Second derivatives of scalar PH losses

Let K have finitely many nonempty faces, and let F be a family of real filter functions on a real normed parameter space. On an open set U, assume each face value is C² and the complete weak ordering of face values agrees with that at x₀ ∈ U. Neither a finite ambient vertex type nor an inner product on the parameter space is needed.

The fixed full-template lift L is C² on the same U and represents the sublevel diagram throughout U. It is the composite R ∘ G, where G collects all face values and R = faceValueSelection is fixed and continuous linear. Thus both derivatives select the corresponding face derivatives; there is no differentiation of a changing sort or pairing.

C2ThroughAllRepresentatives V D₀ requires the pullback of V to be C² near every ordered representative of D₀, for every pair of coordinate counts. The domain includes reversed and diagonal finite pairs. At z = L(x₀), write φ = V ∘ orderedQ. The coordinate neighborhood supplied by this condition may shrink the scalar neighborhood: the scalar function ℓ(x) = V(dgmₚ(F(x))) is C² on an open W with x₀ ∈ W ⊆ U.

Face formula

Use the convention D²f(x)[h,k] = D(Df)(x)[h][k]. The second derivative is a continuous bilinear map, represented in Lean by a curried continuous linear map. Let uⱼ be the existing coordinate units and s(j) the existing injective slot-to-face assignment. Keep β from faceLossCoefficient and define

\[ \Gamma_{\sigma\tau} = D^2\varphi(z)[R e_\sigma,R e_\tau], \]

where eσ is a unit in the complete face-value space. This equals D²φ(z)[uⱼ,uₗ] for σ = s(j) and τ = s(l); an unused face gives a zero row and column. The corresponding first-order evaluation is the existing βσ.

The bilinear form faceCoordinateLossHessian is

\[ B[h,k] = \sum_\sigma\sum_\tau \Gamma_{\sigma\tau}\,DF_\sigma(x_0)[h]\,DF_\tau(x_0)[k] + \sum_\sigma\beta_\sigma\,D^2F_\sigma(x_0)[h,k]. \]

sublevelDiagramScalarLoss_hasFDerivAt_fderiv proves that the derivative map Dℓ has derivative B at x₀. This establishes existence before identifying the nested fderiv, so the theorem cannot be satisfied by its default value at a nondifferentiable function. The C² condition implies the existing C¹ condition, and the first derivative retains the earlier face formula.

The proof differentiates the ordinary composition rule. Its two terms are D²φ(z)[DL h,DL k] and Dφ(z)[D²L[h,k]]. Expanding in the face basis twice and once, respectively, gives the displayed sums. Symmetry of C² second derivatives gives Γστ = Γτσ and B[h,k] = B[k,h]. The standard curried operator norm satisfies

\[ \|D^2\ell(x_0)\| \le \sum_\sigma\sum_\tau |\Gamma_{\sigma\tau}| \|DF_\sigma(x_0)\|\,\|DF_\tau(x_0)\| + \sum_\sigma |\beta_\sigma|\,\|D^2F_\sigma(x_0)\|. \]

Empty face and coordinate sets are included. If the face second derivatives vanish at x₀, the single sum vanishes. An affine family near x₀ therefore has the usual Hessian pullback by its linear part. This is an explicit specialization; a nonlinear filtration generally needs the curvature sum.

Presentations and ties

Any two admissible C² local lifts of the same diagram family produce the same final bilinear form because both differentiate the same scalar function. Coordinate counts may differ. Neither the coefficient vectors nor the ambient Hessian arrays are asserted to be unique across presentations.

A persistent diagonal pair retains both coordinates. Equal face-value paths have equal first and second derivatives at the base point. Cancellation uses these equalities, without assuming that either ambient coefficient vanishes. The squared signed endpoint-difference loss has a nonzero ambient Hessian at a diagonal pair even though its restriction to the diagonal is zero.

Examples

  • A singleton with value t², with loss essentialBirthSum², has scalar loss t⁴. At t = 1, the second derivative is 12: the coordinate-loss term is 8 and the face-curvature term is 4.
  • An edge has vertex values 0, 1+s and edge value 3+t+s² on |s|, |t| < 1/4. Its degree-zero diagram has essential birth 0 and finite interval (1+s,3+t+s²), obtained from the existing homology fixture by grade transport. For squared signed endpoint difference, the loss is (2+t−s+s²)². At the origin the Hessian is [[10,−2],[−2,2]]; its coordinate contribution is [[2,−2],[−2,2]] and its face-curvature contribution is [[8,0],[0,0]].
  • An edge with every face valued t² retains a finite template pair. The signed loss's nonzero birth and death weights cancel against equal face second derivatives. For its square, a nonzero ambient Hessian pulls back to zero. Both scalar losses have zero second derivative.
  • A singleton with face value 1+2*t² has minimal, padded, and nontrivially relabeled C² lifts. The larger lifts retain two unequal moving diagonal pairs; all three yield the genuine scalar Hessian 4*h*k. Additional presentation examples swap unequal essential paths. The absolute endpoint-difference loss at the empty diagram fails the C² condition, because that condition implies the already-disproved C¹ condition there.

The squared losses are coordinate polynomials on every ordered coordinate space, including representatives augmented by diagonal finite pairs. Their regularity is not restricted to a positive cone.

Scope and sources

The local-lift and all-representatives conditions specialize Leygonie–Oudot–Tillmann, Definitions 3.3 and 3.10, Proposition 3.14 and Theorem 4.7. The displayed face Hessian formula is a library consequence of ordinary second-order calculus, not a formula attributed to that paper. Its analytic ingredients are pinned mathlib's composition of varying continuous linear maps and symmetry of second derivatives.

All conclusions concern ordinary scalar derivatives. No topology or Hessian on Diagram is installed, and no derivative at an ordering frontier is claimed. The examples do not depend on the simplicial reducer adapter or close the separate filled-triangle fixture under #45.