Physics-driven Comparative Analysis of Various Statistical Distance Metrics and Normalizing Functions

N. Fuad · 2026

Read Physics-driven Comparative Analysis of Various Statistical Distance Metrics and Normalizing Functions by N. Fuad (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-13512.txt.

Introduction

Fuad tests distance metrics and scaling functions using electron and photon events from decaying krypton-83, taken with an HPGe spectrometer under cryo-vacuum setup. He covers Hellinger, 1D Wasserstein, square-root Jensen-Shannon, L-inf, Kolmogorov-Smirnov, and Fisher-Rao metrics. First setting up a unit-free PoI, he makes spreads from the data and studies how steady PoI stays under sample size, binning length, and scaling functions. He also lists rules that a scaling function should have and uses them in the test.

Catalog record

CREATORN. Fuad
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18257
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.13512

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-driven Comparative Analysis of Various Statistical Distance Metrics and Normalizing Functions N. Fuad Abstract Comparison of two probability density/mass functions (PDF/PMFs) is ubiquitous in various forms of scientific analysis, including machine learning, optimization problems, and hypothesis tests. A copious amount of distance metrics have already been proposed and are regularly being used in this regard. In this document, we report a data-driven systematic comparison among a few of such metrics. The metrics considered here are Hellinger distance, Wasserstein distances (1D), J ​ S \sqrt{JS} distance, L ∞ L_{\infty} norm, Kolmogorov-Smirnov distance, and Fisher-Rao metric. We perform this comparison using electron and photon events from a decaying 83 Kr isotope, collected through an HPGe spectrometer operating under cryo-vacuum conditions. To accomplish this, first, a dimensionless Parameter of Interest (PoI) was established, then PDF/PMFs were generated from the data, and finally the stabilities of the PoI under various criteria, such as sample size, discretization length, and normalizing functions, were studied and the results were summarized. In this report, we also propose a list of properties that a normalizing function should have and utilize them in the comparison. I Introduction Necessity of a standardized measurement of dissimilarity between Probability Density/Mass Functions (PDF/PMFs) is ubiquitous across many fields of scientific analysis. A copious amount of parameters to do so have already been introduced and are regularly used 1 ; 2 ; 13 ; 14 ; 9 . Many of these parameters have also been shown to be connected to each other (e.g. Hellinger distance can be written in terms of Bhattacharya coefficient 12 ; 11 or Chernoff divergence 22 , Fisher information metric is proportional to \sqrt{} JS 4 ; 7 ; 24 ; 21 , Wasserstein-2 takes shape of \chi -distance in 1-D 23 ; 26 ). Commonly used such parameters can be broadly put into two categories - metrics and non-metrics . They are sometimes referred to as distances and divergences respectively also [e.g. 24 ]. A function d(x,y) is called a metric if it satisfies the following properties 3 - • Non-negativity : d(x,y)\geq 0 • Identity : d(x,x)=0 • Symmetry : d(x,y)=d(y,x) • Triangle inequality : d(x,y)+d(y,z)\geq d(x,z) In this analysis, we compare following metrics between two discretized PDF/PMFs, p(x) and q(x) (List I )- 1. Hellinger distance 1 H(p,q)=\sqrt{\sum\frac{1}{2}\left(\sqrt{p}-\sqrt{q}\right)^{2}} (1) 2. Wasserstein-1 distance 27 ; 5 ; 23 W_{1}(p,q)=\sum|F_{p}^{-1}-F_{q}^{-1}| (2) where, F_{p}^{-1} and F_{q}^{-1} are the quantile functions of p and q 3. Wasserstein-2 distance 26 W_{2}(p,q)=\sqrt{\sum\left(F_{p}^{-1}-F_{q}^{-1}\right)^{2}} (3) where, F_{p}^{-1} and F_{q}^{-1} are the quantile functions of p and q 4. \sqrt{} Jensen-Shannon 9 ; 4 \sqrt{JS(p,q)}=\sqrt{\sum\frac{1}{2}\left(p\log\frac{2p}{p+q}+q\log\frac{2q}{p+q}\right)} (4) 5. L_{\infty} norm/Chebyshev distance L_{\infty}(p,q)=\lim_{r\rightarrow\infty}\left(\sum|p-q|^{r}\right)^{\frac{1}{r}}=\max(|p-q|) (5) 6. Kolmogorov-Smirnov distance KS(p,q)=\max(F_{p}-F_{q}) (6) where, F_{p} and F_{q} are the CDFs of p and q 7. Fisher-Rao distance Closed-form expression of Fisher-Rao distance for a categorical distribution was used. A few other closed-form results can be found in 18 . FR(p,q)=\frac{2}{\pi}\arccos\left(\sum\sqrt{pq}\right) (7) where, p(x) and q(x) in are either PDF or PMF in the usual sense - p(x) and p_{m}(x) are PDF and PMF respectively on the parameter space \chi={\{x\}} if they satisfy following properties - • p(x)\geq 0;0\leq p_{m}(x)\leq 1\forall x\in\chi • \sum p_{m}(x)=\sum p(x)dx=1 Since p(x) is unbounded from upper side, distance measurement using p(x) can not be guaranteed to be upper bounded. But p_{m}(x) are bounded in [0,1) by construction. We utilize this notion of normalization to as an additional comparison of our proposed normalized functions. We also present another measure of goodness of normalization which is independent of this. Metrics, such as Fisher-Rao (FR), which include \arccos are bounded by domain : [0,1] ; to tackle these cases, normalizing function is applied right after summation is applied, which is essentially \sum\rightarrow n(\sum) or \max\rightarrow\ n(\max) . Here we propose a list of properties that a normalizing function should have - Definition 1 . We define a function, n(x) , to be a normalizing function if n(x) satisfies the following properties - • Bounds: \displaystyle\lim_{x\rightarrow 0}n(x)=0 and \displaystyle\lim_{x\rightarrow\infty}n(x)=1 ; \mathrm{Domain:[0,\infty)},\mathrm{Range:[0,1)} • Bijectivity: Ensures existence of an inverse function, n^{-1}(x) • Inverse bounds: \displaystyle\lim_{x\rightarrow 0}n^{-1}(x)=0 and \displaystyle\lim_{x\rightarrow 1}n^{-1}(x)=\infty ; \mathrm{Domain^{-1}:[0,1)} and \mathrm{Range^{-1}:[0,\infty)} . • Monotonicity: x_{1}>x_{2}\Rightarrow n(x_{1})>n(x_{2}) • Metric preservation: If d(x,y) is a metric, then n\circ d must also be a metric 6 Note 1. If d(x,y) is a metric, and n(x) is monotonic, then n\circ d is guaranteed to be a metric iff n(0)\geq 0 and n(x) is concave 6 ; 20 . Note 2. If n\circ d is metric preserving through concavity of n(x) , then n^{-1}(x) will necessarily be convex and n^{-1}\circ d will not be metric preserving. In this analysis, we consider the following list of functions that follow Def. 1 - • n_{1}(x)=\frac{\log(1+x)}{1+\log(1+x)} • n_{2}(x)=\frac{x}{1+x} • n_{3}(x)=1-e^{-x} • n_{4}(x)=\frac{2}{\pi}\arctan(x) Notes 1. True domains and ranges of the n(x) s are supersets of domains and ranges defined in Def. 1 , not equal; e.g. true domain of n_{1}(x) is (-1,\infty) and of n_{2}(x) is (-\infty,\infty)\setminus\{-1\} , not [0,\infty) . Figure 1: n(x)\forall x:x\in\{-5\leq\log(x)\leq 10\} . It can be observed that n_{2-4}(x) saturate much faster compared to n_{1}(x) and reaches 1 within x=\mathcal{O}(10^{2}) . So any distance metric that produces value >>10^{2} will not be distinguished by n_{2-4}(x) . In the unsaturated region, n(x) s behave distinctively until it reaches \mathcal{O}(10^{-2}) II PDF/PMF Generation The data were collected using a High Purity Ge based spectrometer exposed to an 83 Kr source. The detector used in our study is known as PPC type detector 8 ; 15 due to its geometry and type of charge collection system. Ge is a semiconductor, so when some energy deposition takes place inside the detector, it creates a number of electron-hole pairs which is dependent on deposition energy, operating temperature, geometry of the detector leading to different amount of energy loss through other mechanism such as ionization, phonon creation or recombination etc. These charges are collected at a point on the surface by applying an electric field through the detector, typically in \mathcal{O}(\mathrm{kV}) . These charges change the voltage of the circuit, which is written to disk using an ADC and converted to a parameter with dimension of energy based on the effective capacitance of the detector. The clocking resolution of our apparatus was 10 ns, meaning the above mentioned parameter was estimated every 10 ns. A set of values of this parameter around the time of energy deposition constitute our signals, s(t) , which are called waveforms 17 . Fig. 2 shows a randomly selected set of such waveforms from one of our data-taking campaigns. These data were taken at cryogenic vacuum pressure ( \mathcal{O}(\mathrm{nbar}) ) and Liquid Nitrogen temperature ( \sim 88 K). Figure 2: Examples of signals in HPGe detectors. Each signal consists of samples taken at 10 ns intervals over a 0.2 – 2~\mu s window. Signal values from well before to well after the trigger ( \mathcal{O}(10~\mu\mathrm{s}) ) are also recorded to analyze the quality of the signal. The height of the signals corresponds to the amount of energy deposited. The energy region of the daughter particles of 83 Kr fall at or below \sim 32 keV 19 . There are also photon events below this energy from 83 Kr, but due to a well known property of

Read the complete text