Skip to content

Latest commit

 

History

History
66 lines (52 loc) · 5.93 KB

paper.md

File metadata and controls

66 lines (52 loc) · 5.93 KB
title tags authors affiliations date bibliography
PyNumDiff: A Python package for numerical differentiation of noisy time-series data
Python
numerical differentiation
denoising
dynamics
time series
machine learning
name affiliation
Floris Van Breugel^[corresponding author]
1
name affiliation
Yuying Liu
2
name affiliation
Bingni W. Brunton
3
name affiliation
J. Nathan Kutz
2
name index
Department of Mechanical Engineering, University of Nevada at Reno
1
name index
Department of Applied Mathematics, University of Washington
2
name index
Department of Biology, University of Washington
3
10 July 2021
paper.bib

Statement of need

The numerical computation of derivatives is ubiquitous in every scientific discipline and engineering application because derivatives express fundamental relationships among many quantities of interest. As a result, a large number of diverse algorithms have been developed to differentiate numerical data. These efforts are challenging because, in reality, practitioners often have sparse and noisy measurements and data, which undermine the ability to estimate accurate derivatives. Among the diversity of mathematical approaches that have been formulated, many are ad hoc in nature and require significant bespoke tuning of multiple parameters to produce reasonable results. Thus, at a practical level, it is often unclear which method should be used, how to choose parameters, and how to compare results from different methods.

Regardless of application domain, scientists of various levels of mathematical expertise would benefit from a unified toolbox for differentiation techniques and parameter tuning. To address these needs, we built the open-source package PyNumDiff, with two primary goals in mind: (1) to develop a unified source for a diversity of differentiation methods using a common API, and (2) to provide an objective approach for choosing optimal parameters with a single universal hyperparameter (gamma) that functions similarly for all differentiation methods [@van2020numerical]. By filling these needs, PyNumdiff facilitates easy computations of derivatives on diverse time-series data sets.

State of the field

Currently, practitioners in need of numerical differentiation tools must often implement a number of methods themselves, before selecting one that is appropriate for their application. High-quality data can leverage computationally efficient and algorithmically simple methods such as the finite-difference, as implemented by standard packages such as NumPy [@harris2020array], SciPy [@scipy; @Virtanen2020], or specialized packages like findiff [@findiff]. Data that are sparse and noisy, however, require more sophisticated algorithms that pracitioners must build themselves based on routines implemented across modules found in disparate packages such as SciPy, PyKalman [@pykalman], PyDMD [@demo18pydmd], or stand alone scripts such as these implementations of total variation regularization [@rudin1992nonlinear; @chartrand2011numerical]. At present, there is no centralized repository that offers a diverse range of vetted numerical differentiation tools under a unified API in Python, or other software languages.

Summary

PyNumDiff is a Python package that implements methods for computing numerical derivatives of noisy data. In this package, we implement four commonly used families of differentiation methods whose mathematical formulations have different underlying assumptions, including both global and local methods [@ahnert2007numerical]. The first family of methods usually start by applying a smoothing filter to the data, followed by a finite difference calculation [@butterworth1930theory]. The second family relies on building a local model of the data through linear regression, and then analytically calculating the derivative based on the model [@belytschko1996meshless; @schafer2011savitzky; @savitzky1964smoothing]. The third family we consider is the Kalman filter [@kalman1960new; @henderson2010fundamentals; @aravkin2017generalized; @crassidis2004optimal], with unknown noise and process characteristics. The last family is an optimization approach based on total variation regularization (TVR) method [@rudin1992nonlinear; @chartrand2011numerical]. For more technical details, refer to @van2020numerical. Individual methods under each family are accessed through the API as pynumdiff.family.method.

Applying PyNumDiff usually takes three steps: (i) pick a differentiation method, (ii) obtain optimized parameters, and (iii) apply the differentiation. Step (ii) can be skipped if one wants to manually assign the parameters, which is recommended when computation time is limited and the timeseries is long. Alternatively for long timeseries, optimal parameters can be chosen using a short but representative subset of the data. This optimization routine is provided as a sub-module (pynumdiff.optimize) with the same structure of differentiation families (i.e. pynumdiff.optimize.family.method). By default, the package performs the optimization using the open source CVXOPT package. Faster solutions can be achieved by using proprietary solvers such as MOSEK.

The software package includes tutorials in the form of Jupyter notebooks. These tutorials demonstrate the usage of the aforementioned features. For more detailed information, there is a more comprehensive Sphinx documentation associated with the repository.

Acknowledgements

The work of J. Nathan Kutz was supported by the Air Force Office of Scientific Research under Grant FA9550-19-1-0011 and FA9550-19-1-0386. The work of F. van Breugel was supported by NIH grant P20GM103650, Air Force Research Lab award FA8651-20-1-0002 Airforce Office of Scientific Research FA9550-21-0122. BWB acknowledges support from the Air Force Office of Scientific Research award FA9550-19-1-0386.

References