Quantum Harmonic Analysis and the Structure in Data: Augmentation

Monika Dörfler, Franz Luef, Henry McNulty · 2025

Read Quantum Harmonic Analysis and the Structure in Data: Augmentation by Monika Dörfler, Franz Luef, Henry McNulty (2025) 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-2509-19474.txt.

Introduction

Adding varied copies of training data is a common trick to help models learn. Monika Doerfler, Franz Luef, and Henry McNulty ask what that trick does to the main patterns inside data. They use tools from quantum harmonic analysis, a field that studies signals with operator math. Their result says the key patterns of enriched data stay smooth and unbroken in a strict math sense.

Tests on made-up data and on audio clips back the theory. Smooth patterns are easier for learning methods to find and use. The authors suggest that planned, informed enrichment beats random tweaks. Their short note links a lab habit to a firm math base. It also hints at better ways to build training sets for sound and other signals.

Catalog record

CREATORMonika Dörfler, Franz Luef, Henry McNulty
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-19474

Browse the full library · About the Tower of Babel

Full text

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

Quantum Harmonic Analysis and the Structure in Data: Augmentation Monika Dörfler, Franz Luef, Henry McNulty Abstract In this short note, we study the impact of data augmentation on the smoothness of principal components of high-dimensional datasets. Using tools from quantum harmonic analysis, we show that eigenfunctions of operators corresponding to augmented data sets lie in the modulation space M 1 ​ ( ℝ d ) M^{1}(\mathbb{R}^{d}) , guaranteeing smoothness and continuity. Numerical examples on synthetic and audio data confirm the theoretical findings. While interesting in itself, the results suggest that manifold learning and feature extraction algorithms can benefit from systematic and informed augmentation principles. 1 Introduction Expansions of functions in terms of basic building blocks, such as trigonometric functions or wavelets, are central in analysis and signal processing. In modern machine learning, features learned by deep networks often resemble such expansions, suggesting that data lie near low-dimensional structures within high-dimensional ambient spaces. Manifold learning formalizes this observation: data often concentrate on a low-dimensional manifold, allowing efficient representations. Data augmentation, a common practice in machine learning, can be understood as a way to smooth operators associated with empirical data, revealing intrinsic structure and stabilizing principal components. This paper studies the smoothing effect of augmentation on data operators, showing how convolution with augmentation measures regularizes principal components and aids feature extraction. To do so, we leverage recent insights and methods from quantum harmonic analysis (QHA), cf. [ 7 , 3 , 4 ] as well as some important results on embeddings of modulation spaces [ 6 , 1 ] and the relation between modulation space membership of symbols and operator properties [ 5 ] . 2 Preliminaries and Notation Let \mathcal{H}=L^{2}(\mathbb{R}^{d}) be the Hilbert space of square-integrable functions. For f,g\in\mathcal{H} , the rank-one operator f\otimes g is defined by (f\otimes g)(h)=\langle h,g\rangle f,\quad h\in\mathcal{H}. A bounded linear operator T:\mathcal{H}\to\mathcal{H} is trace-class if \|T\|_{\mathcal{T}}=\mathrm{tr}(|T|)<\infty, and Hilbert-Schmidt if \|T\|_{\mathrm{HS}}^{2}=\mathrm{tr}(T^{*}T)<\infty. The trace of a trace-class operator T is defined by \mathrm{tr}(T)=\sum_{n}\langle Te_{n},e_{n}\rangle, where (e_{n}) is any orthonormal basis of \mathcal{H} ; the sum converges absolutely and is independent of the choice of basis. We also consider the Feichtinger algebra M^{1}(\mathbb{R}^{d}) , consisting of functions whose short-time Fourier transform is integrable. Operators with kernels in M^{1}(\mathbb{R}^{2d}) have convenient convolution and Fourier transform properties used later. 2.1 Time-Frequency Representations Let \pi(z) denote a time-frequency shift defined by \pi(z)g(t)=e^{2\pi it\omega}g(t-x) . For \psi,\phi\in L^{2}(\mathbb{R}^{d}) we define: • Short-time Fourier transform (STFT) V_{\phi}\psi of \psi with window \phi : V_{\phi}\psi(z)=\langle\psi,\pi(z)\phi\rangle,\quad z\in\mathbb{R}^{2d}, where \pi(z) denotes the time-frequency shift. • Cross-Wigner distribution of \psi and \phi : W(\psi,\phi)(x,\omega)=\int_{\mathbb{R}^{d}}\psi\Big(x+\frac{t}{2}\Big)\overline{\phi\Big(x-\frac{t}{2}\Big)}e^{-2\pi i\omega\cdot t}\,dt. 3 Operators, Translations and Fourier Transform Operators act on functions, respectively elements of vector spaces. Thus they are potentially higher-dimensional than the objects they act on: they need to encode the latter’s structure and the structure of the interaction between the isolated objects. We can understand functions as operators by looking at their action on other functions f by means of projections: (g\otimes h)(f)=\langle f,h\rangle g. If we think of the simple case of finite dimensional operators, which can be represented as finite matrices, we immediately see, that these simple projections form very special cases of operators. In many cases, though, the complicated action of an operator on functions is again described with the help of more functions, e.g. the eigenfunctions in the case of a compact selfadjoint operator T on X : for some orthonormal set \{g_{n}\} in X we have T=\sum_{n}s_{n}(T)g_{n}\otimes g_{n}. Now one of the reasons of the importance of Fourier transforms (and their relatives) in signal processing and other applications is the fact, that they diagonalize time-invariant systems. Systems, hence, which are invariant under translation. Can there be similar helpful connections for operators? What is the appropriate concept of translation, and of Fourier transform? In order to understand the corresponding concepts for operators, first introduced by Werner [ 7 ] , let us consider the properties of translation in the context of harmonic analysis. Indeed, Fourier transform turns translation into multiplication by a complex number with absolute value 1 . Also, the complex exponential functions e_{\omega_{0}}(x)=e^{2\pi ix\omega_{0}} , for some fixed frequency \omega_{0} , are functions on \mathbb{R} . So, in the same spirit, we look for operators, whose inner products with operators turn translations of operators into multiplications with complex numbers. It turns out, that the following translation of operators provides the desired properties if paired with an equally appropriate Fourier transform for operators. 3.1 Translation of Operators For z\in\mathbb{R}^{2d} and S a trace-class operator, define the operator translation \alpha_{z}(S)=\pi(z)S\pi(z)^{*}, where \pi(z) is the time-frequency shift defined by \pi(z)g(t)=e^{2\pi it\omega}g(t-x) . Since we are interested in defining a convolution for operators, which features properties similar to the classical convolution of functions in light of the Fourier transform, we must first find an equivalent for the integral of functions. Consider the defining properties of operator trace \operatorname{tr}(T) : for two trace class operators T,T^{\prime} , and r c\in\mathbb{C} it holds that • \operatorname{tr}(T+T^{\prime})=\operatorname{tr}(T)+\operatorname{tr}(T^{\prime}) • \operatorname{tr}(cT)=c\operatorname{tr}(T) • \operatorname{tr}(TT^{\prime})=\operatorname{tr}(TT^{\prime}) . In light of these connections, the following substitutions in the definition of convolutions, first proposed by Werner [ 7 ] , seem intuitive: FUNCTIONS OPERATORS real functions selfadjoint operators integral trace translations automorphism \alpha_{z} L^{p} -spaces Schatten classes \mathcal{S}^{p} continuous functions vanishing at infinity compact operators We will now introduce an analogue of the Fourier transform for a trace class operator S . 3.2 Fourier-Wigner and Weyl Transform The Fourier-Wigner transform of a trace-class operator S is \mathcal{F}_{W}(S)(z)=e^{-\pi ix\cdot\omega}\mathrm{tr}(\pi(-z)S),\quad z=(x,\omega)\in\mathbb{R}^{2d}. It satisfies \mathcal{F}_{\sigma}(S\star T)=\mathcal{F}_{W}(S)\mathcal{F}_{W}(T), linking operator convolution and Fourier analysis. Example 1 . Let S=\varphi_{2}\otimes\varphi_{1} with \varphi_{1},\varphi_{2}\in L^{2}(\mathbb{R}^{d}) . Then the Fourier-Wigner transform of S is given by \mathcal{F}_{W}(\varphi_{2}\otimes\varphi_{1})(z)=A(\varphi_{2},\varphi_{1})(z), where A(\varphi_{2},\varphi_{1})(z) is the cross-ambiguity function . In particular, for the Gaussian \varphi(t)=2^{d/4}e^{-\pi t\cdot t},\qquad t\in\mathbb{R}^{d}, and the operator S=\varphi\otimes\varphi . Then its Fourier-Wigner transform satisfies \mathcal{F}_{W}\,S(z)=e^{\pi ix\cdot\omega}V_{\varphi}\varphi(z), where V_{\varphi}\varphi is the short-time Fourier transform of \varphi with itself. Explicitly, we find \mathcal{F}_{W}(\varphi\otimes\varphi)(z)=e^{2\pi ix\cdot\omega}\,e^{-\frac{\pi}{2}z\cdot z},\qquad z\in\mathbb{R}^{2d}. Using the cross-Wigner distribution, we may introduce the Weyl calculus: For \sigma\in S_{0}(\mathbb{R}^{2d}) and \psi,\phi\in S(\mathbb{R}^{d}) , the Weyl transform L_{\sigma} of \sigma is the operator defined by \langle L_{\sigma}\psi,\phi\rangle_{S_{0},S}=\langle\sigma,W(\psi,\phi)\rangle_{S_{0},S}. Here, \sigma is called the Weyl symbol of the operator L_{\sigma} . Remark 1 . Let \mathcal{F}_{s} denote the symplectic Fourier transform \displaystyle\mathcal{F}_{s}f(z):=\int_{\mathbb{R}^{2d}}f(z^{\prime})e^{-2\pi i\Theta(z,z^{\prime})}\,dz^{\prime} (1) where \Theta is the symplectic form \displaystyle\Theta(z,z^{\prime}):=x^{\prime}\omega-x\omega^{\prime}. (2) Then the Fourier-Wigner transform and Weyl symbol of an operator S are related by: \mathcal{F}_{W}(S)(z)=\mathcal{F}_{s}(S)(z). 3.3 Convolutions Convolution is an utterly important concept in signal processing. It is used in digital signal processing to study and design linear time-invariant systems such as digital filters. In image processing, convolutional filtering can be used to implement algorithms such as edge detection, image sharpening, and image blurring, to name a few. Convolution of two appropriately chosen functions f and g is defined by f\ast g(x)=\int_{\mathbb{R}}f(y)g(x-y)\,dy=\int_{\mathbb{R}}f(x-y)g(y)\,dy Observe that f\ast g(x)=\int_{\mathbb{R}}f(y)T_{y}\check{g}(x)\,dy , where \check{g}(x)=g(-x) denotes the flip/parity operator, also denoted by Pg(x)=g(-x) . In other words, the convolution is an operator that integrates the translates of the function f weighted by the function g : f\ast g=\int_{\mathbb{R}}g(y)T_{y}f\,dy Given the beautiful properties and tremendous practical importance of convolution, the question of generalisation to operators arises. 3.3.1 Convolution of Operators For two operators S,T on L^{2}(\mathbb{R}) , their

Read the complete text