Files
ConformalLabpp/code/include/inversive_distance_functional.hpp
Tarik Moussa 2dc4ddcc32 perf(inv-dist): B1 port — block-FD Hessian for newton_inversive_distance
The Inversive-Distance solver built its Hessian inline by full finite
differences: n perturbations × a full O(F) gradient eval = O(n·F) per Newton
iteration (quadratic in mesh size), with no fast path at all (api-performance
audit B1, second half).

Port the per-face block-FD scheme already used by HyperIdeal (Phase 9b):
the gradient decomposes by face (G_v = Θ_v − Σ_{f∋v} α_v, and each face's
angles depend only on its 3 vertex DOFs), so the Hessian decomposes into
per-face 3×3 blocks.  Cost drops to O(F) face evaluations, a ≈ n/6 speed-up.

- inversive_distance_functional.hpp: add the pure 3→3 kernel
  inversive_distance_face_grad_contribs (returns the per-face contribution
  −α to G; mirrors the gradient's face-skip on ℓ²≤0 exactly).
- inversive_distance_hessian.hpp (new): full-FD baseline + block-FD + sym
  variants, mirroring hyper_ideal_hessian.hpp.
- newton_solver.hpp: drop the inline full-FD lambda; call
  inversive_distance_hessian_block_fd_sym.
- test: InversiveDistance_BlockFDHessianMatchesFullFD cross-validates the
  two Hessians entry-wise on a perturbed (off-equilibrium) config.

291/291 CGAL tests pass; all Inversive-Distance convergence tests unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-05-31 19:44:50 +02:00

429 lines
19 KiB
C++
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#pragma once
// Copyright (c) 2024-2026 Tarik Moussa.
// SPDX-License-Identifier: MIT
// inversive_distance_functional.hpp
//
// Phase 9a.2 — Inversive-distance circle-packing functional (Luo 2004).
//
// VERTEX-based circle packing. Each vertex carries a circle of radius
// r_i = exp(u_i). The inversive distance I_ij between two adjacent
// circles is a constant of the edge, derived once from the initial
// geometry via Bowers-Stephenson 2004.
//
// This is the FACE-DUAL of CPEuclideanFunctional (Phase 9a.1). The
// correspondence is I_ij = cos θ_e (Glickenstein 2011 §5).
//
// ┌──────────────────────────────────────────────────────────────────────────┐
// │ Mathematical model │
// │ ────────────────── │
// │ │
// │ Variables: u_i = log r_i (per vertex; r_i is the radius) │
// │ Constants: I_ij (per edge; inversive distance) │
// │ Θ_v (per vertex; target cone angle) │
// │ │
// │ Bowers-Stephenson (init from initial geometry): │
// │ I_ij = ( _ij² r_i² r_j² ) / ( 2 r_i r_j ) │
// │ │
// │ Edge length (Luo 2004 §3, Glickenstein 2011 eq. 2.1): │
// │ _ij(u)² = exp(2 u_i) + exp(2 u_j) + 2 I_ij exp(u_i + u_j) │
// │ = r_i² + r_j² + 2 I_ij r_i r_j │
// │ │
// │ Triangle angles: same half-tangent law of cosines as the │
// │ Euclidean functional (numerically stable). │
// │ │
// │ Gradient (Luo 2004 Lemma 3.1): │
// │ ∂E/∂u_v = Θ_v Σ_{T ∋ v} α_v(T) │
// │ │
// │ Energy: path integral E(u) = ∫₀¹ ⟨G(tu), u⟩ dt │
// │ (Luo's 1-form is closed; we use 10-point Gauss-Legendre │
// │ quadrature, identical to euclidean_functional.hpp) │
// │ │
// │ Hessian: finite-difference for the MVP port; an analytic form is │
// │ given in Glickenstein 2011 eq. (4.6) and may be added │
// │ later for performance. │
// └──────────────────────────────────────────────────────────────────────────┘
//
// Relation to euclidean_functional.hpp
// ────────────────────────────────────
// The two are structurally identical in:
// • DOF layout (per vertex), DOF index sentinel (1 = pinned)
// • Gradient pattern (Θ Σ α)
// • Energy via path integral (same Gauss-Legendre constants)
// • Halfedge convention (h0/h1/h2, source pattern, α opposite-edge)
//
// They differ ONLY in:
// • Per-edge constant: λ°_ij (log²-length) vs I_ij (inversive distance)
// • Edge-length formula:
// Euclidean: _ij = exp((λ°_ij + u_i + u_j) / 2)
// Inversive distance: _ij² = exp(2u_i) + exp(2u_j)
// + 2 I_ij exp(u_i + u_j)
//
// In particular at the tangential limit I_ij = 1 the inversive-distance length
// reduces to (exp(u_i) + exp(u_j))² ⇒ _ij = r_i + r_j (tangential circles),
// which is *different* from the Euclidean-conformal length even at the same
// initial geometry. The two functionals describe distinct geometric objects.
//
// Property-map name prefix: "iv:" (vertex) and "ie:" (edge).
#include "conformal_mesh.hpp"
#include "constants.hpp"
#include "gauss_legendre.hpp"
#include "euclidean_geometry.hpp" // euclidean_angles(λ12, λ23, λ31)
#include <CGAL/boost/graph/iterator.h>
#include <vector>
#include <cmath>
#include <cstdint>
#include <iostream>
namespace conformallab {
// ── Property-map type aliases ────────────────────────────────────────────────
/// Property map vertex → `int` for the Inversive-Distance functional.
using IDVMapI = ConformalMesh::Property_map<Vertex_index, int>;
/// Property map vertex → `double` for the Inversive-Distance functional.
using IDVMapD = ConformalMesh::Property_map<Vertex_index, double>;
/// Property map edge → `double` for the Inversive-Distance functional.
using IDEMapD = ConformalMesh::Property_map<Edge_index, double>;
// ── Persistent map bundle ─────────────────────────────────────────────────────
/// Bundle of the four property maps consumed by the Inversive-Distance
/// circle-packing functional (Luo 2004 / Bowers-Stephenson 2004).
struct InversiveDistanceMaps {
IDVMapI v_idx; ///< DOF index per vertex (1 = pinned / u_v = 0)
IDVMapD theta_v; ///< target cone angle Θ_v (default 2π)
IDVMapD r0; ///< initial radius r_i^(0) (default 1)
IDEMapD I_e; ///< inversive distance I_ij (per edge, constant)
};
/// Attach the four inversive-distance property maps to `mesh` and
/// return their handles.
///
/// Defaults are intentionally trivial — every real use of this
/// functional must call `compute_inversive_distance_init_from_mesh()`
/// next to populate `r0` and `I_e` from the input geometry.
/// * `v_idx[v] = -1` (all vertices pinned initially)
/// * `theta_v[v] = 2π` (regular interior vertex)
/// * `r0[v] = 1.0` (placeholder)
/// * `I_e[e] = 1.0` (tangential default — overwritten by init step)
///
/// The maps use the `"iv:"` / `"ie:"` prefix so they do not collide
/// with the Euclidean / Spherical / HyperIdeal / CP-Euclidean maps.
inline InversiveDistanceMaps setup_inversive_distance_maps(ConformalMesh& mesh)
{
InversiveDistanceMaps m;
m.v_idx = mesh.add_property_map<Vertex_index, int> ("iv:idx", -1 ).first;
m.theta_v = mesh.add_property_map<Vertex_index, double>("iv:theta", TWO_PI ).first;
m.r0 = mesh.add_property_map<Vertex_index, double>("iv:r0", 1.0 ).first;
m.I_e = mesh.add_property_map<Edge_index, double>("ie:I", 1.0 ).first;
return m;
}
/// Assign sequential DOF indices `0..n-1` to every vertex.
///
/// **Note:** this overload assigns indices to ALL vertices unconditionally.
/// Any `v_idx` set before the call is overwritten. To pin a gauge vertex,
/// either use the two-argument overload below, or set `m.v_idx[v] = -1`
/// **after** this call. Pinning before this call has no effect.
///
/// For a closed mesh, exactly one pin is required to remove the
/// global rotational mode.
inline int assign_inversive_distance_vertex_dof_indices(ConformalMesh& mesh,
InversiveDistanceMaps& m)
{
int idx = 0;
for (auto v : mesh.vertices()) m.v_idx[v] = idx++;
return idx;
}
/// Assign sequential DOF indices to all vertices, pinning `gauge`
/// (`m.v_idx[gauge] = -1`). Use this overload on closed meshes to fix
/// the rotational gauge mode in a single call.
///
/// \returns The number of free DOFs assigned (`num_vertices 1`).
inline int assign_inversive_distance_vertex_dof_indices(ConformalMesh& mesh,
InversiveDistanceMaps& m,
Vertex_index gauge)
{
int idx = 0;
for (auto v : mesh.vertices())
m.v_idx[v] = (v == gauge) ? -1 : idx++;
return idx;
}
/// Count the free DOFs (vertices with `v_idx >= 0`).
inline int inversive_distance_dimension(const ConformalMesh& mesh,
const InversiveDistanceMaps& m)
{
int dim = 0;
for (auto v : mesh.vertices()) if (m.v_idx[v] >= 0) ++dim;
return dim;
}
/// Two-phase initialisation from initial mesh geometry. Mirrors the
/// role of `compute_lambda0_from_mesh` in the Euclidean functional, but
/// adapted to Luo's vertex-based radius parametrisation.
///
/// **Phase 1.** Pick a positive radius per vertex:
/// \code
/// r_i^(0) = (1/3) · min{_e : e adjacent to v_i}
/// \endcode
/// This is a heuristic — the user may override `m.r0[v]` for any
/// vertex between `setup_inversive_distance_maps()` and this call.
///
/// **Phase 2.** Compute the per-edge inversive distance via the
/// Bowers-Stephenson 2004 identity:
/// \code
/// I_ij = ( _ij² r_i² r_j² ) / ( 2 r_i r_j )
/// \endcode
///
/// \pre Every edge has positive 3-D length.
/// \pre Radii produced in Phase 1 are positive (degenerate isolated
/// vertices fall back to `r_i = 1`).
/// \post Every `I_e[e] > -1` for a valid packing. The chosen
/// Phase-1 heuristic keeps `I_e > 0` for most real meshes.
inline void compute_inversive_distance_init_from_mesh(ConformalMesh& mesh,
InversiveDistanceMaps& m)
{
// Phase 1: r_i = (1/3) · min adjacent edge length.
for (auto v : mesh.vertices()) {
double min_len = std::numeric_limits<double>::infinity();
for (auto h : CGAL::halfedges_around_target(v, mesh)) {
auto p1 = mesh.point(mesh.source(h));
auto p2 = mesh.point(mesh.target(h));
double dx = p1.x() - p2.x();
double dy = p1.y() - p2.y();
double dz = p1.z() - p2.z();
double len = std::sqrt(dx*dx + dy*dy + dz*dz);
if (len < min_len) min_len = len;
}
m.r0[v] = (std::isfinite(min_len) && min_len > 1e-15)
? min_len / 3.0
: 1.0;
}
// Phase 2: I_ij from initial geometry.
for (auto e : mesh.edges()) {
auto h = mesh.halfedge(e);
auto vi = mesh.source(h);
auto vj = mesh.target(h);
auto p1 = mesh.point(vi);
auto p2 = mesh.point(vj);
double dx = p1.x() - p2.x();
double dy = p1.y() - p2.y();
double dz = p1.z() - p2.z();
double l2 = dx*dx + dy*dy + dz*dz;
double ri = m.r0[vi];
double rj = m.r0[vj];
m.I_e[e] = (l2 - ri*ri - rj*rj) / (2.0 * ri * rj);
}
}
// ── Internal helpers ──────────────────────────────────────────────────────────
namespace id_detail {
inline double dof_val(int idx, const std::vector<double>& x) noexcept
{
return idx >= 0 ? x[static_cast<std::size_t>(idx)] : 0.0;
}
// halfedge_to_index is defined in conformal_mesh.hpp.
inline std::size_t hidx(Halfedge_index h) noexcept { return halfedge_to_index(h); }
// Inversive-distance edge length squared: ℓ² = exp(2u_i) + exp(2u_j) + 2 I r_i r_j
// where r_i = exp(u_i), so: ℓ² = r_i² + r_j² + 2 I r_i r_j.
// Returns -1 if the result is non-positive (degenerate; the caller skips the face).
inline double edge_length_squared(double u_i, double u_j, double I_ij) noexcept
{
double ri = std::exp(u_i);
double rj = std::exp(u_j);
double l2 = ri*ri + rj*rj + 2.0 * I_ij * ri * rj;
return l2 > 0.0 ? l2 : -1.0;
}
} // namespace id_detail
/// Inversive-Distance gradient `G_v = Θ_v Σ_faces α_v(face)`. Same
/// half-edge corner-angle storage convention as `euclidean_gradient`.
inline std::vector<double> inversive_distance_gradient(
const ConformalMesh& mesh,
const std::vector<double>& x,
const InversiveDistanceMaps& m)
{
const int n = inversive_distance_dimension(mesh, m);
std::vector<double> G(static_cast<std::size_t>(n), 0.0);
const std::size_t nh = mesh.number_of_halfedges();
std::vector<double> h_alpha(nh, 0.0);
// Pass 1 — per face, compute corner angles via the law of cosines.
// We reuse euclidean_angles(λ12, λ23, λ31) which takes 2·log() per edge.
for (auto f : mesh.faces()) {
Halfedge_index h0 = mesh.halfedge(f);
Halfedge_index h1 = mesh.next(h0);
Halfedge_index h2 = mesh.next(h1);
Vertex_index v1 = mesh.source(h0);
Vertex_index v2 = mesh.source(h1);
Vertex_index v3 = mesh.source(h2);
Edge_index e12 = mesh.edge(h0);
Edge_index e23 = mesh.edge(h1);
Edge_index e31 = mesh.edge(h2);
double u1 = id_detail::dof_val(m.v_idx[v1], x);
double u2 = id_detail::dof_val(m.v_idx[v2], x);
double u3 = id_detail::dof_val(m.v_idx[v3], x);
double l12sq = id_detail::edge_length_squared(u1, u2, m.I_e[e12]);
double l23sq = id_detail::edge_length_squared(u2, u3, m.I_e[e23]);
double l31sq = id_detail::edge_length_squared(u3, u1, m.I_e[e31]);
// A non-real circle configuration (ℓ² ≤ 0) has no limiting angle — skip it.
if (l12sq <= 0 || l23sq <= 0 || l31sq <= 0) continue;
// euclidean_angles expects 2·log() per edge — feed log(ℓ²).
// For a triangle-inequality-violating face euclidean_angles returns the
// *limiting* angles (π opposite the over-long edge, 0/0 otherwise) with
// valid=false. We deliberately do NOT skip on !fa.valid: using those
// limiting angles is the convex C¹ extension onto the infeasible region,
// so Newton can pass through a flip instead of stalling (Finding 9 —
// mirrors the Euclidean/Spherical fix in Finding 1). The angles come
// from genuine Euclidean side lengths, so the extension is geometric.
auto fa = euclidean_angles(std::log(l12sq), std::log(l23sq), std::log(l31sq));
h_alpha[id_detail::hidx(h0)] = fa.alpha3;
h_alpha[id_detail::hidx(h1)] = fa.alpha1;
h_alpha[id_detail::hidx(h2)] = fa.alpha2;
}
// Pass 2 — accumulate vertex gradient.
for (auto v : mesh.vertices()) {
int iv = m.v_idx[v];
if (iv < 0) continue;
double sum_alpha = 0.0;
for (auto h : CGAL::halfedges_around_target(v, mesh)) {
if (mesh.is_border(h)) continue;
sum_alpha += h_alpha[id_detail::hidx(mesh.prev(h))];
}
G[static_cast<std::size_t>(iv)] = m.theta_v[v] - sum_alpha;
}
return G;
}
/// Per-face contribution to the Inversive-Distance gradient, as a pure
/// 3→3 kernel of the three local vertex DOFs `(u1,u2,u3)` and the three
/// edge inversive distances `(I12,I23,I31)`. No mesh, no property maps —
/// used by the per-face block-FD Hessian in `inversive_distance_hessian.hpp`.
///
/// Returns the three values this face *adds* to the global gradient at
/// `(v1,v2,v3)`. Since `G_v = Θ_v Σ_faces α_v`, the per-face contribution
/// is the **negative** corner angles `(−α₁,−α₂,−α₃)`. A non-real circle
/// configuration (any `ℓ² ≤ 0`) contributes nothing — exactly mirroring the
/// face-skip (`continue`) in `inversive_distance_gradient`, so the block-FD
/// Hessian and the full-FD Hessian see the same per-face support.
struct IDFaceGradContribs {
double g1; ///< contribution to G at v₁ (= −α₁)
double g2; ///< contribution to G at v₂ (= −α₂)
double g3; ///< contribution to G at v₃ (= −α₃)
};
inline IDFaceGradContribs inversive_distance_face_grad_contribs(
double u1, double u2, double u3,
double I12, double I23, double I31)
{
double l12sq = id_detail::edge_length_squared(u1, u2, I12);
double l23sq = id_detail::edge_length_squared(u2, u3, I23);
double l31sq = id_detail::edge_length_squared(u3, u1, I31);
// Same face-skip as inversive_distance_gradient: a non-real circle
// configuration has no limiting angle, so the face contributes 0.
if (l12sq <= 0.0 || l23sq <= 0.0 || l31sq <= 0.0)
return {0.0, 0.0, 0.0};
// euclidean_angles deliberately keeps its limiting (valid=false) angles
// here — the convex C¹ extension — matching the gradient's choice not to
// skip on !fa.valid. Corner at v_k = fa.alpha_k (see gradient Pass-2 trace).
auto fa = euclidean_angles(std::log(l12sq), std::log(l23sq), std::log(l31sq));
return {-fa.alpha1, -fa.alpha2, -fa.alpha3};
}
/// Inversive-Distance energy `E(u) = ∫₀¹ ⟨G(t·u), u⟩ dt`, evaluated
/// with 10-point Gauss-Legendre (constants shared with `euclidean_energy`).
inline double inversive_distance_energy(
const ConformalMesh& mesh,
const std::vector<double>& x,
const InversiveDistanceMaps& m)
{
const double* gl_s = gl10_nodes();
const double* gl_w = gl10_weights();
const std::size_t n = x.size();
double E = 0.0;
std::vector<double> tx(n);
for (int k = 0; k < 10; ++k) {
double t = (1.0 + gl_s[k]) * 0.5;
double wt = gl_w[k] * 0.5;
for (std::size_t i = 0; i < n; ++i) tx[i] = t * x[i];
auto G = inversive_distance_gradient(mesh, tx, m);
double dot = 0.0;
for (std::size_t i = 0; i < n; ++i) dot += G[i] * x[i];
E += wt * dot;
}
return E;
}
/// FD gradient check for the Inversive-Distance functional (central diff).
/// Uses the same **relative** error criterion as every other gradient check:
/// `|analytic fd| / max(1, |analytic|) < tol`.
inline bool gradient_check_inversive_distance(
const ConformalMesh& mesh,
const std::vector<double>& x,
const InversiveDistanceMaps& m,
double eps = 1e-5,
double tol = 1e-4)
{
auto G = inversive_distance_gradient(mesh, x, m);
const std::size_t n = G.size();
bool ok = true;
for (std::size_t i = 0; i < n; ++i) {
std::vector<double> xp = x, xm = x;
xp[i] += eps;
xm[i] -= eps;
double Ep = inversive_distance_energy(mesh, xp, m);
double Em = inversive_distance_energy(mesh, xm, m);
double fd = (Ep - Em) / (2.0 * eps);
double err = std::abs(G[i] - fd);
double scale = std::max(1.0, std::abs(G[i]));
if (err / scale > tol) {
std::cerr << "[inversive-distance] FD gradient mismatch at DOF " << i
<< ": analytic=" << G[i]
<< " FD=" << fd
<< " rel-err=" << (err / scale) << "\n";
ok = false;
}
}
return ok;
}
/// Newton equilibrium check: returns `true` iff the gradient at `x`
/// is below `tol` in infinity norm (Σ adj-face angles equal Θ_v).
inline bool is_inversive_distance_equilibrium(
const ConformalMesh& mesh,
const std::vector<double>& x,
const InversiveDistanceMaps& m,
double tol = 1e-8)
{
auto G = inversive_distance_gradient(mesh, x, m);
for (double g : G)
if (std::abs(g) > tol) return false;
return true;
}
} // namespace conformallab