---
title: Jansson Derivative Method Overview
url: https://www.emergentmind.com/topics/jansson-derivative-method
type: topic
---

# Jansson Derivative Method Overview

The Jansson derivative method, in the literature summarized here, denotes a singular-value-decomposition-based procedure for numerical differentiation from noisy discrete data. In its classical form, it differentiates noisy data by inverting the integration operator via its SVD and truncating the inversion at a finite index to control noise amplification. In the formulation developed in "Closed-Form Projection Method for Regularizing a Function Defined by a Discrete Set of Noisy Data and for Estimating its Derivative and Fractional Derivative" [1805.09849], this idea becomes a closed-form, finite-dimensional projection method with measurement weighting, QR-based coefficient extraction, residual diagnostics, and an extension from first derivatives to fractional derivatives.

## 1. Terminology and scope

Within the supplied sources, the term has a primary and a secondary status. Its primary meaning is the SVD-based derivative method described explicitly as the classical “Jansson derivative method” in the context of differentiating an antiderivative or fractional integral corrupted by noise [1805.09849]. Two later papers discuss different derivative constructions and explicitly note that they do not themselves use the name.

| Context | Description | Paper |
|---|---|---|
| Noisy-data differentiation | Classical “Jansson derivative method” via SVD inversion and truncation | [1805.09849] |
| Jacobi-polynomial order derivatives | The name is not used; if encountered elsewhere, it corresponds there to the parameter-derivative/Wronskian approach | [2606.28619] |
| Derivatives of matrix functions | The paper does not use or mention the term; the method is the Mathias/Fréchet block-matrix approach | [1611.00195] |

The dominant encyclopedic meaning in the present corpus is therefore the inverse-problem method of [1805.09849]. In that setting, the problem begins with noisy samples $\{(x_i,y_i)\}$ of either the antiderivative $g$ of an unknown $f$, with $g(x)=\int_a^x f(t)\,dt$ or more generally $g=I_a^\mu f$ for $0<\mu<1$, or of the function $f$ itself. The measurement model is
$$
y_i = h(x_i) + \epsilon_i,
$$
with $E[\epsilon]=0$ and $\operatorname{Cov}(\epsilon)=S^2=\operatorname{diag}(s_1^2,\ldots,s_m^2)$ [1805.09849].

A plausible implication is that the phrase “Jansson derivative method” should be read carefully across subfields: in inverse problems it refers to TSVD differentiation of integral data, whereas in the other two cited papers the term is only discussed as absent or as a possible external naming variant.

## 2. Operator-theoretic formulation

The method is built on explicit singular systems of smoothing operators. For standard differentiation on $[0,1]$, one uses the integration operator
$$
K f(x)=\int_0^x f(t)\,dt,
$$
with $g(0)=0$ and the singular system
$$
\sigma_j=\frac{1}{\pi c_j},\qquad
v_j(x)=\sqrt{2}\cos(c_j\pi x),\qquad
u_j(x)=\sqrt{2}\sin(c_j\pi x),
$$
$$
c_j=j-\frac12,\qquad j=1,2,\ldots
$$
satisfying $K v_j=\sigma_j u_j$, $K^*u_j=\sigma_j v_j$, orthonormality in $L^2[0,1]$, and $\sigma_j\sim j^{-1}$ [1805.09849].

For fractional differentiation on $[-1,1]$, the relevant operator is the left-sided Riemann–Liouville fractional integral
$$
I_a^\mu f(x)=\frac{1}{\Gamma(\mu)}\int_a^x (x-y)^{\mu-1}f(y)\,dy,\qquad 0<\mu<1.
$$
On $\Omega=[-1,1]$, with $w_1(x)=1$ and $w_2(x)=(1-x^2)^{-\mu}$, a closed-form SVD is
$$
\sigma_j=\sqrt{\frac{\Gamma(j-\frac12-\mu)}{\Gamma(j-\frac12+\mu)}}\sim j^{-\mu},
$$
$$
v_j(x)=\sqrt{2j-1}\,P_{j-1}(x),
$$
$$
u_j(x)=c_j(1+x)^\mu P_{j-1}^{(-\mu,\mu)}(x),
$$
$$
c_j=\sqrt{\frac{(j-\frac12)[\Gamma(j)]^2}{\Gamma(j-\mu)\Gamma(j+\mu)}}.
$$
Here the $v_j$ form an orthonormal basis for $L^2(\Omega,w_1)$ and the $u_j$ are orthonormal in $L^2(\Omega,w_2)$, with $I_{-1}^\mu v_j=\sigma_j u_j$ and $(I_{-1}^\mu)^*u_j=\sigma_j v_j$ [1805.09849].

These singular systems encode the ill-posedness directly. The ideal inversion
$$
f=\sum_{j=1}^\infty \frac{(g,u_j)}{\sigma_j}v_j
$$
is unstable because $\sigma_j\downarrow 0$. Jansson’s idea is therefore to replace the divergent inversion by the truncated-SVD estimate
$$
f_{\mathrm{TSVD}}(x)=\sum_{j=1}^{k_*}\frac{(g,u_j)}{\sigma_j}v_j,
$$
with a truncation index chosen from discrepancy or related criteria [1805.09849].

## 3. Closed-form projection and coefficient recovery

The 2018 formulation converts this operator-theoretic inversion into a finite-dimensional procedure on discrete noisy data. The data are first scaled by $S^{-1}$ to account for heteroscedasticity. Writing $b=S^{-1}g$ when an antiderivative is measured, or $b=S^{-1}f$ when the measured object is the function itself, the scaled data satisfy a white-noise model in which residuals can be interpreted statistically [1805.09849].

Given sample nodes $x_i$, one assembles the range-basis design matrix
$$
P_{ij}=u_j(x_i),\qquad i,j=1,\ldots,m.
$$
The algorithm then uses
$$
P=QR,
$$
with $Q$ orthogonal and $R$ upper-triangular, followed by the rotated, scaled coefficients
$$
a=Q^Tb.
$$
Because $Q$ is orthogonal, the noise in $a$ remains white with unit variance, and the components $a_k$ can be interpreted as orthogonalized projections of the scaled data onto increasingly oscillatory basis functions [1805.09849].

Signal extraction is performed by thresholding. One starts with $\tau=3$ and defines
$$
k_{\max}=\max\{k:\ |a_k|>\tau\}.
$$
Then
$$
a_S=(a_1,\ldots,a_{k_{\max}},0,\ldots,0)^T,\qquad a_N=a-a_S,
$$
and
$$
b_S=Qa_S.
$$
From $b_S$ and hence $g_S=S b_S$, one solves
$$
\xi=P^{-1}g_S=R^{-1}Q^T g_S.
$$
This yields a smooth closed-form approximation in the range space,
$$
G(x)=\sum_{j=1}^{k_{\max}}\xi_j u_j(x).
$$
The source estimate is then obtained by SVD inversion:
$$
f_m(x)=\sum_{j=1}^{k_{\max}}\frac{\xi_j}{\sigma_j}v_j(x).
$$
For $g=Kf$ on $[0,1]$, this becomes
$$
f_m(x)=\sqrt{2}\pi\sum_{j=1}^{k_{\max}}\xi_j c_j\cos(c_j\pi x),\qquad c_j=j-\frac12.
$$
For $g=I_{-1}^\mu f$ on $[-1,1]$, it becomes
$$
f_m(x)=\sum_{j=1}^{k_{\max}}\frac{\xi_j}{\sigma_j}\sqrt{2j-1}\,P_{j-1}(x).
$$
These are exactly truncated-SVD inversions with hard spectral cut-off [1805.09849].

## 4. Regularization, diagnostics, and error structure

A defining feature of the closed-form implementation is that truncation is not treated as a purely formal spectral step. It is tied to residual diagnostics after the $Q^Tb$ rotation. The paper proposes four diagnostics: a Morozov-type discrepancy criterion, normality of residuals by histogram and chi-square goodness-of-fit, a cumulative periodogram within $95\%$ Kolmogorov–Smirnov confidence bands, and heuristic exclusion of isolated high-frequency coefficients that pass the threshold but degrade the estimate [1805.09849].

The discrepancy principle is stated in the whitened variables as
$$
m-2\sqrt{2m}\le \|b-Qa_S\|_2^2 \le m+2\sqrt{2m}.
$$
Because the residual should behave like white noise under the model, the normality and periodogram checks supply an operational criterion for whether the retained modes have captured the signal while leaving the noise in the residual [1805.09849].

The regularization is equivalent to TSVD on the underlying operator. In the explicit singular systems, the filter factors are
$$
\phi_j=
\begin{cases}
1/\sigma_j, & j\le k_{\max},\\
0, & j>k_{\max}.
\end{cases}
$$
The paper contrasts this with a Tikhonov filter, for which the factor is described as
$$
\phi_j=\frac{\sigma_j}{\sigma_j^2+\lambda}\cdot \sigma_j^{-1}
= \frac{1}{\sigma_j^2+\lambda}\text{ times }\sigma_j.
$$
This distinction is central: the Jansson construction uses a hard cut-off rather than continuous damping [1805.09849].

From the SVD viewpoint, if $A$ denotes $K$ or $I_a^\mu$ and $g^\delta=g+\eta$ with $\|\eta\|\approx\delta$, the TSVD error satisfies
$$
\|f_k-f\|_X^2
=
\sum_{j>k}|(f,v_j)|^2
+
\sum_{j=1}^k \sigma_j^{-2}|(\eta,u_j)|^2.
$$
If $\sigma_j\sim j^{-\alpha}$ and $|(f,v_j)|\lesssim j^{-(s+1/2)}$, then the bias decays like $k^{-2s}$ and the variance grows like $\delta^2\sum_{j\le k}j^{2\alpha}$. Balancing yields
$$
k\sim \delta^{-1/(s+\alpha+1/2)},
\qquad
\|f_k-f\|\lesssim \delta^{2s/(2s+2\alpha+1)}.
$$
The paper interprets this as showing that higher $\mu$ makes inversion more ill-posed, because the Abel singular values decay as $j^{-\mu}$ [1805.09849].

## 5. Direct differentiation and fractional generalization

The method is not restricted to recovering $f$ from antiderivative data. When the measured quantity is $f$ itself, the paper proposes a decoupled route: regularize $f$ by a low-degree orthogonal expansion and then differentiate or apply fractional operators analytically [1805.09849].

For this mode, one chooses $\phi_k(x)$ as Legendre or Jacobi polynomials on $[-1,1]$, or sines and cosines on $[0,1]$ or periodic domains, and computes
$$
f_m(x)=\sum_{k=0}^m c_k\phi_k(x),
$$
with coefficients from weighted least squares or interpolation on the signal subspace determined by the thresholded coefficients. The derivative is then
$$
f_m'(x)=\sum_{k=0}^m c_k\phi_k'(x).
$$
For Jacobi polynomials, the identity used is
$$
\frac{d}{dx}P_n^{(a,b)}(x)=\frac{n+a+b+1}{2}\,P_{n-1}^{(a+1,b+1)}(x).
$$
In the trigonometric SVD basis on $[0,1]$, the derivative identities are
$$
\frac{d}{dx}\big[\sqrt{2}\sin(c_j\pi x)\big]=\sqrt{2}c_j\pi\cos(c_j\pi x),
$$
$$
\frac{d}{dx}\big[\sqrt{2}\cos(c_j\pi x)\big]=-\sqrt{2}c_j\pi\sin(c_j\pi x),
$$
so that if $G(x)=\sum \xi_j u_j(x)$ then the inversion formula produces
$$
f(x)=\sqrt{2}\pi\sum \xi_j c_j\cos(c_j\pi x)
$$
with $c_j=j-\frac12$ [1805.09849].

For fractional derivatives with $\alpha\in(0,1)$, the paper gives two routes. In the inversion route, one interprets the measurement as $g=I_a^\alpha f$ and computes
$$
f_m(x)=\sum_{j\le k_{\max}} (g,u_j)_Y\,\sigma_j^{-1}v_j(x),
$$
or its collocation/QR version. In the decoupled route, one expands $f$ in a Jacobi basis and applies known fractional operators termwise whenever closed-form RL or Caputo actions are available. The paper states that, in its examples, the inversion route with the Abel SVD was emphasized [1805.09849].

This generalization is described explicitly as a principled extension of Jansson’s method from $\alpha=1$ to $\alpha\in(0,1)$. The central change is to replace the ordinary integration operator $K$ by the fractional operator $I_a^\mu$ and then use its closed-form SVD to estimate $D_a^\mu g$ by TSVD [1805.09849].

## 6. Practical behavior, limitations, and distinct later usages

The experimental summaries in [1805.09849] characterize the method as producing short closed-form sums for both the regularized data and the recovered source. In the Craig–Brown example on $[0,1]$, the method retained $3$–$5$ low-frequency modes out of $m=250$; the residual energy satisfied the discrepancy bounds, and the cumulative periodogram fell within $95\%$ white-noise bands. In a polynomial example where $f$ itself was measured, a $3$-term Legendre expansion sufficed to recover $f$ and $f'$ accurately. In the fractional example with $\mu=1/2$ and $g=I_{-1}^{1/2}f$ on $[-1,1]$, only $3$ modes were needed to recover $f$ to graphical accuracy, while QR stabilized Jacobi collocation up to approximately $90$ columns, with the chosen $k_{\max}$ remaining well below the instability threshold [1805.09849].

The practical considerations stated in the paper are closely tied to the bases and operators. For $K$ on $[0,1]$, $u_j(0)=0$ aligns with $g(0)=0$, while $v_j(1)=0$ implies possible edge effects at $x=1$. For $I_{-1}^\mu$, the boundary condition $g(-1)=0$ is required, and the $u_j$ are not differentiable at $x=-1$ when $\mu>0$, matching the Abel kernel’s endpoint behavior. The nodes need not be uniform, because the matrix $P$ is evaluated directly at the data sites; orthogonality is continuous, but QR stabilizes the discrete system. The paper also states that the $S^{-1}$ scaling is essential for heteroscedasticity and that robust variants such as Huber weights could be incorporated [1805.09849].

The listed limitations include conditioning of $P$ for very high degrees, especially in Jacobi bases; possible boundary artifacts when the basis enforces endpoint zeros; sensitivity of $\tau$ and hence $k_{\max}$; and vulnerability to outliers or non-Gaussian noise. The extensions named in the paper are adaptive or localized bases, Bayesian priors on coefficients, multidimensional generalizations using separable polynomial or trigonometric bases, and Tikhonov or GSVD filters in place of hard truncation [1805.09849].

The term’s later appearances in the supplied corpus point to different mathematical settings. "Derivative of the Jacobi polynomials with respect to their order and applications to indefinite integration" [2606.28619] derives explicit formulas for $\partial P_\mu^{(\alpha,\beta)}/\partial\mu$ and uses a Wronskian integral formula based on parameter derivatives; it states that the name “Jansson Derivative Method” is not used in the paper and that, if encountered elsewhere, it corresponds there to the same parameter-derivative/Wronskian approach. "A method to compute derivatives of functions of large complex matrices" [1611.00195] presents a block-matrix Fréchet-derivative method due to Mathias, combined with Krylov approximations and generalized deflation, and likewise states that it does not use or mention the term. This suggests that the stable encyclopedic identification of the Jansson derivative method remains the TSVD differentiation framework of [1805.09849], while other uses are naming variants external to those papers rather than established internal terminology.

Source: https://www.emergentmind.com/topics/jansson-derivative-method