ωN}T=\operatorname{span}\{\omega_1,\ldots,\omega_N\}

be the translation space, let PTP_T be the L2(SN−1)L^2(\mathbb{S}^{N-1}) projection onto TT, and let PXP_X be the projection onto the orthogonal complement of constants and TT.

Lemma 5.1 (Smooth mean-curvature map). For ψ\psi in a sufficiently small neighborhood of 00 in X3,αX^{3,\alpha}, the pulled-back mean-curvature map

ψ⟼H(ψ)\psi\longmapsto H(\psi)

is C∞C^\infty from X3,αX^{3,\alpha} to C1,α(SN−1)C^{1,\alpha}(\mathbb{S}^{N-1}). Its differential at the sphere is

DH(0)[ϕ]=JRϕ,JR=−Δ∂BR−N−1R2.DH(0)[\phi]=J_R\phi,\qquad J_R=-\Delta_{\partial B_R}-\frac{N-1}{R^2}.

Proof. The metric, second fundamental form, inverse metric and unit normal of a radial graph depend smoothly on ψ\psi, ∇ψ\nabla\psi and ∇2ψ\nabla^2\psi as long as R+a(ψ)+ψR+a(\psi)+\psi stays positive. Since a:X3,α→Ra:X^{3,\alpha}\to\mathbb{R} is smooth, the resulting mean curvature is a smooth quasilinear second-order operator from C3,αC^{3,\alpha} to C1,αC^{1,\alpha}. The standard first-variation formula for mean curvature at a sphere gives the displayed Jacobi operator with the convention H=div⁡NH=\operatorname{div}N.

Inversion of the higher spherical modes

At ψ=0\psi=0,

H(0)=N−1R.H(0)=\frac{N-1}{R}.

The linearized mean-curvature operator is

JR=−Δ∂BR−N−1R2=−1R2(ΔSN−1+N−1).(26)J_R=-\Delta_{\partial B_R}-\frac{N-1}{R^2} =-\frac{1}{R^2}\left(\Delta_{\mathbb{S}^{N-1}}+N-1\right). \tag*{(26)}

On degree-ℓ\ell spherical harmonics,

JRYℓ=ℓ(ℓ+N−2)−(N−1)R2Yℓ.(27)J_RY_\ell=\frac{\ell(\ell+N-2)-(N-1)}{R^2}Y_\ell. \tag*{(27)}

Thus constants form the volume mode, degree one is the translation kernel, and JRJ_R is positive and invertible on all degrees ℓ≥2\ell\ge2.

Proposition 5.2 (Higher-mode equation). There is a neighborhood of (x0,0)(x_0,0) and a unique C2C^2 map

Ψ:U×(−ε0,ε0)⟶X3,α\Psi:U\times(-\varepsilon_0,\varepsilon_0)\longrightarrow X^{3,\alpha}

such that

PXM(x,Ψ(x,ε),ε)=0,Ψ(x,0)=0.(28)P_X\mathcal{M}(x,\Psi(x,\varepsilon),\varepsilon)=0,\qquad\Psi(x,0)=0. \tag*{(28)}

Moreover

JR∂εΨ(x,0)=12PX(gx2),(29)J_R\partial_\varepsilon\Psi(x,0)=\frac{1}{2}P_X(g_x^2), \tag*{(29)}

where gx=g(x,0)g_x=g(x,0).

Proof. The derivative of PXMP_X\mathcal{M} with respect to ψ\psi at (x,0,0)(x,0,0) is JRJ_R restricted to X3,αX^{3,\alpha}, which is an isomorphism onto X1,αX^{1,\alpha} by (27). The implicit-function theorem gives (28). Differentiating in ε\varepsilon at 00 gives (29).

Desingularized translation equation

Define

G(x,ε)=PTM(x,Ψ(x,ε),ε).(30)G(x,\varepsilon)=P_T\mathcal{M}(x,\Psi(x,\varepsilon),\varepsilon). \tag*{(30)}

Since G(x,0)=0G(x,0)=0 for every xx, the direct xx-linearization is singular. Define for ε≠0\varepsilon\ne0

G~(x,ε)=G(x,ε)ε.(31)\widetilde{G}(x,\varepsilon)=\frac{G(x,\varepsilon)}{\varepsilon}. \tag*{(31)}

Lemma 5.3 (Desingularization). The map G~\widetilde{G} extends to a C1C^1 map at ε=0\varepsilon=0, with

G~(x,0)=−12PT(gx2).(32)\widetilde{G}(x,0)=-\frac{1}{2}P_T(g_x^2). \tag*{(32)}

More precisely,

G~(x,0)(ω)=NσN−1RN−1∇h(x)⋅ω.(33)\widetilde{G}(x,0)(\omega)=\frac{N}{\sigma_{N-1}R^{N-1}}\nabla h(x)\cdot\omega. \tag*{(33)}

Proof. Because GG is C2C^2 and G(x,0)=0G(x,0)=0,

G~(x,ε)=∫01∂εG(x,tε) dt,\widetilde{G}(x,\varepsilon)=\int_0^1\partial_\varepsilon G(x,t\varepsilon)\,dt,

which extends C1C^1 to ε=0\varepsilon=0. At ε=0\varepsilon=0,

∂εG=PT(JR∂εΨ−12gx2).\partial_\varepsilon G=P_T\left(J_R\partial_\varepsilon\Psi-\frac{1}{2}g_x^2\right).

The first term belongs to XX and has zero translation projection, which gives (31). Since

PTϕ=NσN−1(∫SN−1ϕ(ω)ω dω)⋅ω,P_T\phi=\frac{N}{\sigma_{N-1}}\left(\int_{S^{N-1}}\phi(\omega)\omega\,d\omega\right)\cdot\omega,

Lemma 3.1 yields (32).

Theorem 5.4 (Preparatory large-perimeter branch near a selected ball). Assume Assumption 3.2. Then there exists ε0>0\varepsilon_0>0 and a unique C1C^1 curve

ε↦x(ε),x(0)=x0,\varepsilon\mapsto x(\varepsilon),\qquad x(0)=x_0,

such that, with

ψ(ε)=Ψ(x(ε),ε),\psi(\varepsilon)=\Psi(x(\varepsilon),\varepsilon),

the exact-volume domain Ωx(ε),ψ(ε)\Omega_{x(\varepsilon),\psi(\varepsilon)} is a constrained critical domain of (1) for every sufficiently small positive ε=μ−1\varepsilon=\mu^{-1}. There exist ξ1∈RN\xi_1\in\mathbb{R}^N and ψ1∈X3,α\psi_1\in X^{3,\alpha} such that

x(ε)=x0+εξ1+o(ε),ψ(ε)=εψ1+oC3,α(ε),(34)\begin{aligned} x(\varepsilon)&=x_0+\varepsilon\xi_1+o(\varepsilon),\\ \psi(\varepsilon)&=\varepsilon\psi_1+o_{C^{3,\alpha}}(\varepsilon), \tag*{(34)} \end{aligned}

and

JRψ1=12PX(g02),g0=g(x0,0).(35)J_R\psi_1=\frac{1}{2}P_X(g_0^2),\qquad g_0=g(x_0,0). \tag*{(35)}

The center coefficient is determined by the differentiated reduced equation:

ξ1=−[DxG~(x0,0)]−1∂εG~(x0,0).(36)\xi_1=-[D_x\widetilde{G}(x_0,0)]^{-1}\partial_\varepsilon\widetilde{G}(x_0,0). \tag*{(36)}

The scaled Lagrange multiplier satisfies

λε=−N−1R+ε2∫SN−1g02 dω+o(ε).(37)\lambda_\varepsilon=-\frac{N-1}{R}+\frac{\varepsilon}{2}\int_{S^{N-1}}g_0^2\,d\omega+o(\varepsilon). \tag*{(37)}

Equivalently,

Λμ=−μN−1R+12∫SN−1g02 dω+o(1).(38)\Lambda_\mu=-\mu\frac{N-1}{R}+\frac{1}{2}\int_{S^{N-1}}g_0^2\,d\omega+o(1). \tag*{(38)}

Proof. By (33), the derivative DxG~(x0,0)D_x\widetilde{G}(x_0,0) is invertible exactly when D2h(x0)D^2h(x_0) is invertible. The finite-dimensional implicit-function theorem therefore gives x(ε)x(\varepsilon) solving

G~(x(ε),ε)=0.\widetilde{G}(x(\varepsilon),\varepsilon)=0.

Together with Proposition 5.2, all nonconstant components of M\mathcal{M} vanish. The remaining constant component is absorbed by a scalar λε\lambda_\varepsilon, so (13) holds.

Since Ψ(x,0)≡0\Psi(x,0)\equiv0, one has DxΨ(x,0)=0D_x\Psi(x,0)=0, hence

ψ1=∂εΨ(x0,0),\psi_1=\partial_\varepsilon\Psi(x_0,0),

and (35) follows from (29). Differentiating G~(x(ε),ε)=0\widetilde{G}(x(\varepsilon),\varepsilon)=0 at ε=0\varepsilon=0 gives (36). The expansion of λε\lambda_\varepsilon is obtained by averaging (13); the average of JRψ1J_R\psi_1 vanishes because ψ1\psi_1 has zero mean. Multiplying by μ=ε−1\mu=\varepsilon^{-1} yields (38).

Remark 5.5 (Correct scale). The generic shape scale is μ−1\mu^{-1}. A μ−1/2\mu^{-1/2} scale would require an additional degeneracy causing the linear restoring term to vanish in the relevant mode; it is not the generic large-perimeter regime treated here.

Openness of OC\mathcal{O}_C

Theorem 6.1 (C1C^1-openness of OC\mathcal{O}_C). Let d0∈C2,α(M)d_0\in C^{2,\alpha}(M) be positive and assume Ω[d0]∈OC\Omega[d_0]\in\mathcal{O}_C.

Then there exists ε∗>0\varepsilon_*>0 such that every positive d∈C2,α(M)d\in C^{2,\alpha}(M) satisfying

∥d−d0∥C1(M)<ε∗(39)\lVert d-d_0\rVert_{C^1(M)}<\varepsilon_* \tag*{(39)}

obeys

Ω[d]∈OC.\Omega[d]\in\mathcal{O}_C.

Moreover, for dd in this neighborhood the first-return time τd\tau_d and return map FdF_d satisfy

∥τd−τ0∥C0+∥Fd−F0∥C0≤C∥d−d0∥C1,(40)\lVert\tau_d-\tau_0\rVert_{C^0}+\lVert F_d-F_0\rVert_{C^0}\leq C\lVert d-d_0\rVert_{C^1}, \tag*{(40)}

and the first return remains uniformly transverse.

Proof. The global graph property is automatic for positive dd, since strict convexity gives unique exterior normal coordinates. It remains to prove persistence of the inward-normal penetration.

From (7)–(8), the maps d↦wdd\mapsto w_d and d↦ndd\mapsto n_d are locally Lipschitz from C1C^1 to C0C^0:

∥nd−n0∥C0≤C∥d−d0∥C1.\lVert n_d-n_0\rVert_{C^0}\leq C\lVert d-d_0\rVert_{C^1}.

Let

F0(c)=Ψ0(c)+τ0(c)n0(c)F_0(c)=\Psi_0(c)+\tau_0(c)n_0(c)

be the first return. Since Ω[d0]∈OC\Omega[d_0]\in\mathcal{O}_C, the same ray continues into int⁡C\operatorname{int} C. Convexity implies

n0(c)⋅ν(F0(c))<0.n_0(c)\cdot\nu(F_0(c))<0.

To justify the compactness step, let ρC\rho_C be the signed distance to MM and write

ρC(Ψ0(c)+τ0(c)n0(c))=0.\rho_C\bigl(\Psi_0(c)+\tau_0(c)n_0(c)\bigr)=0.

At the first entrance point,

∂sρC(Ψ0(c)+sn0(c))∣s=τ0(c)=ν(F0(c))⋅n0(c)<0.\left.\partial_s\rho_C\bigl(\Psi_0(c)+sn_0(c)\bigr)\right|_{s=\tau_0(c)} =\nu(F_0(c))\cdot n_0(c)<0.

The parameter-dependent implicit-function theorem therefore gives τ0∈C1,α(M)\tau_0 \in C^{1,\alpha}(M) and hence continuity of F0F_0. Compactness now yields a uniform constant η0>0\eta_0 > 0 such that

−n0(c)⋅ν(F0(c))≥η0.-n_0(c) \cdot\nu(F_0(c)) \ge\eta_0.

Let ρC\rho_C be the signed distance to MM, negative in CC. Uniform transversality gives a δ∗>0\delta_* > 0 and m0>0m_0 > 0 such that

ρC(F0(c)+δ∗n0(c))≤−m0\rho_C(F_0(c) + \delta_* n_0(c)) \le-m_0

for all cc. For nearby dd, define

yd(c)=Ψd(c)+(τ0(c)+δ∗)nd(c).y_d(c) = \Psi_d(c) + (\tau_0(c) + \delta_*)n_d(c).

Then

∥yd−y0∥C0≤C∥d−d0∥C1.\|y_d-y_0\|_{C^0} \le C\|d-d_0\|_{C^1}.

For sufficiently small (39), yd(c)∈int⁡Cy_d(c) \in\operatorname{int} C for every cc, so every perturbed inward normal ray reaches the interior. Hence Ω[d]∈OC\Omega[d] \in\mathcal{O}_C.

The quantitative estimates and persistence of transversality follow by applying the implicit-function theorem uniformly to

Hd(c,s)=ρC(Ψd(c)+snd(c))H_d(c,s) = \rho_C(\Psi_d(c) + sn_d(c))

near (c,τ0(c))(c,\tau_0(c)). This produces, for each nearby dd, a unique transverse zero τ^d(c)\widehat{\tau}_d(c) in a fixed neighborhood of τ0(c)\tau_0(c). It remains to identify that zero with the first entrance time. Since CC is convex and the ray contains the later interior point yd(c)y_d(c), the set of ray parameters for which it lies in CC is a nonempty compact interval. The transverse zero τ^d(c)\widehat{\tau}_d(c) has points immediately before it outside CC and points immediately after it inside CC. Hence it is the entering endpoint of that interval, so no earlier intersection can occur and τ^d=τd\widehat{\tau}_d=\tau_d. The same uniform implicit-function estimates then give (40).

Coordinate transfer from the selected sphere to the core

Assume now

x0∈int⁡C,C∈BR(x0).(41)x_0 \in\operatorname{int} C,\qquad C \in B_R(x_0). \tag*{(41)}

Every core-normal ray meets Σ0=∂BR(x0)\Sigma_0=\partial B_R(x_0) exactly once. Let d0d_0 be defined by

Ψ0(c)=c+d0(c)ν(c)∈∂BR(x0).(42)\Psi_0(c)=c+d_0(c)\nu(c)\in\partial B_R(x_0). \tag*{(42)}

Set

ω0(c)=Ψ0(c)−x0R,θ0(c)=ω0(c)⋅ν(c).(43)\omega_0(c)=\frac{\Psi_0(c)-x_0}{R},\qquad\theta_0(c)=\omega_0(c)\cdot\nu(c). \tag*{(43)}

Then θ0>0\theta_0>0 uniformly.

Proposition 7.1 (The selected sphere belongs to OC\mathcal{O}_C). Under (41), the sphere BR(x0)B_R(x_0) is an element of OC\mathcal{O}_C. More precisely, each outward core-normal ray meets ∂BR(x0)\partial B_R(x_0) exactly once and the inward sphere normal reaches x0∈int⁡Cx_0\in\operatorname{int} C.

Proof. Fix c∈Mc\in M and consider

qc(r)=∣c+rν(c)−x0∣2−R2.q_c(r)=\lvert c+r\nu(c)-x_0\rvert^2-R^2.

Since c∈C∈BR(x0)c\in C\in B_R(x_0), one has qc(0)<0q_c(0)<0, while qc(r)→+∞q_c(r)\to+\infty as r→∞r\to\infty. The product of the two roots of this quadratic equals qc(0)<0q_c(0)<0, so exactly one root is positive. This gives the unique outer intersection and hence the thickness d0(c)d_0(c). At the intersection point the inward sphere normal points along the segment toward x0x_0, which lies in int⁡C\operatorname{int} C; therefore the inward normal ray enters the core.

Moreover, taking the scalar product of

c+d0(c)ν(c)−x0=Rω0(c)c + d_0(c)\nu(c) - x_0 = R\omega_0(c)

with ν(c)\nu(c) gives

Rθ0(c)=(c−x0)⋅ν(c)+d0(c)>0.R\theta_0(c) = (c-x_0)\cdot\nu(c) + d_0(c) > 0.

Compactness yields a positive uniform lower bound. In fact, by the exact normal-graph formula,

θ0=N0⋅ν=χ0−1,χ0=1+∣w0∣2.(44)\theta_0 = N_0\cdot\nu= \chi_0^{-1}, \qquad\chi_0 = \sqrt{1+|w_0|^2}. \tag*{(44)}

Theorem 7.2 (Coordinate transfer). For (x,ψ)(x,\psi) sufficiently close to (x0,0)(x_0,0) in RN×X3,α\mathbb{R}^N \times X^{3,\alpha}, the hypersurface Σx,ψ\Sigma_{x,\psi} is a unique normal graph over MM:

Σx,ψ={c+d(x,ψ)(c)ν(c):c∈M},\Sigma_{x,\psi} = \{c+d(x,\psi)(c)\nu(c):c\in M\},

with d(x,ψ)∈C2,α(M)d(x,\psi)\in C^{2,\alpha}(M). The map

(x,ψ)↦d(x,ψ)(x,\psi)\mapsto d(x,\psi)

is C1C^1, and

∥d(x,ψ)−d0∥C2,α≤C(∣x−x0∣+∥ψ∥C3,α).(45)\|d(x,\psi)-d_0\|_{C^{2,\alpha}} \leq C(|x-x_0|+\|\psi\|_{C^{3,\alpha}}). \tag*{(45)}

Its derivative at (x0,0)(x_0,0) is

Dd(x0,0)[ξ,ϕ](c)=ω0(c)⋅ξ+ϕ(ω0(c))θ0(c).(46)Dd(x_0,0)[\xi,\phi](c)=\frac{\omega_0(c)\cdot\xi+\phi(\omega_0(c))}{\theta_0(c)}. \tag*{(46)}

Proof. For yy near Σ0\Sigma_0 define

rx(y)=∣y−x∣,ωx(y)=y−x∣y−x∣r_x(y)=|y-x|,\qquad\omega_x(y)=\frac{y-x}{|y-x|}

and the defining function

ρx,ψ(y)=rx(y)−R−a(ψ)−ψ(ωx(y)).\rho_{x,\psi}(y)=r_x(y)-R-a(\psi)-\psi(\omega_x(y)).

Then ρx,ψ=0\rho_{x,\psi}=0 precisely on Σx,ψ\Sigma_{x,\psi}. Define

G(x,ψ,d)(c)=ρx,ψ(c+d(c)ν(c)).G(x,\psi,d)(c)=\rho_{x,\psi}(c+d(c)\nu(c)).

With one derivative of regularity in reserve, GG is C1C^1 from

RN×X3,α×C2,α(M) to C2,α(M).\mathbb{R}^N\times X^{3,\alpha}\times C^{2,\alpha}(M)\ \text{to}\ C^{2,\alpha}(M).

Indeed, after restricting to a fixed neighborhood of Σ0\Sigma_0, the maps y↦rx(y)y\mapsto r_x(y) and y↦ωx(y)y\mapsto\omega_x(y) are smooth uniformly for xx near x0x_0. The Hölder composition map (ψ,z)↦ψ∘z(\psi,z)\mapsto\psi\circ z is C1C^1 from C3,α×C2,αC^{3,\alpha}\times C^{2,\alpha} to C2,αC^{2,\alpha} on this neighborhood (this is a standard consequence of the Hölder algebra property and the chain rule; see for instance [7 Lemma A.1] for the precise statement adapted to the present setting), while d↦c+dνd\mapsto c+d\nu is affine into C2,αC^{2,\alpha}. This is the reason for retaining one derivative beyond the target regularity. At the reference state, G(x0,0,d0)=0G(x_0,0,d_0)=0, and

DdG(x0,0,d0)[η]=θ0η.D_dG(x_0,0,d_0)[\eta]=\theta_0\eta.

Since θ0\theta_0 is bounded away from zero, multiplication by θ0\theta_0 is an isomorphism on C2,α(M)C^{2,\alpha}(M). The implicit-function theorem gives d=d(x,ψ)d=d(x,\psi) and (44). Since d0>0d_0>0 on the compact manifold MM, shrinking the neighborhood if necessary also gives d(x,ψ)>0d(x,\psi)>0.

For a strictly convex core, the exterior normal-coordinate map (c,r)↦c+rν(c)(c,r)\mapsto c+r\nu(c) is one-to-one for r>0r>0. Hence the positive graph c↦c+d(c)ν(c)c\mapsto c+d(c)\nu(c) is an embedded hypersurface. Its image is open and closed in the connected hypersurface Σx,ψ\Sigma_{x,\psi}, hence equals it. Uniqueness follows from the same global exterior normal coordinates.

Finally, differentiate G(x,ψ,d(x,ψ))=0\mathcal{G}(x,\psi,d(x,\psi))=0. At the reference state,

Dxρ[ξ]=−ω0⋅ξ,Dψρ[ϕ]=−ϕ(ω0)D_x\rho[\xi]=-\omega_0\cdot\xi,\qquad D_\psi\rho[\phi]=-\phi(\omega_0)

because Da(0)=0D_a(0)=0. Therefore

0=−ω0⋅ξ−ϕ(ω0)+θ0η,(47)0=-\omega_0\cdot\xi-\phi(\omega_0)+\theta_0\eta, \tag*{(47)}

which is (45).

Corollary 7.3 (Core-normal expansion of the critical branch). Let x(ε)x(\varepsilon) and ψ(ε)\psi(\varepsilon) be the branch of Theorem 5.4. Then

dε=d0+εη1+oC2,α(ε),(48)d_\varepsilon=d_0+\varepsilon\eta_1+o_{C^{2,\alpha}}(\varepsilon), \tag*{(48)}

where

η1(c)=ω0(c)⋅ξ1+ψ1(ω0(c))θ0(c)=χ0(c)[ω0(c)⋅ξ1+ψ1(ω0(c))].(49)\eta_1(c)=\frac{\omega_0(c)\cdot\xi_1+\psi_1(\omega_0(c))}{\theta_0(c)} =\chi_0(c)[\omega_0(c)\cdot\xi_1+\psi_1(\omega_0(c))]. \tag*{(49)}

Consequently

dμ=d0+μ−1η1+oC2,α(μ−1).(50)d_\mu=d_0+\mu^{-1}\eta_1+o_{C^{2,\alpha}}(\mu^{-1}). \tag*{(50)}

Remark 7.4. The translation coefficient ξ1\xi_1 contributes at the same order as the genuine higher-mode deformation ψ1\psi_1. The shorthand η1=χ0ψ1\eta_1=\chi_0\psi_1 is correct only in the special case ξ1=0\xi_1=0.

Fréchet differentiability of the return map

Let d0∈C2,α(M)d_0\in C^{2,\alpha}(M) with Ω[d0]∈OC\Omega[d_0]\in\mathcal{O}_C. By Theorem 6.1, all nearby thicknesses remain admissible. Write

Fd(c)=Ψd(c)+τd(c)nd(c),(51)F_d(c)=\Psi_d(c)+\tau_d(c)n_d(c), \tag*{(51)}

where τd(c)\tau_d(c) is the first entrance time into CC.

Differentials of the graph slope and normal

For a perturbation dε=d0+εηd_\varepsilon=d_0+\varepsilon\eta,

δw[η]=L0−1(∇Mη+ηSw0).(52)\delta w[\eta]=L_0^{-1}(\nabla_M\eta+\eta S w_0). \tag*{(52)}

Differentiating (8) gives

δn[η]=1χ0(I−n0⊗n0)L0−1(∇Mη+ηSw0).(53)\delta n[\eta]=\frac{1}{\chi_0}(I-n_0\otimes n_0)L_0^{-1}(\nabla_M\eta+\eta S w_0). \tag*{(53)}

Lemma 8.1 (Banach-space first-hit). Let d0∈C2,α(M)d_0\in C^{2,\alpha}(M) with Ω[d0]∈OC\Omega[d_0]\in\mathcal{O}_C. There are neighborhoods U⊂C2,α(M)U\subset C^{2,\alpha}(M) of d0d_0 and V⊂C1,α(M)V\subset C^{1,\alpha}(M) of τ0\tau_0 such that

H:U×V⟶C1,α(M),H(d,τ)(c)=ρC(Ψd(c)+τ(c)nd(c)),\mathcal{H}:U\times V\longrightarrow C^{1,\alpha}(M),\qquad \mathcal{H}(d,\tau)(c)=\rho_C(\Psi_d(c)+\tau(c)n_d(c)),

is C1C^1. Moreover,

DτH(d0,τ0)[σ]=[ν(F0)⋅n0]σD_\tau\mathcal{H}(d_0,\tau_0)[\sigma]=[\nu(F_0)\cdot n_0]\sigma

is an isomorphism of C1,α(M)C^{1,\alpha}(M).

Proof. Choose a tubular neighborhood T\mathcal{T} of MM on which the signed distance ρC\rho_C is smooth with the regularity supplied by M∈C4,αM \in C^{4,\alpha}. Shrink U\mathcal{U} and V\mathcal{V} so that

Ψd+τnd∈T\Psi_d+\tau n_d\in\mathcal{T}

pointwise for every (d,τ)∈U×V(d,\tau)\in\mathcal{U}\times\mathcal{V}. The map d↦Ψd=c+dνd\mapsto\Psi_d=c+d\nu is affine from C2,αC^{2,\alpha} to C2,αC^{2,\alpha}. Since

wd=(I−dS)−1∇Md,nd=−ν+wd1+∣wd∣2,w_d=(I-dS)^{-1}\nabla_Md,\qquad n_d=\frac{-\nu+w_d}{\sqrt{1+\lvert w_d\rvert^2}},

and I−dS≥II-dS\ge I (because S≤0S\le0 and d≥0d\ge0), the inverse (I−dS)−1(I-dS)^{-1} is uniformly bounded by 11 in operator norm on U\mathcal{U}; the Hölder algebra property and smoothness of matrix inversion then imply that d↦wdd\mapsto w_d and d↦ndd\mapsto n_d are C1C^1 from C2,α(M)C^{2,\alpha}(M) to C1,α(M)C^{1,\alpha}(M), with derivative continuous in the operator norm L(C2,α,C1,α)\mathcal{L}(C^{2,\alpha},C^{1,\alpha}). Consequently

(d,τ)⟼Ψd+τnd(d,\tau)\longmapsto\Psi_d+\tau n_d

is C1C^1 into C1,α(M;RN)C^{1,\alpha}(M;\mathbb{R}^N). Composition with the fixed smooth function ρC\rho_C is C1C^1 on maps with image in T\mathcal{T}, which proves the asserted mapping property of H\mathcal{H}.

At the reference state, the chain rule gives

DτH(d0,τ0)[σ]=∇ρC(F0)⋅n0σ=[ν(F0)⋅n0]σ.D_\tau\mathcal{H}(d_0,\tau_0)[\sigma]=\nabla\rho_C(F_0)\cdot n_0\sigma=[\nu(F_0)\cdot n_0]\sigma.

By Theorem 6.1, first-return transversality is uniform: −ν(F0)⋅n0≥η0>0-\nu(F_0)\cdot n_0\ge\eta_0>0. The multiplier belongs to C1,α(M)C^{1,\alpha}(M), as does its reciprocal. Multiplication by it is therefore a bounded isomorphism of C1,α(M)C^{1,\alpha}(M).

Theorem 8.2 (Principal return-response theorem). There is a neighborhood U\mathcal{U} of d0d_0 in C2,α(M)C^{2,\alpha}(M) such that

R:U→C1,α(M;RN),R(d)=Fd,\mathcal{R}:\mathcal{U}\to C^{1,\alpha}(M;\mathbb{R}^N),\qquad\mathcal{R}(d)=F_d,

is C1C^1 in the Fréchet sense, with derivative continuous in the operator norm L(C2,α(M),C1,α(M;RN))\mathcal{L}(C^{2,\alpha}(M),C^{1,\alpha}(M;\mathbb{R}^N)). For η∈C2,α(M)\eta\in C^{2,\alpha}(M) define

Vη=ην+τ0δn[η].(54)V_\eta=\eta\nu+\tau_0\delta n[\eta]. \tag*{(54)}

Then

DR[d0]η=Vη−ν(F0)⋅Vην(F0)⋅n0n0.(55)D\mathcal{R}[d_0]\eta=V_\eta-\frac{\nu(F_0)\cdot V_\eta}{\nu(F_0)\cdot n_0}n_0. \tag*{(55)}

Equivalently,

DR[d0]η=Π0ret[ην+τ0χ0(I−n0⊗n0)L0−1(∇Mη+ηSw0)].(56)D\mathcal{R}[d_0]\eta=\Pi_0^{\mathrm{ret}}\left[\eta\nu+\frac{\tau_0}{\chi_0}(I-n_0\otimes n_0)L_0^{-1}\left(\nabla_M\eta+\eta Sw_0\right)\right]. \tag*{(56)}

where

Π0retv=v−ν(F0)⋅vν(F0)⋅n0n0.(57)\Pi_0^{\mathrm{ret}}v=v-\frac{\nu(F_0)\cdot v}{\nu(F_0)\cdot n_0}n_0. \tag*{(57)}

In particular,

DR[d0]η(c)∈TF0(c)M.(58)D\mathcal{R}[d_0]\eta(c)\in T_{F_0(c)}M. \tag*{(58)}

Proof. By Lemma 8.1, the implicit-function theorem gives a unique C1C^1 map

d⟼τ^(d)∈C1,α(M)d\longmapsto\widehat{\tau}(d)\in C^{1,\alpha}(M)

near d0d_0 such that H(d,τ^(d))=0\mathcal{H}(d,\widehat{\tau}(d))=0 and τ^(d0)=τ0\widehat{\tau}(d_0)=\tau_0. We must still verify that this local root is the geometric first entrance time. Fix d∈Ud\in\mathcal{U} and c∈Mc\in M, and set hd,c(s)=ρC(Ψd(c)+snd(c))h_{d,c}(s)=\rho_C(\Psi_d(c)+sn_d(c)). By the first-hit claim, there is a unique transverse zero s^=τ^(d)(c)\widehat{s}=\widehat{\tau}(d)(c) in a fixed neighborhood of τ0(c)\tau_0(c), with

∂shd,c(s^)=ν(Fd(c))⋅nd(c)≤−η0/2<0\partial_s h_{d,c}(\widehat{s})=\nu(F_d(c))\cdot n_d(c)\le-\eta_0/2<0

by Theorem 6.1 (uniform transversality). Consequently hd,c(s^−δ)>0h_{d,c}(\hat{s}-\delta)>0 and hd,c(s^+δ)<0h_{d,c}(\hat{s}+\delta)<0 for some δ>0\delta>0; by compactness of MM and continuity of (d,c)↦hd,c(d,c)\mapsto h_{d,c} in the C1C^1 topology, δ\delta can be chosen uniform in c∈Mc\in M. Thus, immediately before s^\hat{s} the ray is outside CC, and immediately after s^\hat{s} it is inside CC.

Since CC is convex and every inward normal ray reaches int⁡C\operatorname{int} C (Theorem 6.1), the set of parameters ss for which Ψd(c)+snd(c)∈C\Psi_d(c)+sn_d(c)\in C is a nonempty compact interval [s−,s+][s_-,s_+]. Since ∂shd,c(s^)≠0\partial_s h_{d,c}(\hat{s})\ne0, the point s^\hat{s} is a transverse crossing, not an endpoint of the interval; and because hd,ch_{d,c} changes sign from positive to negative at s^\hat{s}, we must have s^=s−\hat{s}=s_-. Hence τ^(d)(c)=τd(c)\hat{\tau}(d)(c)=\tau_d(c) for every c∈Mc\in M, so d↦τdd\mapsto\tau_d is C1C^1 as a map from U⊂C2,α(M)U\subset C^{2,\alpha}(M) to C1,α(M)C^{1,\alpha}(M).

Since d↦Ψdd\mapsto\Psi_d and d↦ndd\mapsto n_d have the corresponding C1C^1 regularity, (49) also shows directly that R:U→C1,α(M;RN)\mathcal{R}:U\to C^{1,\alpha}(M;\mathbb{R}^N) is C1C^1.

Differentiate the first-hit identity. Since δΨ=ην\delta\Psi=\eta\nu,

0=ν(F0)⋅[ην+(Dτ[d0]η)n0+τ0δn[η]].0=\nu(F_0)\cdot[\eta\nu+(D\tau[d_0]\eta)n_0+\tau_0\delta n[\eta]].

Thus

Dτ[d0]η=−ν(F0)⋅Vην(F0)⋅n0.D\tau[d_0]\eta=-\frac{\nu(F_0)\cdot V_\eta}{\nu(F_0)\cdot n_0}.

Differentiating (49) gives (53). The projected form (54) follows from (51). Equation (56) follows by taking the scalar product with ν(F0)\nu(F_0).

Remark 8.3 (Link with the inverse-paper compatibility system). The nonlinear transport system of [6] can be written in the present notation as T(d,F)=0\mathcal{T}(d,F)=0, where T\mathcal{T} encodes the geometric compatibility between the thickness dd and its return map FF. The differential DFTD_F\mathcal{T} along (d0,F0)(d_0,F_0) is an isomorphism by uniform transversality, hence

DR[d0]=−(DFT)−1DdT.D\mathcal{R}[d_0]=-(D_F\mathcal{T})^{-1}D_d\mathcal{T}.

In other words, DR[d0]D\mathcal{R}[d_0] is exactly the linearization of the inverse-paper transport system along the solution (d0,F0)(d_0,F_0), solved for the variation of FF. This identification will be used in future work to address the range characterization stated in Conjecture 12.1.

Corollary 8.4 (First return-map response of the critical branch). Under the assumptions of Theorems 5.4, 6.1, 7.2, and 8.2,

Fμ=F0+μ−1DR[d0]η1+oC1,α(μ−1).(59)F_\mu=F_0+\mu^{-1}D\mathcal{R}[d_0]\eta_1+o_{C^{1,\alpha}}(\mu^{-1}). \tag*{(59)}

Proof. Combine (48) with the Fréchet expansion of R\mathcal{R} at d0d_0.

Consistency with the PDE-generated deformation

The superlevel-set paper [7] obtained

dt=d0−tb0+oC1(t),Ft=F0+tA0+o(t).d_t=d_0-tb_0+o_{C^1}(t),\qquad F_t=F_0+tA_0+o(t).

For the direction η=−b0\eta=-b_0, (50) becomes

δw[−b0]=−L−1(∇Mb0+b0Sw0).\delta w[-b_0]=-L^{-1}(\nabla_M b_0+b_0S w_0).

Substitution into (51) and (54) gives exactly the return-map first-variation coefficient A0A_0 derived in [7]; the calculation is reproduced in Appendix § 14 to make the identification explicit. Thus

DR[d0](−b0)=A0.(60)D\mathcal{R}[d_0](-b_0)=A_0. \tag*{(60)}

This identity provides an independent sign and projection check for Theorem 8.2.

Combined main theorem

Theorem 9.1 (Main shape-to-dynamics theorem in OC\mathcal{O}_C). Let C⊂RNC \subset\mathbb{R}^N be compact and strictly convex with ∂C∈C4,α\partial C \in C^{4,\alpha}. Assume:

(H1)(H1) f∈Cc4,α(int⁡C)f \in C^{4,\alpha}_c(\operatorname{int} C), f≥0f \ge0, f≠0f \ne0;

(H2)(H2) RR is fixed by ∣BR∣=V0|B_R| = V_0;

(H3)(H3) there is x0∈int⁡Cx_0 \in\operatorname{int} C such that C⋐BR(x0)C \Subset B_R(x_0);

(H4)(H4) x0x_0 is a nondegenerate critical point of the translated-ball energy hh, i.e.

∇h(x0)=0,det⁡D2h(x0)≠0.\nabla h(x_0) = 0,\qquad\det D^2h(x_0) \ne0.

Then, for all sufficiently large μ\mu, there exists a volume-constrained critical domain Ωμ\Omega_\mu of (1) such that

Ωμ∈OC.\Omega_\mu\in\mathcal{O}_C.

Let d0d_0 be the core-normal thickness of BR(x0)B_R(x_0) and F0F_0 its return map. Then

dμ=d0+μ−1η1+oC2,α(μ−1),(61)d_\mu= d_0 + \mu^{-1}\eta_1 + o_{C^{2,\alpha}}(\mu^{-1}), \tag*{(61)}

where

η1(c)=ω0(c)⋅ξ1+ψ1(ω0(c))ω0(c)⋅ν(c),(62)\eta_1(c) = \frac{\omega_0(c)\cdot\xi_1 + \psi_1(\omega_0(c))}{\omega_0(c)\cdot\nu(c)}, \tag*{(62)}

ψ1\psi_1 is uniquely determined by

JRψ1=12PX(g02),(63)J_R\psi_1 = \frac{1}{2}P_X(g_0^2), \tag*{(63)}

and the first center correction is explicitly characterized by

ξ1=−[DxG~(x0,0)]−1∂εG~(x0,0).(64)\xi_1 = -[D_x\widetilde{G}(x_0,0)]^{-1}\partial_\varepsilon\widetilde{G}(x_0,0). \tag*{(64)}

Finally,

Fμ=F0+μ−1DR[d0]η1+oC1,α(μ−1),(65)F_\mu= F_0 + \mu^{-1}D\mathcal{R}[d_0]\eta_1 + o_{C^{1,\alpha}}(\mu^{-1}), \tag*{(65)}

with DR[d0]D\mathcal{R}[d_0] given explicitly by (56).

Proof. Theorem 5.4 gives the selected critical branch in spherical coordinates, parametrized by ε=μ−1∈(0,ε0)\varepsilon= \mu^{-1} \in(0,\varepsilon_0). Theorem 7.2 converts the branch to a core-normal thickness and yields (61)–(60). Since $d_\mu \to d_0$ in $C^1$, Theorem 6.1 implies $\Omega_\mu \in \mathcal{O}_C$ for large $\mu$. Theorem {xref}8.2 then gives {xref}(63)`. □

Remark 9.2 (On assumption (H3)). Assumption (H3) is structural: it ensures that the selected sphere belongs to OC\mathcal{O}_C (Proposition 7.1) and that θ0>0\theta_0 > 0 uniformly (equation (44)). It is not a consequence of (H4). Two regimes are relevant. If x0x_0 is the center of the minimal ball containing CC (or more generally a point for which C⋐BR(x0)C \Subset B_R(x_0) holds by construction), then (H3) is satisfied automatically and (H4) becomes a condition on ff through the reduced energy hh. A concrete non-spherical instance is obtained by taking CC to be an ellipsoid with center x0x_0 and ff radial about x0x_0 with sufficiently small non-radial perturbations; then x0x_0 remains a nondegenerate critical point of hh (see the perturbative example in § 11) and (H3)–(H4) hold simultaneously. If instead x0x_0 is off-center relative to CC, (H3) imposes a geometric relation between CC, ff and x0x_0 that is not generic. Theorem 9.1 should therefore be read as conditional on (H3)–(H4) jointly; the existence question for a non-spherical pair (C,f)(C,f) satisfying both is addressed perturbatively in § 11 and left open in full generality in § 12(ii).

PDE boundary-layer interpretation

The preceding theorem concerns optimization, not superlevel motion. Nevertheless, the first-order forcing can be expressed through a quantity already identified in [7].

Suppose the reference sphere Σ0\Sigma_0 is represented over the core by d0d_0, and let

χ0=1+∣w0∣2.\chi_0=\sqrt{1+\lvert w_0\rvert^2}.

For the small superlevel domains of the fixed reference PDE problem, [7] gives

dt=d0−tb0+oC1(t),b0(c)=1−∂N0u0(Ψ0(c))N0(Ψ0(c))⋅ν(c).d_t=d_0-tb_0+o_{C^1}(t),\qquad b_0(c)=\frac{1}{-\partial_{N_0}u_0(\Psi_0(c))N_0(\Psi_0(c))\cdot\nu(c)}.

Since N0⋅ν=χ0−1N_0\cdot\nu=\chi_0^{-1},

(g02)∘Ψ0=χ02b02.(66)(g_0^2)\circ\Psi_0=\frac{\chi_0^2}{b_0^2}. \tag*{(66)}

Under the stronger smoothness hypotheses of the present paper, the same boundary-flux formula gives b0∈C2,α(M)b_0\in C^{2,\alpha}(M), so −b0-b_0 is an admissible direction for Theorem 8.2. Thus the PDE level-set speed does not equal the optimization direction. Instead, it determines the optimization forcing after inversion and projection:

level speed b0⟼12g02=12(χ0b0)2⟼ψ1=12J−1PX(g02).\text{level speed }b_0\longmapsto\frac{1}{2}g_0^2=\frac{1}{2}\left(\frac{\chi_0}{b_0}\right)^2\longmapsto\psi_1=\frac{1}{2}\mathcal{J}^{-1}P_X(g_0^2).

The two deformations therefore encode different geometric responses to the same boundary flux.

Benchmarks and checks

Radial benchmark

Take

C=Br(x0),Ω0=BR(x0),0<r<R,C=B_r(x_0),\qquad\Omega_0=B_R(x_0),\qquad0<r<R,

and assume ff is radial about x0x_0. Then g0g_0 is constant on ∂BR(x0)\partial B_R(x_0), so

PX(g02)=0P_X(g_0^2)=0

and hence

ψ1=0.\psi_1=0.

The Euler–Lagrange equation is constant on the sphere for every μ\mu, with the multiplier absorbing both the curvature and flux terms. Thus the sphere is an exact critical shape. Moreover

d0≡R−rd_0\equiv R-r

and

F0=Id⁡∂C.F_0=\operatorname{Id}_{\partial C}.

Rotational symmetry makes the reduced translated-ball energy radial about x0x_0, and the uniqueness of the local reduced branch implies x(ε)≡x0x(\varepsilon)\equiv x_0; hence ξ1=0\xi_1=0. Together with ψ1=0\psi_1=0, formula (49) gives η1=0\eta_1=0. The first return-map response therefore vanishes, consistently with (59).

Perturbative nonradial example

We now exhibit a nonradial setting in which η1≠0\eta_1\ne0 to first order, confirming that Theorem 9.1 is not vacuous. Work in N=2N=2 with

C=B1(0),V0=πR2with R>1,C=B_1(0),\qquad V_0=\pi R^2\quad\text{with }R>1,

and choose a source of the form

f=f0+δf1,f1(x)=χ(x)Y2(θ),f=f_0+\delta f_1,\qquad f_1(x)=\chi(x)Y_2(\theta),

where f0≥0f_0\geq0 is radial about 00, Y2(θ)=cos⁡(2θ)Y_2(\theta)=\cos(2\theta) (a degree-2 spherical harmonic in N=2N=2), and χ∈Cc∞(B1/2(0))\chi\in C^\infty_c(B_{1/2}(0)) is a nonnegative radial cutoff equal to 11 on B1/4(0)B_{1/4}(0). Then f∈Cc4,α(int⁡C)f\in C^{4,\alpha}_c(\operatorname{int} C) and supp⁡f⋐B1/2(0)⋐int⁡C\operatorname{supp} f\Subset B_{1/2}(0)\Subset\operatorname{int} C, so (H1) holds for all sufficiently small δ\delta.

Nondegeneracy of hh at 00. The reduced energy hδh_\delta depends smoothly on δ\delta; at δ=0\delta=0 it equals the radial energy h0h_0, which satisfies D2h0(0)=−c0ID^2h_0(0)=-c_0I with c0>0c_0>0 (the radial case is a strict maximum of the reduced energy, by the same computation as in the radial benchmark). Since δ↦D2hδ(0)\delta\mapsto D^2h_\delta(0) is continuous, D2hδ(0)D^2h_\delta(0) remains invertible for ∣δ∣<δ1|\delta|<\delta_1 with δ1=c0/(2sup⁡∣δ∣≤1∥D2hδ(0)∥)\delta_1=c_0/(2\sup_{|\delta|\leq1}\|D^2h_\delta(0)\|). Also ∇hδ(0)=0\nabla h_\delta(0)=0 for every δ\delta because f1f_1 is even under x↦−xx\mapsto-x, so hδh_\delta is even. Hence (H4) holds at x0=0x_0=0 for ∣δ∣<δ1|\delta|<\delta_1.

Center correction. Because G~(⋅,ε)\widetilde{G}(\cdot,\varepsilon) is odd under x↦−xx\mapsto-x for this symmetric perturbation, ∂εG~(0,0)=0\partial_\varepsilon\widetilde{G}(0,0)=0, hence ξ1=0\xi_1=0 by (36).

Shape coefficient. The flux splits as g0=g0(0)+δg0(1)+O(δ2)g_0=g_0^{(0)}+\delta g_0^{(1)}+O(\delta^2), where g0(0)g_0^{(0)} is the constant radial flux. Therefore

PX(g02)=2δg0(0)PX(g0(1))+O(δ2).P_X(g_0^2)=2\delta g_0^{(0)}P_X(g_0^{(1)})+O(\delta^2).

For a generic cutoff χ\chi, the degree-2 component of g0(1)g_0^{(1)} is nonzero (it is the trace of f1f_1 on ∂BR(0)\partial B_R(0) under the Dirichlet problem, which is nonzero by elliptic regularity and unique continuation), so PX(g0(1))≠0P_X(g_0^{(1)})\neq0. Hence

ψ1=12JR−1PX(g02)=δψ1(1)+O(δ2),ψ1(1)=g0(0)JR−1PX(g0(1))≠0.\psi_1=\frac{1}{2}J_R^{-1}P_X(g_0^2)=\delta\psi_1^{(1)}+O(\delta^2),\qquad\psi_1^{(1)}=g_0^{(0)}J_R^{-1}P_X(g_0^{(1)})\neq0.

where JR−1J_R^{-1} acts on the degree-2 mode via (27). Since ξ1=0\xi_1=0, formula (49) gives

η1=χ0ψ1∘ω0=δχ0ψ1(1)∘ω0+O(δ2)≠0.\eta_1=\chi_0\psi_1\circ\omega_0=\delta\chi_0\psi_1^{(1)}\circ\omega_0+O(\delta^2)\neq0.

Thus η1≠0\eta_1\neq0 to first order in δ\delta, and consequently DR[d0]η1≠0D\mathcal{R}[d_0]\eta_1\neq0 generically (the kernel of DR[d0]D\mathcal{R}[d_0] is conjecturally trivial modulo volume; see Conjecture 12.1). This confirms that the first-order response is nonzero in a genuinely nonradial, nondegenerate situation satisfying (H1)–(H4).

Reproducible nonradial computational protocol

No numerical result is asserted in this draft, but the first-order theorem is directly testable. A reproducible two-dimensional protocol is:

  • (P1) choose a circular core C=B1(0)C=B_1(0) and a smooth compactly supported nonradial source inside B1/2(0)B_{1/2}(0), for example a sum of two unequal C∞C^\infty bumps;

  • (P2) for the prescribed area πR2\pi R^2, compute the reduced translated-ball energy h(x)h(x) on a neighborhood of a candidate center, locate a critical point x0x_0, and verify numerically that D2h(x0)D^2h(x_0) is nonsingular;

  • (P3) solve the Dirichlet problem on BR(x0)B_R(x_0), compute the boundary flux g0g_0, project g02g_0^2 onto Fourier modes, and solve the diagonal Jacobi equation for ψ1\psi_1;

  • (P4) compute the center coefficient ξ1\xi_1 from the desingularized finite-dimensional equation, transfer (ξ1,ψ1)(\xi_1,\psi_1) to the core-normal coefficient η1\eta_1, and evaluate the explicit derivative DR[d0]η1D\mathcal{R}[d_0]\eta_1;

  • (P5) for a sequence μk→∞\mu_k\to\infty, solve the constrained Euler–Lagrange system numerically and compare the normalized errors

μk(dμk−d0)−η1,μk(Fμk−F0)−DR[d0]η1\mu_k(d_{\mu_k}-d_0)-\eta_1,\qquad\mu_k(F_{\mu_k}-F_0)-D\mathcal{R}[d_0]\eta_1

in the norms predicted by Theorem 9.1.

A numerical section should report meshes, nonlinear tolerances, projection resolution, center-search tolerance, and the measured convergence rates. Until such data are produced, the protocol is illustrative rather than evidential.

Mode separation

For a nonradial source, the first correction is obtained by decomposing g02g_0^2 into spherical harmonics. Its constant component changes the multiplier, its degree-one component determines the center drift ξ1\xi_1 via the desingularized equation (36) (and does not contribute to the shape deformation ψ1\psi_1), and only degrees ℓ≥2\ell\ge2 deform the shape via

ψ1=12JR−1PX(g02).\psi_1 = \frac{1}{2}J_R^{-1}P_X(g_0^2).

On such a mode,

JR−1Yℓ=R2ℓ(ℓ+N−2)−(N−1)Yℓ.J_R^{-1}Y_\ell= \frac{R^2}{\ell(\ell+N-2)-(N-1)}Y_\ell.

This gives a direct spectral interpretation of the first-order smoothing effect of perimeter regularization.

Discussion and open problems

What has been proved

The large-μ\mu branch is used as a preparatory input rather than claimed as a new general existence mechanism. The new geometric content begins when that branch is transferred to the normal-return class OC\mathcal{O}_C. In particular, the return map is promoted from a pointwise geometric construction to a Fréchet-differentiable shape observable. The large-μ\mu problem itself is a singular perturbation around the finite-dimensional manifold of balls of fixed volume. The translation kernel is not an obstruction to be removed by an artificial coercivity assumption; it is resolved by the reduced center equation. The higher modes are governed by the invertible part of the spherical Jacobi operator. The resulting branch can be transferred to an arbitrary strictly convex coordinate core, remains inside OC\mathcal{O}_C, and induces a first-order return-map response, which is generically nonzero by the perturbative example of § 11.2.

What remains open

Several questions are not addressed here.

(i) Minimality and Morse index. The branch consists of critical domains. Determining its local minimizing or saddle character requires the constrained second variation of the full functional along the branch.

(ii) Existence of a selected center. The theorem assumes a nondegenerate critical point of hh inside the admissible center set. Genericity of (H4) alone follows from Sard’s lemma applied to ∇h\nabla h for generic ff; the open question is the joint compatibility with (H3) for non-spherical cores. The perturbative example of § 11.2 provides one non-spherical instance, but a systematic existence theory is open.

(iii) Second-order expansion. The implicit identity (36) determines ξ1\xi_1, but an explicit formula for it in terms of PDE data, together with a μ−2\mu^{-2} shape/return expansion, requires the next-order reduced equation and the second derivative of the return operator.

(iv) Range characterization. Theorem 8.2 establishes the derivative of the forward return map. Whether its range equals the full affine compatibility space of the linearized inverse problem is a separate question, stated precisely in Conjecture 12.1 below.

(v) Nonspherical large-μ\mu reference states. For Euclidean perimeter, the leading-order reference is spherical. Nonspherical reference states would require a modified leading geometric energy, for example anisotropic perimeter.

Conjecture 12.1 (Range). Let d0∈C2,α(M)d_0 \in C^{2,\alpha}(M) with Ω[d0]∈OC\Omega[d_0] \in\mathcal{O}_C. Then DR[d0]D\mathcal{R}[d_0] is injective on C2,α(M)/RC^{2,\alpha}(M)/\mathbb{R} (modulo the volume mode already absorbed by the chart (20)) but not surjective. Its image is a closed subspace of C1,α(M;RN)C^{1,\alpha}(M;\mathbb{R}^N) of codimension at most N+1N+1, corresponding to the modes ℓ=0\ell=0 (multiplier) and ℓ=1\ell=1 (translations) absorbed by the reduced center equation and the Lagrange multiplier. Generically, the codimension equals N+1N+1.

Remark 12.2. The conjecture is consistent with the structure of the first-hit equation: the return map is determined by dd through a scalar equation, so one expects injectivity; and the admissible return maps must be compatible with the geometry of CC, which excludes at least the N+1N+1 directions corresponding to the volume mode and to translations. The identification of the range with the compatibility space of [6] is the natural next step.

Conclusion

The classical Dirichlet-perimeter theory provides a local critical-shape branch near a nondegenerately selected ball when the perimeter weight is large. The contribution of the present paper is to place such a branch inside the geometric normal-return framework of OC\mathcal{O}_C and to determine how its discrete return dynamics responds.

Three geometric steps are essential. First, admissibility in OC\mathcal{O}_C is open under small C1C^1 thickness perturbations. Second, the translated spherical branch admits a unique core-normal representation whose first coefficient contains both center drift and genuine shape deformation. Third, the forward return operator R:d↦Fd\mathcal{R}:d\mapsto F_d is Fréchet C1C^1 and has an explicit derivative determined by the first-hit equation. Their combination yields the first-order shape-to-dynamics law

Fμ−F0=μ−1DR[d0]η1+oC1,α(μ−1).F_\mu-F_0=\mu^{-1}D\mathcal{R}[d_0]\eta_1+o_{C^{1,\alpha}}(\mu^{-1}).

The perturbative nonradial example of § 11.2 shows that the response coefficient η1\eta_1 is generically nonzero, confirming that the theorem is not vacuous.

This response formula is the main new result. It complements earlier work on the fixed return map, inverse reconstruction, and PDE-generated superlevel deformations by providing the general forward sensitivity operator with respect to boundary shape.

The next mathematical questions are consequently sharper than a generic shape-optimization extension: characterize the range and possible loss of rank of DR[d]D\mathcal{R}[d], obtain second-order return sensitivities, and determine how these operators interact with the affine/holonomy compatibility structure of the inverse problem. Numerical validation of the nonradial protocol in § 11 is also left for subsequent work.

Consistency with the PDE-generated deformation

We verify identity (60) in detail. Let η=−b0\eta=-b_0, where b0b_0 is the level-set speed of [7], related to g0g_0 by (66). We compute the three terms in (56) and compare with the coefficient A0A_0 of [7].

Term 1: ην=−b0ν\eta\nu=-b_0\nu. This is the direct boundary displacement.

Term 2: τ0δη[−b0]\tau_0\delta\eta[-b_0]. From (53) and (52),

δη[−b0]=−1χ0(I−n0⊗n0)L0−1(∇Mb0+b0Sw0).\delta\eta[-b_0]=-\frac{1}{\chi_0}(I-n_0\otimes n_0)L_0^{-1}(\nabla_M b_0+b_0S w_0).

Hence

τ0δη[−b0]=−τ0χ0(I−n0⊗n0)L0−1(∇Mb0+b0Sw0).\tau_0\delta\eta[-b_0]=-\frac{\tau_0}{\chi_0}(I-n_0\otimes n_0)L_0^{-1}(\nabla_M b_0+b_0S w_0).

Term 3: projection. The projection Π0ret\Pi_{0}^{\mathrm{ret}} removes the component along n0n_0 at F0(c)F_0(c), and is exactly the projection used in [7 equation for A_0] (the normal direction at the return point is n0(c)n_0(c), and the tangent space at F0(c)F_0(c) is TF0(c)MT_{F_0(c)}M).

Combining the three terms,

DR[d0](−b0)=Π0ret[−b0ν−τ0χ0(I−n0⊗n0)L0−1(∇Mb0+b0Sw0)].DR[d_0](-b_0)=\Pi_{0}^{\mathrm{ret}}\left[-b_0\nu-\frac{\tau_0}{\chi_0}(I-n_0\otimes n_0)L_0^{-1}(\nabla_M b_0+b_0Sw_0)\right].

This is exactly the expression for A0A_0 derived in [7 Section 8], up to the identification (66) of b0b_0 with the boundary flux. The sign is consistent with the convention η=−b0\eta=-b_0 (the superlevel deformation moves inward, whereas DRDR is computed for an outward perturbation; the two signs cancel in the projection). This establishes (60).

References

References

  1. [1]A. D. Alexandrov, A characteristic property of spheres, Ann. Mat. Pura Appl. 58 (1962), 303–315.DOI
  2. [2]V. Amato, C. Nitsch, C. Trombetti, and F. Villone, On some functionals involving torsional rigidity, principal eigenvalue and perimeter, Potential Analysis 65 (2026), article 25. doi:10.1007/s11118-026-10328-2.DOI
  3. [3]M. Barkatou, Some geometric properties for a class of non-Lipschitz domains, New York J. Math. 8 (2002), 189–213.
  4. [4]M. Barkatou, D. Seck, and I. Ly, An existence result for an interior electromagnetic casting problem, Cent. Eur. J. Math. 4 (2006), no. 4, 573–584. doi:10.2478/s11533-006-0026-5.DOI
  5. [5]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. doi:10.62762/JNDA.2026.569839.DOI
  6. [6]M. El Morsalani and M. Barkatou, Inverse problems for the return map in the class O_C: reconstruction and identifiability, preprint (2026), arXiv:2604.24915.arxiv.org/abs/2604.24915
  7. [7]M. El Morsalani and M. Barkatou, Elliptic superlevel sets in the class O_C: boundary-layer geometry and stability of the return map, preprint (2026), arXiv:2607.16912.arxiv.org/abs/2607.16912
  8. [8]G. Buttazzo, F. P. Maiale, D. Mazzoleni, G. Tortone, and B. Velichkov, Regularity of the optimal sets for a class of integral shape functionals, Arch. Ration. Mech. Anal. 248 (2024), article 44. doi:10.1007/s00205-024-01984-y.DOI
  9. [9]F. Caubet, M. Dambrine, and J. Darde, On the penalization by the perimeter in shape optimization applied to Dirichlet inverse obstacle problem, ESAIM Control Optim. Calc. Var. 32 (2026), article 7. doi:10.1051/cocv/2025093.DOI
  10. [10]A. Celentano, C. Nitsch, and C. Trombetti, On a Serrin type overdetermined problem, Milan J. Math. 92 (2024), 579–590. doi:10.1007/s00032-024-00405-9.
  11. [11]M. C. Delfour and J.-P. Zolesio, Shapes and Geometries: Metrics, Analysis, Differential Calculus, and Optimization, SIAM, Philadelphia, 2001.
  12. [12]F. Feppon, G. Allaire, and C. Dapogny, A variational formulation for computing shape derivatives of geometric constraints along rays, ESAIM Math. Model. Numer. Anal. 54 (2020), no. 1, 181–228. doi:10.1051/m2an/2019056.DOI
  13. [13]M. Hayouni and A. Novruzi, Sufficient condition for the existence of solutions of a free boundary problem, Quart. Appl. Math. 60 (2002), no. 3, 425–435. doi:10.1090/qam/1914434.DOI
  14. [14]A. Henrot and M. Pierre, Variation et optimisation de formes: une analyse géométrique, Springer, Berlin, 2005.
  15. [15]F. Immler and C. Traut, The flow of ODEs: formalization of variational equation and Poincaré map, J. Autom. Reason. 62 (2019), 215–236. doi:10.1007/s10817-018-9449-5.DOI
  16. [16]A. Novruzi, C^{2,α} existence result for a class of shape optimization problems, SIAM J. Control Optim. 43 (2004), no. 1, 174–193. doi:10.1137/S0363012902382401.DOI
  17. [17]P. I. Plotnikov and J. Sokolowski, Geometric framework for gradient flow in shape optimization, Uspekhi Mat. Nauk 81 (2026), no. 3, 51–150. doi:10.4213/rm10296.DOI
  18. [18]J. Sokolowski and J.-P. Zolesio, Introduction to Shape Optimization: Shape Sensitivity Analysis, Springer, Berlin, 1992.

Paper details

vlnc:W_afmwh9f3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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