A coefficient vector replaces the moment-first step
Homogeneous numerical integration combines Euler's identity and Stokes' theorem to integrate homogeneous polynomials over a polytope P. Existing algorithms return either a scalar integral or reusable moment families. The paper's move is different: it propagates functionals through the face complex once for a given (P, q) and returns a signed vertex-derivative coefficient functional on H_q, the space of degree-q forms.
Concretely, the construction produces a polynomial-independent vector C(q). If E_q is its evaluation matrix on a monomial basis of H_q, then b_q = E_q C(q) is the resulting degree-q moment vector. The paper stresses that the novelty is not vertex support, point support, or the mere existence of such a coefficient vector, but constructing one directly through local Euler–Stokes propagation while retaining face-incidence and anchor structure.
The face descent aggregates equal terminal functionals
The underlying recurrence is a facewise Euler–Stokes identity. For a k-face F and multi-index α with |α| ≤ q, the integral I(F, α) of D^α p over F satisfies a relation that sums contributions from relative facets G ≺ F plus a term involving the anchor z_F and derivatives of one higher order. At a vertex v, I({v}, α) reduces to D^α p(v).
The coefficient-propagation algorithm stores, for each state (F, α), a sparse map from terminal pairs (v, β) to coefficients. States are processed by decreasing dim F + q − |α|. A vertex state is initialized as {(v, α) ↦ 1}; otherwise the recurrence is applied formally, child states are memoized, scaled maps are added, duplicate terminal pairs are collected, and exact zeros are removed. The returned map at (P, 0) is C(q).
Anchors change coefficients but not the induced functional
Anchors are chosen points in each positive-dimensional face, and they enter the recurrence through the height h_{F,G}. The paper proves that if two anchor choices produce exact outputs C and C̃, then J(C − C̃) = 0 when both are embedded in a common vertex-derivative information set. So anchors can change the displayed coefficients, but not the functional they induce on H_q.
The choice can also change the number of coefficients. If an anchor lies on a relative facet, the corresponding height h_{F,G} vanishes and that boundary branch disappears. Coherent face-vertex anchors can therefore produce much sparser formal outputs, although the paper does not claim this solves the separate sparsest-representation problem.
A Riesz framework measures restricted-formula error
The paper equips H_q with the L2(P) inner product and defines the moment matrix M_q, the moment vector b_q, and a reproducing kernel K_{P,q}. The Riesz representative R_{P,q} satisfies ∫_P p = ⟨p, R_{P,q}⟩_P. This is described as the exact-degree kernel of H_q, not the usual degree-at-most Christoffel setting; in particular K_{P,q}(0,0) = 0 for q > 0.
Theorem 3.1 gives the worst-case L2(P) error of any coefficient vector C as the norm of R_{P,q} minus the sum of C_j κ_j. It also characterizes exactness as JC = ρ, shows that all exact vectors form C_0 + ker J, and gives a weighted minimum C_W = W^{-1} J^T (J W^{-1} J^T)^† ρ when ρ lies in the range of J. The paper notes that a coefficient norm is meaningful only after derivatives of different orders are scaled, so W must encode a normalization or evaluation cost.
Degree-q and degree-2q vectors satisfy a compatibility relation
The paper reports that the independently generated degree-q and degree-2q coefficient vectors satisfy an explicit compatibility relation. If both HNI formulas are exact, then b_q = E_q C(q), and the moment matrix entries can be written using the degree-2q coefficients applied to products of basis elements.
With r = M_q^{-1} b_q recovering R_{P,q}, and P(r) defined by v_q(x) R_{P,q}(x) = P(r) v_{2q}(x), the coefficient vectors satisfy P(r) E_{2q} C(2q) = E_q C(q). For fixed and known r, this is a linear relation between the two HNI vectors. Conversely, exact HNI vectors at degrees q and 2q determine b_q and M_q, and hence determine r.
Reported examples show pruning without a conditioning claim
The paper reports term counts for several domains. For the cube at q = 4, the propagated/restricted terms are 248 center, 22 vertex, and 12 order-≤2, against an unrestricted baseline of 15. For the four-cube at q = 4, the counts are 1104 center, 35 vertex, and 31 order-≤2, against a baseline of 35. For the non-convex L-polygon at q = 4, sparse affine anchors give 54 terms and a finite search over 6·2^6 = 384 assignments gives 30, against a baseline of 5.
For the seeded non-symmetric convex hexagon, the same class reduces 90 to 30 terms at q = 4 and 270 to 90 at q = 8. A separate benchmark at q = 4 propagates every one of the n 2^n assignments on seeded perturbed convex polygons; on an Apple M3 it grows from 64 assignments in 0.005 seconds for n = 4 to 10,240 in 5.47 seconds for n = 10, while reducing midpoint-anchor counts from 60 to 10 and from 150 to 60. The paper states this supports offline use for modest planar elements, not polynomial-time scalability, and that fewer terms do not imply better conditioning because S(C) sometimes increases after support minimization.
Scope covers oriented non-convex domains in any dimension
The construction covers convex polytopes and oriented non-convex polyhedral domains in arbitrary dimension. For the non-convex case, one must additionally supply the oriented boundary complex, including boundary components, incidence signs, and a consistent outward side; it is not a decomposition of P into convex elementary cells. Convexity is not used in the divergence identity, but the supplied boundary cells are assumed to form a consistently oriented finite polyhedral complex.
The paper explicitly does not claim lower cost or greater general reusability than optimized moment-table methods, nor global cardinality optimality or improved conditioning. The exact-rational reductions establish algebraic exactness, not a general stability result. The comparison with unknown-node positive moment-fitting cubature does not apply to every cubature construction, in particular not to fixed-node linear methods or compression of an already available positive formula.
The reusable functional matters most when integrands arrive as vertex data
The paper's own priority is the structured lift: a reusable functional adapted to vertex-derivative data, with anchor dependence and nonuniqueness exposed, exact restricted-information post-processing and Riesz certification enabled, and a degree-q/degree-2q coefficient relation made explicit. The four-cube example shows dimension is not restricted to 2 or 3, while the planar tests show finite searches over vertex anchors can prune this structured output offline.
For readers deciding whether this matters, the paper's own scenario split is the useful threshold: the structured lift is described as particularly useful when integrands are naturally supplied through local vertex data rather than global monomial coefficients. The reported pruning is an offline preprocessing effect for modest planar elements, and the paper leaves scalable anchor selection with simultaneous sparsity and conditioning control to future work.
Sources used in this answer
Direct Vertex-Derivative Coefficient Propagation forHomogeneous Numerical Integration (HNI)
The paper constructs a signed vertex-derivative coefficient functional on H_q directly through Euler–Stokes face propagation for a fixed (P, q), without first forming the moment vector, and shows that anchors change coefficients but not the induced functional, that a Riesz framework gives an intrinsic L2(P) error for restricted formulas, and that degree-q and degree-2q coefficient vectors satisfy an explicit compatibility relation.