---
title: Fast Newton Transform (FNT)
url: https://www.emergentmind.com/topics/fast-newton-transform-fnt
type: topic
---

# Fast Newton Transform (FNT)

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 [2505.14909].

## 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 \(|A|\) [2505.14909].

If
\[
n_i=\max\{\alpha_i:\alpha\in A\},
\]
then the maximal degree in coordinate \(i\) is \(n_i\), the mean degree is
\[
\bar n=\frac{n_1+\cdots+n_m}{m},
\]
and the overall maximum degree is \(n=\max_i n_i\). These quantities enter directly into the transform complexity.

A central family of examples is given by the \(\ell^p\)-sets
\[
A_{m,n,p}=\{\alpha\in\mathbb N_0^m:\|\alpha\|_p\le n\},
\]
with
\[
\|\alpha\|_p=\left(\sum_i \alpha_i^p\right)^{1/p}\quad (p<\infty),
\]
and \(\max_i \alpha_i\) for \(p=\infty\). In this terminology, \(p=1\) gives total degree, \(p=2\) Euclidean degree, and \(p=\infty\) max-degree. The 2025 analysis of interpolation in polynomial spaces of \(p\)-degree extends the FNT framework to the spaces \(\Pi_{m,n,p}\) induced by these sets and retains the same structural reliance on downward closedness [2507.13640].

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 \(i=1,\dots,m\), one chooses one-dimensional nodes
\[
\xi_0^{(i)},\dots,\xi_{n_i}^{(i)}\in[-1,1],
\]
for example Chebyshev-Lobatto nodes. The interpolation grid associated with \(A\) is the non-tensorial set
\[
G=\{\xi_\alpha=(\xi_{\alpha_1}^{(1)},\dots,\xi_{\alpha_m}^{(m)}):\alpha\in A\}\subset[-1,1]^m.
\]
Thus the geometry of the node set is inherited from \(A\): even though the coordinate nodes are one-dimensional, only those multi-indices in the downward closed set are realized [2505.14909].

The multivariate Newton basis, written in co-lex order on \(A\), is
\[
Q_\alpha(x)=\prod_{i=1}^m\prod_{k=0}^{\alpha_i-1}(x_i-\xi_k^{(i)}).
\]
For Chebyshev-Lobatto nodes,
\[
\xi_k^{(i)}=\cos(k\pi/n_i),\qquad k=0,\dots,n_i.
\]
The same framework also accommodates Leja ordering. Given a candidate set \(Z\), Leja ordering picks \(\xi_1=\max_{z\in Z}|z|\), then
\[
\xi_j=\arg\max_{z\in Z\setminus\{\xi_1,\dots,\xi_{j-1}\}}\prod_{\ell<j}|z-\xi_\ell|.
\]
According to the source description, this spreads out the nodes and controls the Lebesgue constant [2505.14909].

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 \(\ell^p\)-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 \(A\) consists of three components: precomputing one-dimensional Vandermonde factors
\[
V^{(i)}_{j,k}=Q_k^{(i)}(\xi_j^{(i)})
\]
and their LU decompositions; computing the “tube,” “sub-tube” and “volume” projections of \(A\); and computing ordinal embedding matrices \(\Phi[A,R]\), described as croppings from the full box \(R=\times_{i=1}^m\{0,\dots,n_i\}\) [2505.14909].

Coefficient extraction is then cast as solving \(V\,c=f\). The vectors \(c\) and \(f\) are organized blockwise according to the volume projection of \(A\) in the last coordinate. For \(j=1,\dots,m\), the algorithm processes each block by solving the lower-triangular system associated with \(V^{(j)}\) against the current block representation via the “lower \(A\)-block product.” After the final pass, the output is the vector of Newton coefficients. Recursive evaluation, mapping coefficients back to sampled values \(q\mapsto q(\xi_\alpha)\), uses exactly the same scheme but applies \(V^{(j)}\) blockwise by forward substitution [2505.14909].

The related analysis of \(\Pi_{m,n,p}\) emphasizes the same hierarchy in a slightly different language. At stage \(i\), the current vector is partitioned into blocks according to the value of \(\alpha_i\), one applies the same univariate Newton step in direction \(i\) to each block, and then blocks are merged in the pattern dictated by downward closedness. Two specific utilities are highlighted: tube projections \(\mathcal T_i(A)\), which determine how blocks grow and merge, and ordinal embeddings or selection matrices \(\Phi\in\{0,1\}^{r\times s}\), which extract subvectors as partial identity matrices. Theorem 3.6 in that presentation states that all \(\mathcal T_i(A)\) can be computed from \(\mathcal T_1(A)\) in linear time \(O(|A|)\) via a “length\(\to\)sum” reduction [2507.13640].

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 \(\ell^p\)-set geometry

The stated complexity of the transform is
\[
\sum_{j=1}^m O\bigl((n_j+1)\,|A|\,\kappa\bigr)=O(|A|\,m\,\bar n\,\kappa),
\]
and, ignoring the mild shape factor \(\kappa(A)\), this becomes \(O(|A|\,m\,\bar n)\). The reason given is that downward closedness induces sparsity in the multi-index set and allows hierarchical block updates rather than full \(|A|^2\) operations [2505.14909].

For the specific family \(A_{m,n,p}\), the later analysis states the complexity as
\[
O\bigl(|A_{m,n,p}|\,m\,n\bigr),
\]
which is the same scaling when all coordinatewise maximal degrees are bounded by a common \(n\). That work introduces the density ratio
\[
\rho_{m,n,p}:=\frac{|A_{m,n,p}|}{(n+1)^m},
\]
so that the FNT cost can be read as a tensor-product cost multiplied by \(\rho_{m,n,p}\). It further states that, for fixed \(p\in(0,\infty)\), \(\rho_{m,n,p}\) decays super-exponentially with increasing spatial dimension when \(m\lesssim n^p\) [2507.13640].

The combinatorial contrast among set families is explicit. The full tensor grid \(A_{m,n,\infty}\) has \(|A|=(n+1)^m\). The total-degree set \(A_{m,n,1}\) has \(|A|=\binom{m+n}{n}=O(n^m/m!)\). For the Euclidean-degree set \(A_{m,n,2}\),
\[
|A|\approx \operatorname{vol}(\{\|\alpha\|_2\le n\})\simeq \frac{\pi^{m/2}}{\Gamma(m/2+1)}\,n^m,
\]
and the ratio satisfies
\[
\frac{|A|}{(n+1)^m}=O(m^{-m/2})=O(e^{-m\log\sqrt m}),
\]
hence is exponentially small in \(m\) in the sense emphasized by the source. The abstract formulation of the original paper summarizes this point by stating that the \(\ell^2\)-Euclidean case is the pivotal choice, leading to
\[
|A|/(n+1)^m\in O(e^{-m})
\]
[2505.14909].

Within the provided material, \(\ell^2\)-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
\[
\Omega_\rho^q=\left\{z\in\mathbb C^m:\left\| \bigl(\log|z_i+\sqrt{z_i^2-1}|\bigr)_i \right\|_q<\log\rho\right\},
\qquad \rho>1,\quad \frac1p+\frac1q=1.
\]
If \(f\) is analytic in this poly-ellipse, then the best approximation error by \(\Pi_{A_{m,n,p}}\) decays like \(O(\rho^{-n})\). Interpolation on Leja-ordered Chebyshev-Lobatto or pure Leja grids inherits the same rate up to mild powers [2505.14909].

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 \(m\) dimensions has complexity
\[
O((n+1)^m\,m\,\log n),
\]
whereas the FNT has complexity \(O(|A|\,m\,\bar n)\). The abstract states that, when \(A\) is chosen as an \(\ell^2\)-set, the FNT time complexity is less than the FFT time complexity in a range of \(n\) behaving as \(O(e^m)\). The same comparison is carried through for differentiation: differentiating via FNT costs \(O(|A|\,m\,\bar n)\), whereas spectral differentiation via FFT has the same complexity as the forward and backward transforms, \(O((n+1)^m\,m\,\log n)\). In the source phrasing, the FNT advantage persists for computing \(\partial_i q\) [2505.14909].

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 \(m\) up to \(6\)–\(10\) in practice; surrogate modeling and uncertainty quantification in parametric PDEs; and spectral methods for boundary-value problems in \(3\)–\(6\) dimensions. The recommended practical choices are also explicit: use downward-closed \(\ell^2\)-sets \(A_{m,n,2}\), 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 [2505.14909].

A more specialized downstream application appears in global sensitivity analysis. There, one forms the gradient covariance matrix
\[
\mathbf C=\int_{[-1,1]^m} (\nabla f(x))(\nabla f(x))^T\,\rho(x)\,dx,
\]
computes eigenpairs \((\lambda_j,w_j)\), and defines activity scores by
\[
\theta_i=\sum_{j=1}^k \lambda_j [w_j]_i^2,
\qquad k\text{ chosen so that }\lambda_k\gg \lambda_{k+1}.
\]
Once \(f\) 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 \((n\approx 8\)–\(10)\) achieve uniform RMSE \(<10^{-4}\) with \(|A|\) ranging from \(10^4\) to \(10^6\) points, and that the resulting activity scores match Monte Carlo \(\pm 1\,\sigma\) on \(10^5\)–\(10^7\) samples while correctly ranking parameter importance [2507.13640].

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 \(GF(p)\), used for Reed-Solomon erasure codes with complexity \(O(n\log n)\) [0907.1788]. In a distinct 3D alignment context, a 2026 report describes a “Fast Newton Transform” as a frequency-marched Newton method on \(\mathrm{SO}(3)\), combined with alternating FFT-based translation updates, for subtomogram alignment [2603.15285]. 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 [2507.13640].

Source: https://www.emergentmind.com/topics/fast-newton-transform-fnt