Fast Newton Transform (FNT)

• FNT is a fast algorithm for multivariate Newton interpolation in downward closed polynomial spaces, efficiently computing Newton coefficients using a hierarchy of 1D solves.
• It constructs non-tensorial interpolation grids with Leja-ordered Chebyshev-Lobatto nodes to maintain sparsity while preserving approximation quality for non-periodic functions.
• The method offers lower complexity compared to FFT by leveraging downward closed index sets, making it practical for high-dimensional interpolation and surrogate modeling.

Searching arXiv for the cited Fast Newton Transform papers and closely related work. arxiv_search(query="The Fast Newton Transform: Interpolation in Downward Closed Polynomial Spaces", max_results=5) The Fast Newton Transform (FNT) is an algorithm for performing multivariate Newton interpolation in downward closed polynomial spaces. In the formulation introduced for interpolation, the ambient polynomial space is indexed by a finite downward closed set $A\subset \mathbb N_0^m$, and the transform computes Newton coefficients, or evaluates the corresponding interpolant on its interpolation grid, with time complexity $\mathcal O(|A|\,m\,\bar n)$ up to a mild shape factor $\kappa(A)$. The construction is intended for non-tensorial polynomial approximation, especially for non-periodic functions, and is presented as a complement to the Fast Fourier Transform (FFT) rather than a replacement for periodic spectral methods (Hofmann et al., 20 May 2025).

1. Downward closed polynomial spaces

The algebraic setting begins with a finite set $A\subset \mathbb N_0^m$. The set is called downward closed if, for every $\alpha\in A$ and every $\beta\in\mathbb N_0^m$ with $0\le \beta_i\le \alpha_i$ for $i=1,\dots,m$, one has $\beta\in A$. The associated polynomial space is

$$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$

and its dimension is $\mathcal O(|A|\,m\,\bar n)$0 (Hofmann et al., 20 May 2025).

If

$\mathcal O(|A|\,m\,\bar n)$1

then the maximal degree in coordinate $\mathcal O(|A|\,m\,\bar n)$2 is $\mathcal O(|A|\,m\,\bar n)$3, the mean degree is

$\mathcal O(|A|\,m\,\bar n)$4

and the overall maximum degree is $\mathcal O(|A|\,m\,\bar n)$5. These quantities enter directly into the transform complexity.

A central family of examples is given by the $\mathcal O(|A|\,m\,\bar n)$6-sets

$\mathcal O(|A|\,m\,\bar n)$7

with

$\mathcal O(|A|\,m\,\bar n)$8

and $\mathcal O(|A|\,m\,\bar n)$9 for $\kappa(A)$0. In this terminology, $\kappa(A)$1 gives total degree, $\kappa(A)$2 Euclidean degree, and $\kappa(A)$3 max-degree. The 2025 analysis of interpolation in polynomial spaces of $\kappa(A)$4-degree extends the FNT framework to the spaces $\kappa(A)$5 induced by these sets and retains the same structural reliance on downward closedness (Hofmann et al., 18 Jul 2025).

The significance of this formulation is that downward closedness encodes a sparse multivariate polynomial model without abandoning hierarchical one-dimensional structure. This suggests a route to multivariate interpolation that is neither full tensor-product nor purely combinatorial: the index set remains sparse, but the transform still decomposes into coordinatewise Newton steps.

2. Newton basis, nodes, and non-tensorial interpolation grids

For each coordinate $\kappa(A)$6, one chooses one-dimensional nodes

$\kappa(A)$7

for example Chebyshev-Lobatto nodes. The interpolation grid associated with $\kappa(A)$8 is the non-tensorial set

$\kappa(A)$9

Thus the geometry of the node set is inherited from $A\subset \mathbb N_0^m$0: even though the coordinate nodes are one-dimensional, only those multi-indices in the downward closed set are realized (Hofmann et al., 20 May 2025).

The multivariate Newton basis, written in co-lex order on $A\subset \mathbb N_0^m$1, is

$A\subset \mathbb N_0^m$2

For Chebyshev-Lobatto nodes,

$A\subset \mathbb N_0^m$3

The same framework also accommodates Leja ordering. Given a candidate set $A\subset \mathbb N_0^m$4, Leja ordering picks $A\subset \mathbb N_0^m$5, then

$A\subset \mathbb N_0^m$6

According to the source description, this spreads out the nodes and controls the Lebesgue constant (Hofmann et al., 20 May 2025).

The interpolation-theoretic role of these node choices is twofold. First, they provide a nested or at least hierarchically usable one-dimensional structure for Newton interpolation. Second, when combined with downward closed $A\subset \mathbb N_0^m$7-sets, they generate non-tensorial grids that preserve approximation quality while reducing the number of points relative to a full tensor grid.

3. Hierarchical transform construction

The FNT computes Newton coefficients by successive one-dimensional Newton solves organized around the combinatorics of the index set. In the detailed exposition, the data setup performed once per $A\subset \mathbb N_0^m$8 consists of three components: precomputing one-dimensional Vandermonde factors

$A\subset \mathbb N_0^m$9

and their LU decompositions; computing the “tube,” “sub-tube” and “volume” projections of $\alpha\in A$0; and computing ordinal embedding matrices $\alpha\in A$1, described as croppings from the full box $\alpha\in A$2 (Hofmann et al., 20 May 2025).

Coefficient extraction is then cast as solving $\alpha\in A$3. The vectors $\alpha\in A$4 and $\alpha\in A$5 are organized blockwise according to the volume projection of $\alpha\in A$6 in the last coordinate. For $\alpha\in A$7, the algorithm processes each block by solving the lower-triangular system associated with $\alpha\in A$8 against the current block representation via the “lower $\alpha\in A$9-block product.” After the final pass, the output is the vector of Newton coefficients. Recursive evaluation, mapping coefficients back to sampled values $\beta\in\mathbb N_0^m$0, uses exactly the same scheme but applies $\beta\in\mathbb N_0^m$1 blockwise by forward substitution (Hofmann et al., 20 May 2025).

The related analysis of $\beta\in\mathbb N_0^m$2 emphasizes the same hierarchy in a slightly different language. At stage $\beta\in\mathbb N_0^m$3, the current vector is partitioned into blocks according to the value of $\beta\in\mathbb N_0^m$4, one applies the same univariate Newton step in direction $\beta\in\mathbb N_0^m$5 to each block, and then blocks are merged in the pattern dictated by downward closedness. Two specific utilities are highlighted: tube projections $\beta\in\mathbb N_0^m$6, which determine how blocks grow and merge, and ordinal embeddings or selection matrices $\beta\in\mathbb N_0^m$7, which extract subvectors as partial identity matrices. Theorem 3.6 in that presentation states that all $\beta\in\mathbb N_0^m$8 can be computed from $\beta\in\mathbb N_0^m$9 in linear time $0\le \beta_i\le \alpha_i$0 via a “length$0\le \beta_i\le \alpha_i$1sum” reduction (Hofmann et al., 18 Jul 2025).

The algorithmic importance of this hierarchy is that it avoids forming or inverting a full multivariate Vandermonde system. Instead, it reuses one-dimensional Newton solves across a structured sparse multi-index set.

4. Complexity, sparsity, and $0\le \beta_i\le \alpha_i$2-set geometry

The stated complexity of the transform is

$0\le \beta_i\le \alpha_i$3

and, ignoring the mild shape factor $0\le \beta_i\le \alpha_i$4, this becomes $0\le \beta_i\le \alpha_i$5. The reason given is that downward closedness induces sparsity in the multi-index set and allows hierarchical block updates rather than full $0\le \beta_i\le \alpha_i$6 operations (Hofmann et al., 20 May 2025).

For the specific family $0\le \beta_i\le \alpha_i$7, the later analysis states the complexity as

$0\le \beta_i\le \alpha_i$8

which is the same scaling when all coordinatewise maximal degrees are bounded by a common $0\le \beta_i\le \alpha_i$9. That work introduces the density ratio

$i=1,\dots,m$0

so that the FNT cost can be read as a tensor-product cost multiplied by $i=1,\dots,m$1. It further states that, for fixed $i=1,\dots,m$2, $i=1,\dots,m$3 decays super-exponentially with increasing spatial dimension when $i=1,\dots,m$4 (Hofmann et al., 18 Jul 2025).

The combinatorial contrast among set families is explicit. The full tensor grid $i=1,\dots,m$5 has $i=1,\dots,m$6. The total-degree set $i=1,\dots,m$7 has $i=1,\dots,m$8. For the Euclidean-degree set $i=1,\dots,m$9,

$\beta\in A$0

and the ratio satisfies

$\beta\in A$1

hence is exponentially small in $\beta\in A$2 in the sense emphasized by the source. The abstract formulation of the original paper summarizes this point by stating that the $\beta\in A$3-Euclidean case is the pivotal choice, leading to

$\beta\in A$4

(Hofmann et al., 20 May 2025).

Within the provided material, $\beta\in A$5-sets are therefore singled out as the most consequential isotropic choice. A plausible implication is that the transform’s practical efficiency depends less on the nominal polynomial degree than on whether the chosen multi-index geometry preserves isotropy while suppressing tensor growth.

5. Approximation rates and comparison with the FFT

For functions analytic in an open Bernstein poly-ellipse, the FNT is tied to geometric approximation rates. The relevant domain is

$\beta\in A$6

If $\beta\in A$7 is analytic in this poly-ellipse, then the best approximation error by $\beta\in A$8 decays like $\beta\in A$9. Interpolation on Leja-ordered Chebyshev-Lobatto or pure Leja grids inherits the same rate up to mild powers (Hofmann et al., 20 May 2025).

This approximation result is the analytic counterpart to the sparsity discussion. The method is not merely a fast solver for a reduced basis; it is paired with index sets for which geometric convergence can still be asserted under analyticity assumptions.

The comparison with the FFT is formulated at the level of non-periodic multivariate approximation. The FFT on a full tensor grid in $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$0 dimensions has complexity

$$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$1

whereas the FNT has complexity $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$2. The abstract states that, when $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$3 is chosen as an $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$4-set, the FNT time complexity is less than the FFT time complexity in a range of $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$5 behaving as $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$6. The same comparison is carried through for differentiation: differentiating via FNT costs $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$7, whereas spectral differentiation via FFT has the same complexity as the forward and backward transforms, $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$8. In the source phrasing, the FNT advantage persists for computing $$\Pi_A=\operatorname{span}\{x^\alpha=x_1^{\alpha_1}\cdots x_m^{\alpha_m}:\alpha\in A\},$$9 (Hofmann et al., 20 May 2025).

The intended distinction is therefore methodological rather than adversarial. FFT-based tensor-product schemes remain natural for periodic settings, while the FNT is presented as the corresponding fast transform for sparse, downward-closed, non-periodic polynomial expansions.

6. Applications, recommended practice, and terminological ambiguity

The stated applications include high-dimensional non-periodic interpolation, with $\mathcal O(|A|\,m\,\bar n)$00 up to $\mathcal O(|A|\,m\,\bar n)$01–$\mathcal O(|A|\,m\,\bar n)$02 in practice; surrogate modeling and uncertainty quantification in parametric PDEs; and spectral methods for boundary-value problems in $\mathcal O(|A|\,m\,\bar n)$03–$\mathcal O(|A|\,m\,\bar n)$04 dimensions. The recommended practical choices are also explicit: use downward-closed $\mathcal O(|A|\,m\,\bar n)$05-sets $\mathcal O(|A|\,m\,\bar n)$06, employ Leja-ordered Chebyshev-Lobatto nodes in each axis, precompute and cache one-dimensional LU decompositions and ordinal embeddings, and use the open-source Python package lpFun, described as Numba-accelerated for “C-level” speed (Hofmann et al., 20 May 2025).

A more specialized downstream application appears in global sensitivity analysis. There, one forms the gradient covariance matrix

$\mathcal O(|A|\,m\,\bar n)$07

computes eigenpairs $\mathcal O(|A|\,m\,\bar n)$08, and defines activity scores by

$\mathcal O(|A|\,m\,\bar n)$09

Once $\mathcal O(|A|\,m\,\bar n)$10 is approximated by a Newton polynomial, differentiation is done termwise, and Gaussian-Legendre or tensorized Gauss quadrature evaluates the required integrals exactly up to the total degree of the interpolant. The cited numerical benchmarks are a 5D solar cell, a 6D OTL circuit, and a 7D piston model; these are reported to show that FNT-based polynomial surrogates of moderate degree $\mathcal O(|A|\,m\,\bar n)$11–$\mathcal O(|A|\,m\,\bar n)$12 achieve uniform RMSE $\mathcal O(|A|\,m\,\bar n)$13 with $\mathcal O(|A|\,m\,\bar n)$14 ranging from $\mathcal O(|A|\,m\,\bar n)$15 to $\mathcal O(|A|\,m\,\bar n)$16 points, and that the resulting activity scores match Monte Carlo $\mathcal O(|A|\,m\,\bar n)$17 on $\mathcal O(|A|\,m\,\bar n)$18–$\mathcal O(|A|\,m\,\bar n)$19 samples while correctly ranking parameter importance (Hofmann et al., 18 Jul 2025).

The acronym “FNT” is, however, not unique across the literature. In coding theory, it denotes the Fermat Number Transform, a discrete Fourier-type transform over $\mathcal O(|A|\,m\,\bar n)$20, used for Reed-Solomon erasure codes with complexity $\mathcal O(|A|\,m\,\bar n)$21 (0907.1788). In a distinct 3D alignment context, a 2026 report describes a “Fast Newton Transform” as a frequency-marched Newton method on $\mathcal O(|A|\,m\,\bar n)$22, combined with alternating FFT-based translation updates, for subtomogram alignment (Kruse et al., 16 Mar 2026). This makes context essential: in the interpolation literature of 2025, “Fast Newton Transform” refers specifically to the hierarchical transform for downward closed polynomial spaces introduced by Hofmann, Wicaksono, and Hecht (Hofmann et al., 18 Jul 2025).