DerivKit: stable numerical derivatives bridging Fisher forecasts and MCMC

DerivKit is a Python package that provides stable numerical differentiation tools to bridge the gap between fast Fisher-matrix forecasting and computationally intensive MCMC sampling by enabling higher-order likelihood approximations.

The Gist

Imagine you are trying to predict how a complex machine—like a jet engine or a massive weather system—will behave if you turn one of its dials just a tiny bit.

In science, this is called finding a "derivative." It’s essentially asking: "If I nudge this knob by 1%, how much will the output change?"

The problem is that in many high-level sciences (like cosmology or climate science), the "machine" is actually a massive, messy computer simulation. These simulations are often "black boxes"—you can turn the dials and see the result, but you can't see the gears inside to calculate the math perfectly. Even worse, these simulations can be "noisy," meaning if you turn the dial twice, you might get slightly different results just because of digital static.

This paper introduces DerivKit, a new software tool designed to solve this "noisy knob" problem.

---

The Problem: The "Shaky Hand" Dilemma

Imagine you are trying to measure the slope of a hill in the dark using only a flashlight.

1. The Standard Way (Finite Differences): You take a step forward, see how much your height changed, and guess the slope. But if the ground is covered in loose gravel (numerical noise), your footing slips. You might think the hill is steep when you actually just tripped on a rock. In science, this leads to wrong predictions.
2. The "Perfect" Way (Automatic Differentiation): This is like having X-ray vision that lets you see the exact geometry of the hill. It’s perfect, but it only works if you built the hill yourself. If you are studying a mountain that was already there (a "legacy" simulation or a pre-made data table), X-ray vision doesn't work.

The Solution: DerivKit (The "Smart Surveyor")

DerivKit acts like a highly skilled surveyor with advanced GPS and stabilizing gear. Instead of just taking one shaky step, it uses two clever strategies:

• The "Smooth-it-Out" Method (Polynomial Fitting): Instead of looking at just two points, DerivKit looks at a whole neighborhood of points around you. It says, "I know these individual points are a bit jumpy because of the gravel, but if I draw a smooth curve through all of them, I can see the true shape of the hill."
• The "High-Tech Extrapolation" Method: It uses mathematical "cheat codes" (like Richardson extrapolation) to cancel out the errors caused by the noise, effectively "cleaning" the data as it measures it.

Why does this matter? (The Bridge)

The paper describes DerivKit as a bridge between two worlds:

• World 1: The Fast Forecast (Fisher Forecasts): This is like looking at a quick sketch of a map. It’s incredibly fast and tells you roughly where you are going, but it assumes everything is a simple, smooth line. It’s great for a quick guess, but it fails if the terrain gets weird and curvy.
• World 2: The Deep Dive (MCMC Sampling): This is like walking every single inch of the mountain to map it perfectly. It is incredibly accurate, but it takes a massive amount of time and energy.

DerivKit builds the bridge between them. It allows scientists to take the "quick sketch" method and add "curvy" corrections to it. This makes the fast method much more accurate—approaching the quality of the "deep dive" without the massive time cost.

Summary in a Nutshell

DerivKit is a specialized toolkit for scientists. It allows them to take messy, noisy, "black box" computer models and accurately calculate how sensitive those models are to change. It turns "shaky guesses" into "stable measurements," helping us predict everything from the evolution of the universe to the complexities of our climate with much higher confidence.

Technical Summary

Technical Summary: DerivKit

Title: DerivKit: stable numerical derivatives bridging Fisher forecasts and MCMC
Authors: Nikolina Šarčević, Matthijs van der Wild, and Cynthia Trendafilova (2026)

1. Problem Statement

In scientific computing fields such as cosmology, particle physics, and climate science, researchers frequently need to calculate derivatives of model predictions with respect to physical parameters. This is essential for two primary types of statistical inference:

• Fisher Forecasting: A computationally efficient method that assumes Gaussian posteriors and relies on first-order derivatives.
• Sampling-based Inference (e.g., MCMC): A robust method for non-Gaussian posteriors that is computationally expensive and scales poorly with high dimensionality.

The authors identify a critical "gap" between these two methods. While the Derivative Approximation for Likelihoods (DALI) offers a way to bridge them by using higher-order derivatives to capture non-Gaussianity, it is difficult to implement. Existing methods struggle because:

• Automatic Differentiation (Autodiff) cannot be used on "black-box" models, legacy code, or tabulated data (lookup tables).
• Standard Finite-Difference methods are highly sensitive to numerical noise, irregular sampling, and the choice of step size, often leading to unstable or incorrect derivative estimates.

2. Methodology

The paper introduces DerivKit, a modular Python package designed to provide stable, diagnostics-driven numerical differentiation for general models. The software is organized into four specialized "kits":

• DerivativeKit (The Core Engine): Implements two main strategies for differentiation:
  • High-order Finite-Difference: Uses 3-, 5-, 7-, and 9-point stencils (up to 4th order) combined with Richardson extrapolation and Ridders’ method to improve accuracy in smooth models.
  • Polynomial Fitting: For noisy or "stiff" models, it employs local polynomial fits. This includes an adaptive method that uses domain-aware Chebyshev sampling grids, internal offset scaling, and ridge regularization to ensure stability even when data is noisy.
• CalculusKit: A utility layer that assembles raw derivatives into higher-level mathematical structures, such as gradients, Jacobians, Hessians, and higher-order derivative tensors.
• ForecastKit: A high-level inference tool that uses the calculus utilities to automate the construction of Fisher matrices, Fisher bias estimates (to account for parameter shifts due to model bias), and the derivative tensors required for DALI.
• LikelihoodKit: A lightweight module providing basic Gaussian and Poisson likelihood models for end-to-end testing and validation.

3. Key Contributions

• Robustness to Noise: Unlike standard finite-difference methods, DerivKit’s adaptive polynomial fitting can extract accurate derivatives from noisy data (as demonstrated in the paper's Figure 1).
• Implementation of DALI: It provides the first practical, general-purpose software implementation for the Derivative Approximation for Likelihoods, allowing users to move beyond simple Gaussian approximations.
• Diagnostics-Driven Approach: The package includes internal quality checks, reporting metadata on sampling geometry and fit quality, and issuing warnings when derivative estimates are unreliable.
• Versatility: It supports scalar- and vector-valued models and is compatible with tabulated data, making it applicable to emulators and legacy simulation suites where autodiff is impossible.

4. Results and Validation

The paper demonstrates the efficacy of the package through several simulated scenarios:

• Noise Resilience: Comparison plots show that while finite-difference methods produce highly inaccurate median estimates when noise is present ($\sigma = 0.2$), the adaptive-fit method remains precise and centered on the true derivative.
• Inference Accuracy: The authors demonstrate that using the ForecastKit to generate DALI-based likelihoods produces posterior contours that closely match the "ground truth" results obtained from computationally intensive MCMC sampling (e.g., using the emcee sampler).
• Workflow Integration: The package successfully automates complex tasks like calculating the "Fisher bias" (the shift in parameters caused by a biased data vector) and constructing complex derivative tensors.

5. Significance

DerivKit serves as a practical bridge between fast, local approximations (Fisher) and slow, global sampling (MCMC). By providing a stable way to compute higher-order derivatives from "black-box" or noisy models, it enables researchers to perform sophisticated non-Gaussian inference without the massive computational overhead of full MCMC or the requirement to rewrite legacy code for automatic differentiation. This has profound implications for the speed and accuracy of cosmological forecasting and sensitivity analysis in various scientific domains.