betaFeedback

WorksMathematics

Kernel Regression using Persistence Diagrams

Abstract

Persistence diagrams are widely used in topological data analysis to summarize topological and geometric structure in data. We study nonparametric regression where inputs are persistence diagrams and outputs are real-valued. Using kernel-based local averaging estimators defined directly over diagram space with Wasserstein distance, we establish consistency and derive bounds on the rate of convergence under suitable assumptions in the L2L^2-theory. We also present computational experiments that demonstrate strong performance of the proposed estimator, particularly in low data regimes.

Introduction

Topological data analysis (TDA) uses multiscale measurements of topology to summarize the shape of data. Persistent homology is one of the principal tools in this field, producing a finite descriptor called a persistence diagram that encodes topological and geometric information across scales [18]. Persistence diagrams have found widespread application in diverse domains including materials science, neuroscience, and computational biology, where shape information provides discriminative features for downstream statistical tasks [33].

In this paper, we study regression where the predictor variable XX takes values in the space of persistence diagrams and the response variable YY is real-valued. Given independent observations (X1,Y1),…,(Xn,Yn)(X_1,Y_1),\ldots,(X_n,Y_n) from a joint distribution, we wish to estimate the regression function m(x)=E[Y∣X=x]m(x) = \mathbb{E}[Y \mid X = x]. This problem arises naturally when one wishes to predict a real-valued quantity such as a physical parameter, a clinical outcome, or a geometric property from topological summaries of complex data.

A fundamental challenge in regression using persistence diagrams is that the space of persistence diagrams, equipped with natural metrics such as the bottleneck or Wasserstein distance, is not a Hilbert space. This precludes direct application of standard regression techniques that rely on vector space structure. The dominant strategy in the literature has therefore been to embed diagrams into a Hilbert space, either explicitly through finite-dimensional vectorization or implicitly through kernel methods [33].

To this end, several vectorization schemes have been studied. Persistence landscapes [6] embed diagrams into a Banach space of functions that can be discretized and Persistence images [2] discretize a weighted Gaussian representation into fixed-length vectors. Kernel methods provide an alternative route via implicit embedding into a reproducing kernel Hilbert space, with kernels such as the persistence scale-space kernel [34] and the sliced Wasserstein kernel [9] enabling kernel regression and Gaussian process regression [26]. A variety of other vectorization methods exist, surveyed in [33], though their application to regression problems (such as in [29] or [30]) has been less systematic.

Notably, all of these approaches operate by embedding persistence diagrams into a Hilbert space. However, such embeddings cannot preserve the native Wasserstein or bottleneck geometry without distortion [10]. Consequently, neighborhood relationships in diagram space

MSC2020 subject classifications: Primary 62R40, Topological data analysis; secondary 62G05, Nonparametric estimation.

Keywords and phrases: Topological data analysis, Non-parametric Regression. may not be faithfully represented in the embedded space, which is particularly relevant for statistical procedures that rely on local metric structure.

Furthermore, there are compelling reasons to operate directly within diagram space. The bottleneck and Wasserstein distances underpin much of the statistical theory for persistence diagrams. Foundational stability results [13, 14] established that small perturbations in input data induce small changes in persistence diagrams. Mileyko, Mukherjee, and Harer [27] proved that the space of persistence diagrams equipped with the pp-Wasserstein metric is a complete, separable metric space, enabling rigorous probability theory on diagrams. This has supported the development of Fréchet means and laws of large numbers [40], confidence sets via bootstrap methods [19], null hypothesis testing for comparing distributions of diagrams [35], and distance covariance methods for measuring dependence between persistence diagrams and response variables [41]. Hence, various statistical tools have already been developed which operate with respect to the Wasserstein or bottleneck metrics.

Despite this, to our knowledge, methods for regression from persistence diagrams to real values that operate directly in the native metric space without any Hilbert embedding have been essentially absent from the literature. We propose to address this gap by developing a regression framework directly on the space of persistence diagrams. Since there is no natural parametric model for probability distributions on this space, we adopt a non-parametric local averaging approach, which relies only on the underlying metric structure of the space of predictors.

In particular, we study kernel-based local averaging estimators, originally proposed by Nadaraya [28] and Watson [42], which constructs the estimator as an average of the known response values weighted by the proximity of the new input xx to the known predictors:

m^n(x)=∑i=1nKh(d(x,Xi))Yi∑i=1nKh(d(x,Xi))\widehat{m}_n(x) = \frac{\sum_{i=1}^{n} K_h(d(x, X_i))Y_i}{\sum_{i=1}^{n} K_h(d(x, X_i))}

where KhK_h is a smoothing kernel function with bandwidth h>0h > 0 and dd is a metric on the predictor space. In the Euclidean case, the Nadaraya–Watson estimator has been extensively studied and shown to be consistent in mean squared error [17, 36] under suitable conditions. Subsequent work extended these results to infinite-dimensional settings and more generally to regression with predictors taking values in abstract metric spaces. In these broader contexts, mean square consistency has been established under different regularity assumptions placed on the regression function or the probability distributions or the metric space itself [21, 1]. Hence, our goal of establishing consistency for our estimator on diagram space can directly leverage this theory.

Another question remains regarding the rate of convergence for such an estimator on our space. In the Euclidean case, if we assume the regression function mm to be in the class of functions Fp\mathcal{F}_p where:

Fp={m:[0,1]d→R∣m is p-Ho¨lder smooth},\mathcal{F}_p = \left\{m : [0,1]^d \to\mathbb{R} \mid m\text{ is }p\text{-Hölder smooth}\right\},

the optimal minimax rate of convergence is shown to be O(n−2p/(2p+d))O(n^{-2p/(2p+d)}). That is:

Rn(Fp)=inf⁡m^nsup⁡m∈FpE[(m^n(X)−m(X))2]≲O(n−2p/(2p+d))\mathcal{R}_n(\mathcal{F}_p) = \inf_{\widehat{m}_n} \sup_{m\in\mathcal{F}_p} \mathbb{E}\left[(\widehat{m}_n(X)-m(X))^2\right] \lesssim O(n^{-2p/(2p+d)})

where the infimum is taken over all estimators m^n\widehat{m}_n (i.e all measurable functions of the data). When p=1p=1, the Hölder class F1\mathcal{F}_1 reduces to the class of Lipschitz continuous functions. In this case, the optimal minimax rate of convergence is

Rn(F1)≲n−2/(2+d)\mathcal{R}_n(\mathcal{F}_1) \lesssim n^{-2/(2+d)}

which kernel estimators have been proven to achieve [22, 38]. Notably, in the Euclidean setting the convergence rate depends on the ambient dimension dd in the denominator, codifying the curse of dimensionality. When predictors take values in an infinite-dimensional or general metric space, the effective “dimension” governing convergence is no longer the ambient Euclidean dimension, but an intrinsic notion of the space’s size or complexity. For instance, rates can depend on small ball probabilities or the Assouad dimension of the predictor space [20, 5, 25, 16, 24]. In this work, we similarly leverage covering number bounds on a restricted diagram space to establish a corresponding upper bound on the rate of convergence of our kernel estimator.

Our Contributions

Given the discussions above, our main contributions are as follows:

  • (i) Restricted Diagram Space. First, we introduce a restricted space of persistence diagrams DN,R\mathcal{D}^{N,R} corresponding to a finite paradigm that reflects the bounded nature of most practical applications. We establish that this space is compact under the pp-Wasserstein distance and derive explicit bounds on its covering numbers.

  • (ii) Asymptotic Results. Next, we prove consistency of the kernel regression estimator on this metric space and derive upper bounds on the rate of convergence in mean squared error. The rate we obtain is O(n−2/(2N+2))O(n^{-2/(2N+2)}), where NN bounds the number of points in the persistence diagrams. This reveals that the parameter 2N2N (corresponding to NN points with birth-death coordinates) plays the role of dimension in determining the convergence rate, providing a satisfying parallel with classical results where rates depend on the ambient dimension.

  • (iii) Empirical Validation. Through computational experiments, we demonstrate strong performance of the proposed estimator. We find that in low-sample regimes where training data is limited, the performance benefit of operating directly in diagram space is most apparent.

Outline

In Section 2, we provide background on persistent homology and persistence diagrams. In Section 3 we define the restricted diagram space DN,R\mathcal{D}^{N,R} for which we prove compactness, and establish covering number bounds. Then in Section 4, we present the general theory for kernel regression on metric spaces and apply it to obtain consistency and convergence rates for persistence diagram predictors. Finally, in Section 5, we present experimental results benchmarking our method against existing approaches.

Persistent Homology and Persistence Diagrams

Here we will provide a very brief introduction to persistent homology. A more complete coverage can be found in [18]. Although homology can be computed over any ring, to compute persistent homology we need a field. This is usually F2\mathbb{F}_2 for computational purposes.

To define persistent homology, we start with a nested sequence of topological spaces,

A0⊆A1⊆A2⊆⋯An=A.(1)A_0 \subseteq A_1 \subseteq A_2 \subseteq\cdots A_n = A. \tag*{(1)}

Often this sequence arises from the sublevel sets of a function, f:A→Rf:A \to\mathbb{R}, with Ai=f−1((−∞,ti]A_i = f^{-1}((-\infty,t_i] for a sequence −∞<t0≤t1≤⋯≤tn≤∞-\infty< t_0 \le t_1 \le\cdots\le t_n \le\infty. The sequence induces linear maps on homology for any dimension rr

Hr(A0)→Hr(A1)→⋯→Hr(An).H_r(A_0) \to H_r(A_1) \to\cdots\to H_r(A_n).

We are interested in when homology classes appear and disappear in this sequence. Let ϕij:Hr(Ai)→Hr(Aj)\phi_i^j:H_r(A_i) \to H_r(A_j) be the linear map on homology induced by the inclusion Ai→AjA_i \to A_j. Observe that if i<j<ki<j<k then ϕjk∘ϕij=ϕik\phi_j^k\circ\phi_i^j=\phi_i^k. The homology class γ∈Hr(Ai)\gamma\in H_r(A_i) is said to be born at AiA_i if it is in the cokernel of ϕi−1i\phi_{i-1}^i. This same class is said to die at XjX_j if its image in Hr(Aj−1)H_r(A_{j-1}) is not in the image of ϕi−1j−1\phi_{i-1}^{j-1}, but its image in Hr(Aj)H_r(A_j) is in the image of ϕi−1j\phi_{i-1}^j. In the case that the spaces arose from the level sets of a function ff as defined above, we define say that γ\gamma is born at time tit_i, dies at time tjt_j, and its lifetime is tj−tit_j-t_i.

A discrete description of the persistent homology of (1) is the multiset of points in the extended plane where we include (ti,tj)(t_i,t_j) with multiplicity the dimension of the set of homology classes born at tit_i and dying at tjt_j. This is the information recorded as the (rrth dimensional) persistence diagram.

Although we have here restricted our definition of persistent homology to that of a finite nested sequence, it can be defined more generally for nested parameterized families of topological spaces {At:t∈R}\{A_t:t\in\mathbb{R}\} with the condition that As⊆AtA_s\subseteq A_t whenever s≤ts\leq t (alongside some technical finiteness conditions). We can think of this as a filtered space with filtration parameter tt. For “nice” filtered spaces the persistence diagram can be defined analogously to the finite case. For details see [15].

It is worth observing that although we are computing homology we are often capturing geometric information through the filtration parameter. For example, if we are considering a filtration of R2\mathbb{R}^2 by the distance from a circle of radius rr then the 1st dimensional persistent homology will have exactly one class born at 00 (the time the circle first appears) and dying at rr (the time when the circle is filled in).

Motivating Example

Consider a point cloud P⊂[0,1]×[0,1]⊂R2P\subset[0,1]\times[0,1]\subset\mathbb{R}^2 in the square portion of the plane as visualized in Figure 1. Our goal is to capture the overall shape information in this point cloud which we know to be roughly circular. To do this, we effectively blur our vision by placing an ε\varepsilon ball at each point in PP, giving us the set ⋃p∈PB(p,ε)\bigcup_{p\in P}B(p,\varepsilon). The relevant information from the blurred point cloud is the overlapping of the ε\varepsilon-balls which connects or blurs the points together in the shape of a loop. This overlap information can be stored in a simplicial complex constructed on ⋃p∈PB(p,ε)\bigcup_{p\in P}B(p,\varepsilon). The Čech complex is an example of such a construction:

Point cloud, epsilon-blurred point cloud, and Čech complex

Figure 1. The point cloud PP (left), the ϵ\epsilon-blurred point cloud ⋃p∈PB(p,ϵ)\bigcup_{p\in P} B(p,\epsilon) (center) and the resulting simplicial complex (right) Cˇϵ(P)\check{C}_{\epsilon}(P) for ϵ=0.22\epsilon= 0.22.

Definition 2.1 (Čech complex). Let P={p1,p2,…,pn}⊂RdP=\{p_1,p_2,\ldots,p_n\}\subset\mathbb{R}^d be a finite point cloud and ε>0\varepsilon>0 a scale parameter. The Čech complex at scale ε\varepsilon, denoted Cˇε(P)\check{C}_\varepsilon(P), is the abstract simplicial complex

Cˇε(P)={σ⊆X | ⋂pi∈σB(pi,ε)≠∅}\check{C}_\varepsilon(P)=\left\{\sigma\subseteq X\ \middle|\ \bigcap_{p_i\in\sigma}B(p_i,\varepsilon)\ne\varnothing\right\}

That is, a subset of points forms a simplex if and only if the corresponding ε\varepsilon-balls have a non-empty intersection.

Figure 1 illustrates the process of constructing the Čech complex with ε=0.22\varepsilon=0.22, in which we can see a connected loop shape emerge.

Finally, we can leverage the tools from algebraic topology, namely calculating the simplicial homology of Cˇϵ(P)\check{C}_{\epsilon}(P). For ϵ=0.22\epsilon= 0.22, we get a single non-trivial 1-cycle which corresponds to a single generator of the first dimensional homology group H1H_1. On the other hand, if our point cloud dataset was linear, the same method would not yield any 1-cycles in the corresponding Čech complex for any value of ϵ\epsilon. Hence, counting the number of generators for H1H_1 in Cˇϵ(P)\check{C}_{\epsilon}(P) quantifies the circular shape information from PP if the value of ϵ\epsilon is well chosen. This allows us to distinguish circular and linearly shaped point clouds.

It is apparent that the correct choice of ϵ\epsilon is central for the success of the above method. The key idea behind persistent homology is that instead of choosing a particular value for ϵ\epsilon, we calculate the homology information at every value of ϵ\epsilon. The persistence diagram stores exactly this information in a slightly different format. As the filtration parameter ϵ\epsilon varies, it record the values of ϵ\epsilon when a new homology feature is born (birth coordinate) and when that feature disappears from the simplicial complex (death coordinate). Hence, a persistence diagram can just be considered as a multiset of points (since multiple homology features can have the same birth-death coordinates) above the diagonal on the plane (since every homology feature must appear before disappearing). Figure 2 illustrates the persistence diagram for our point cloud PP and how each point reflects the birth and death of a homology feature.

Persistence diagram and Čech complexes at epsilon equals 0.19 and 0.50

Figure 2. The persistence diagram (left) displays the birth-death co-ordinates for every homological feature as ϵ\epsilon varies. A dimension 1 homology feature appears for ϵ=0.19\epsilon= 0.19 as a loop in the Čech complex (center) and disappears at ϵ=0.5\epsilon= 0.5 when the loop is filled in with high order simplices (right)

By varying ϵ\epsilon and capturing the changes in homology features, a persistence diagram captures shape information at all scales or all levels of blurring in a single object. Beyond this example, persistence diagrams can be constructed on a variety of complex data types such as images [29] or graphs [3].

While we have defined the construction of a persistence diagram from a finite nested sequence of topological spaces, it will be important to define the space of persistence diagrams independent of the process of persistent homology.

Let R2+={(a,b)∈R2:a≤b}\mathbb{R}^{2+} = \{(a,b) \in\mathbb{R}^{2} : a \le b\}. This is the part of the plane above the diagonal and points lying here correspond to persistent homology classes with finite lifetimes. Since we wish to also want to consider persistent homology classes of infinite duration we will also want to include lines at infinity. Let L∞:={(a,∞):a∈R}L_{\infty} := \{(a,\infty): a \in\mathbb{R}\}. For tractability we will impose some finiteness conditions, namely that only finitely many classes have infinite lifetimes and that the sum of all the finite lifetimes is finite. This restriction is not onerous in applications where generally we have finite sized data as input.

Definition 2.2. A persistence diagram XX is a countable multi-set of L∞∪R2+L_{\infty} \cup\mathbb{R}^{2+} such that the number of elements in X∣L∞X|_{L_{\infty}} is finite, and ∑(xi,yi)∈X∣R2+(yi−xi)<∞\sum_{(x_i,y_i)\in X|_{\mathbb{R}^{2+}}}(y_i-x_i) < \infty.

Remark 2.1. For full generality, we could extend this definition to also include points with −∞-\infty as the birth time. This is rarely useful in applications and will not be needed in our examples later. We removed this extra generality for clearer exposition, but there is no theoretical impediment to include these extra kinds of points.

Let D\mathcal{D} denote the space of all persistence diagrams. We will consider a family of metrics which are analogous to the pp-Wasserstein distances on the space of probability measures and to the LpL^{p} distances on the space of functions on a discrete set. R2+\mathbb{R}^{2+} inherits natural LpL^{p} distances from R2\mathbb{R}^{2}. For p∈[1,∞)p \in[1,\infty) we have ∥(a1,b1)−(a2,b2)∥pp=∣a1−a2∣p+∣b1−b2∣p\|(a_1,b_1)-(a_2,b_2)\|_{p}^{p}=|a_1-a_2|^{p}+|b_1-b_2|^{p} and ∥(a1,b1)−(a2,b2)∥∞=max⁡{∣a1−a2∣,∣b1−b2∣}\|(a_1,b_1)-(a_2,b_2)\|_{\infty}=\max\{|a_1-a_2|,|b_1-b_2|\}.

Let Δ\Delta denote an abstract element representing the diagonal {(x,x):x∈R}\{(x,x):x \in\mathbb{R}\}. With a slight abuse of notation we write ∥(a,b)−Δ∥p\|(a,b)-\Delta\|_{p} to denote the shortest LpL^{p} distance from (a,b)(a,b) in to a point in the diagonal set in R2\mathbb{R}^{2}. Thus

∥(a,b)−Δ∥p=inf⁡t∈R∥(a,b)−(t,t)∥p=21p−1∣b−a∣\|(a,b)-\Delta\|_{p}=\inf_{t\in\mathbb{R}}\|(a,b)-(t,t)\|_{p}=2^{\frac{1}{p}-1}|b-a|

for p<∞p<\infty, and ∥(a,b)−Δ∥∞=inf⁡t∈R∥(a,b)−(t,t)∥∞=∣y−x∣/2\|(a,b)-\Delta\|_{\infty}=\inf_{t\in\mathbb{R}}\|(a,b)-(t,t)\|_{\infty}=|y-x|/2. L∞L_{\infty} also inherits natural LpL^{p} distances from the LpL^{p} metric on R\mathbb{R}; ∥(a1,∞)−(a,∞)∥p=∣a1−a2∣\|(a_1,\infty)-(a,\infty)\|_{p}=|a_1-a_2|. We should think of L∞L_{\infty} and R2+∪Δ\mathbb{R}^{2+}\cup\Delta as disjoint parts of a larger space; the distance between any point in L∞L_{\infty} to any point in R2+∪Δ\mathbb{R}^{2+}\cup\Delta is infinity.

Distances between persistence diagrams are phrased as optimal transport distances where the allowable transportation plans are called matchings.

Definition 2.3. Given persistence diagrams XX and YY a matching is a triple (X′,Y′,ϕ)(X',Y',\phi) with X′⊂XX' \subset X, Y′⊂YY' \subset Y and M:X′→Y′M:X' \to Y' a bijection.

We define the cost of a matching as a sum of costs of moving points from X′X' to Y′Y' along ϕ\phi and moving the unmatched points to the diagonal. The cost will depend on the choice of p∈[1,∞]p \in[1,\infty].

Definition 2.4. Given a matching (X′,Y′,ϕ)(X',Y',\phi) between persistence diagrams XX and YY, its p-cost, cp(X′,Y′,M)c_p(X',Y',M) is defined by

cp(X′,Y′,ϕ)p=∑x∈X′∥x−ϕ(x)∥pp+∑x∈X∖X′∪Y∖Y′∥x−Δ∥ppc_p(X',Y',\phi)^p = \sum_{x\in X'}\|x-\phi(x)\|_p^p+\sum_{x\in X\setminus X'\cup Y\setminus Y'}\|x-\Delta\|_p^p

for p∈[1,∞)p \in[1,\infty), and

c∞(X′,Y′,ϕ)=max⁡{sup⁡x∈X′∥x−ϕ(x)∥∞,sup⁡x∈X∖X′∪Y∖Y′∥x−Δ∥∞}.c_{\infty}(X',Y',\phi)=\max\left\{\sup_{x\in X'}\|x-\phi(x)\|_{\infty},\sup_{x\in X\setminus X'\cup Y\setminus Y'}\|x-\Delta\|_{\infty}\right\}.

For each pp define

dp(X,Y)=inf⁡matchings (X′,Y′,ϕ)cp(X′,Y′,ϕ).d_p(X,Y)=\inf_{\text{matchings }(X',Y',\phi)}c_p(X',Y',\phi).

These distances may be infinite - for example if XX and YY contain a different number of points in L∞L_{\infty} then dp(X,Y)=∞d_p(X,Y)=\infty for all pp.

The Finite Paradigm

The naked space of persistence diagrams is too large to admit useful properties such as compactness, completeness or separability. Various authors as in [27], [30], [8], [40] impose different restrictions to establish different properties. In this Section, we will define the restricted metric space of persistence diagrams (DN,R,dp)(\mathcal{D}^{N,R},d_p) which represents the finite paradigm inherent in the finite nature of most input data in applications. We derive theoretical properties of this space required for establishing the asymptotic results in § 4.

For our purposes, we will restrict diagram space D\mathcal{D} to the finite paradigm as defined below:

Definition 3.1 (Finite Paradigm Restricted Space). For some N,R>0N,R>0, let DN,R\mathcal{D}^{N,R} be the subset of D\mathcal{D} such that for all diagrams XX in DN,R\mathcal{D}^{N,R} we have:

Condition 1: (Finite Total Multiplicity) ∣X∣≤N|X|\leq N

Condition 2: (Finite Region Bound) x∈Xx\in X implies x∈[−R,R]×([−R,R]∪{∞})x\in[-R,R]\times\left([-R,R]\cup\{\infty\}\right)

This creates a triangular region and an interval within L∞L_{\infty} (depicted in Figure 3) which together we denote TRT_R which contains all the finitely many points in any persistence diagram.

Triangular region and interval $T_R$ in the birth-death plane

Figure 3. The set TRT_R where the points in the persistence diagrams in DN,R\mathcal{D}^{N,R} are allowed to be in.

By restricting to DN,R\mathcal{D}^{N,R} we get a well-behaved space in which we can avoid many pathologies such as diagrams with infinitely many points accumulating away from the diagonal. We find this restriction is not limiting because persistence diagrams, in most practical settings, arise from studying processes with real world domain bounds which naturally translate into the global bounds on multiplicity and birth-death times.

For instance, consider a dataset consisting of point clouds in R2\mathbb{R}^{2} contained within a bounded square region. Hence, the filtration parameter is necessarily bounded above by the diameter RR of the square, providing a natural region bound. Moreover, if the process under study can generate only finitely many points within the square, then the number of homological features across all dimensions is finite, yielding some N<∞N < \infty that bounds the number of points in the persistence diagrams. See § 5.2 for a concrete experimental example. Similar reasoning applies to persistence diagrams constructed from images (finite image size and resolution) and from graphs (finite number of vertices and edges). In such cases, the finite paradigm is a natural reflection of the underlying data constraints rather than an artificial restriction.

If persistence diagrams arise from sublevel set filtrations (or from Rips or Cech filtrations of points clouds), then the points in the persistence diagrams with infinite death (those in L∞\mathcal{L}_{\infty}) correspond to the homology classes of the domain and is generally consistent over the dataset of persistence diagrams. This implies we remain in a single connected component within the space of persistence diagrams.

Remark 3.1. The finite paradigm restriction does have its limitations. For instance, as noted in [30], persistence diagrams with infinitely many points may be necessary to encode phenomena such as fractal structures or attractors arising in dynamical systems.

Now we establish some properties for (DN,R,dp)(\mathcal{D}^{N,R},d_p), most importantly compactness and a bound on the covering number. This bound is directly required for bounding the rate of convergence.

Theorem 3.1 (Compactness). For any choice of p∈[1,∞]p \in[1,\infty], the metric space (DN,R,dp)(\mathcal{D}^{N,R},d_p) is compact.

Proof. Let (Xn)(X_n) be a sequence of persistence diagrams in DN,R\mathcal{D}^{N,R}. We can represent each diagram as a vector in (R∪∞)2N(\mathbb{R} \cup\infty)^{2N} since we have at most NN points, each with a birth and death coordinate. So we can represent each diagram as the vector (x1,…,xm,0,…,0)T(x_1,\ldots,x_m,0,\ldots,0)^{T}. Note that 0s used to fill in the vector can be interpreted as the point (0,0)(0,0) on the diagonal which has infinite multiplicity. Note that the ordering of points within the persistence diagram in terms of the birth-death coordinate points in the vector representation is arbitrary, but fixed for each diagram. Hence, the same argument will apply for any choice of ordering.

Note that the birth coordinate of every point in a diagram XnX_n is within the range [−R,R][-R,R] and the death coordinate is in [−R,R]∪{∞}[-R,R] \cup\{\infty\} since Xn∈DN,RX_n \in\mathcal{D}^{N,R}. Hence, the vector representation of each diagram is contained within the closed region ([−R,R]∪{∞})2N([-R,R] \cup\{\infty\})^{2N}. Hence the sequence (Xn)(X_n) can be represented as a sequence of vectors (xn)(x_n) in the region ([−R,R]∪{∞})2N([-R,R] \cup\{\infty\})^{2N}. We can look at this sequence of vectors coordinate-wise with the following notation:

(xn)n∈N=((xn(1),xn(2),…,xn(i),…,xn(2K))T)n∈N(x_n)_{n\in\mathbb{N}}=\left((x_n^{(1)},x_n^{(2)},\ldots,x_n^{(i)},\ldots,x_n^{(2K)})^T\right)_{n\in\mathbb{N}}

We know [−R,R]∪{∞}[-R,R] \cup\{\infty\} is compact with respect to any lpl_p metric. Hence, as a finite product of compact sets, we get that ([−R,R]∪{∞})2N([-R,R] \cup\{\infty\})^{2N} is compact with respect to the product topology by Tychonoff’s theorem. We also know that the sup-norm metric induces the product topology on ([−R,R]∪{∞})2N([-R,R] \cup\{\infty\})^{2N}. Hence, for any choice of p∈[1,∞]p \in[1,\infty], via sequential compactness we get there exists a subsequence (xnj)(x_{n_j}) converging to some x∗∈([−R,R]∪{∞})2Nx_* \in([-R,R]\cup\{\infty\})^{2N}. That is:

max⁡1≤i≤2N∥xnj(i)−x∗(i)∥p→0as j→∞(2)\max_{1\leq i\leq2N}\|x_{n_j}^{(i)}-x_*^{(i)}\|_p \to0 \quad\text{as } j\to\infty \tag*{(2)}

We can interpret x∗x_* as a persistence diagram X∗X_* which will be our candidate limit. Note that X∗X_* is in DN,R\mathcal{D}^{N,R} since x∗x_* has 2N2N components where each coordinate is within [−R,R]∪{∞}[-R,R]\cup\{\infty\}. Hence, it remains to show Xnj→X∗X_{n_j}\to X_*. For p≠∞p\neq\infty we have:

dp(Xnj,X∗)≤(∑i=12N∥xnj(i)−x∗(i)∥pp)1p(3)d_p(X_{n_j},X_*)\leq\left(\sum_{i=1}^{2N}\left\|x_{n_j}^{(i)}-x_*^{(i)}\right\|_p^p\right)^{\frac{1}{p}} \tag*{(3)}

as dpd_p is an optimal transport distance and the choice to send (xn(2j),xn(2j+1))(x_{n}^{(2j)},x_{n}^{(2j+1)}) to (x∗(2j),x∗(2j+1))(x_*^{(2j)},x_*^{(2j+1)}) defines a matching. By (2) we get:

dp(Xnj,X∗)≤(∑i=12N∥xnj(i)−x∗(i)∥pp)1p≤(2N)1p(max⁡1≤i≤2N∥xnj(i)−x∗(i)∥p)→0as j→∞\begin{aligned} d_p(X_{n_j},X_*)&\leq\left(\sum_{i=1}^{2N}\left\|x_{n_j}^{(i)}-x_*^{(i)}\right\|_p^p\right)^{\frac{1}{p}}\\ &\leq(2N)^{\frac{1}{p}}\left(\max_{1\leq i\leq2N}\left\|x_{n_j}^{(i)}-x_*^{(i)}\right\|_p\right)\to0\quad\text{as }j\to\infty \end{aligned}

Note that p=∞p=\infty similarly follows directly from (2). Hence, for any p∈[1,∞]p\in[1,\infty], (Xn)(X_n) has a converging subsequence, so by sequential compactness DN,R\mathcal{D}^{N,R} is compact. □

Definition 3.2 (Covering Number). Let SS be a subset of the metric space (M,d)(M,d). The ε\varepsilon-covering number C(S,ε)C(S,\varepsilon) is the smallest number of open balls of radius ε\varepsilon that cover the set SS:

C(S,ε)=inf⁡{k≥1:∃x1,…,xk∈S such that S⊆⋃i=1kB(xi,ε)}.C(S,\varepsilon)=\inf\left\{k\geq1:\exists x_1,\ldots,x_k\in S\text{ such that }S\subseteq\bigcup_{i=1}^{k}B(x_i,\varepsilon)\right\}.

Theorem 3.2 (Bound on Covering Number). Let ϵ<R/2\epsilon<R/2. For any choice of the metric dpd_p for p∈[1,∞]p\in[1,\infty], the covering number of DN,R\mathcal{D}^{N,R} as a subset of (D,dp)(\mathcal{D},d_p) is bounded above by:

C(DN,R,ϵ)≤(4R2Nϵ2)NC(\mathcal{D}^{N,R},\epsilon)\leq\left(\frac{4R^2N}{\epsilon^2}\right)^N

Proof. As dp(X1,X2)≤dq(X1,X2)d_p(X_1,X_2)\leq d_q(X_1,X_2) for p≥qp\geq q it is sufficient to prove the covering number bound for the case where p=1p=1. To get an upper bound on the covering number C(DN,R,ϵ)C(\mathcal{D}^{N,R},\epsilon) in the ambient metric space (D,d1)(\mathcal{D},d_1), we will explicitly construct a finite set A⊂DN,RA\subset\mathcal{D}^{N,R} such that:

DN,R⊆⋃X∈SB1(X,ϵ)\mathcal{D}^{N,R}\subseteq\bigcup_{X\in S}B_1(X,\epsilon)

where B1B_1 is the ball with respect to d1d_1. For this, we must ensure that any diagram X∈DN,RX\in\mathcal{D}^{N,R} is within an ϵ\epsilon distance of some diagram in AA.

First we consider the triangular region union a line segment in the extended plane, which we previously denoted TRT_R. We can create a grid of points for the triangular region as depicted in Figure 4 such that l1l_1 balls with radius δ\delta cover TRT_R. This grid contains at most

Grid of points over the triangular region and line segment at infinity

Figure 4.

(22R2δ+1)(2R2δ+1)\left(\frac{2\sqrt{2}R}{2\delta}+1\right)\left(\frac{\sqrt{2}R}{2\delta}+1\right)

locations (with the "+1" to account for taking the ceiling). The line segment [−R,R]×{∞}[-R, R] \times\{\infty\} can be covered by l1l_1 balls with radius δ\delta regularly spaced 2δ2\delta apart - needing at most 2R2δ+1\frac{2R}{2\delta}+1. Let GG be the set of points in this grid over TRT_R and the line segment at infinity. Let δ=ϵN\delta= \frac{\epsilon}{N}.

We can show that TRT_R has a finite cover of the form:

⋃xi∈GBl1(xi,ϵN).\bigcup_{x_i \in G} B_{l_1}\left(x_i,\frac{\epsilon}{N}\right).

Construct the set of persistence diagrams containing at most NN points with each of the points lying in GG, and repetitions are allowed. Denote this set of persistence diagrams SS.

Construct the function γ:R2+∪L∞→G\gamma: \mathbb{R}^{2+} \cup L_\infty\to G that sends each point to the closest grid point, breaking ties arbitrarily. For an arbitrary diagram X∈DN,RX \in D^{N,R} consider the diagram where we snap each point to the closest option in GG. The resulting persistence diagram X^={γ(x)∣x∈X}\hat{X} = \{\gamma(x) \mid x \in X\} is by construction in AA. By considering the matching which pairs each point to the closest grid point we have

d1(X,X^)=inf⁡γ(∑x∈X∥x−γ(x)∥1)≤NϵN=ϵ.d_1(X,\hat{X}) = \inf_\gamma\left(\sum_{x\in X}\lVert x-\gamma(x)\rVert_1\right) \leq N\frac{\epsilon}{N}=\epsilon.

This ensures that any point in an arbitrary diagram X∈DN,RX \in D^{N,R} is close to a element in SS.

Counting the number of elements in ZZ is the same as the number of ways of choosing at most NN points from the collection of ∣G∣|G| points with repetitions allowed. There are (N+∣G∣N)\binom{N+|G|}{N} way to do this, so

∣Z∣=(N+∣G∣N).|Z|=\binom{N+|G|}{N}.

It remains to show that ∣Z∣≤(4R2Nϵ2)N|Z| \leq\left(\frac{4R^2N}{\epsilon^2}\right)^N. It is trivially true for N=1N=1 so we can assume that N≥2N\geq2. Recall we have assumed that ϵ<R/2\epsilon<R/2.

First observe that (a+bb)≤(e(a+b)b)b\binom{a+b}{b} \le\left(\frac{e^{(a+b)}}{b}\right)^b for any positive integers a,ba,b. For N≥2N \ge2 and ϵ≤R/2\epsilon\le R/2 we have

e(∣G∣+N)≤e((2RNϵ+1)(RN2ϵ+1)+RNϵ+1+N)≤e(R2N2ϵ2+3.2RNϵ+N+2)\begin{aligned} e(|G|+N) \le e\left(\left(\frac{\sqrt{2}RN}{\epsilon}+1\right)\left(\frac{RN}{\sqrt{2}\epsilon}+1\right)+\frac{RN}{\epsilon}+1+N\right) \\ &\le e\left(\frac{R^2N^2}{\epsilon^2}+3.2\frac{RN}{\epsilon}+N+2\right) \end{aligned}
≤4R2N2ϵ2.\le4\frac{R^2N^2}{\epsilon^2}.

This implies that ∣Z∣≤(e(∣G∣+N)N)N≤(4R2Nϵ2)N|Z| \le\left(\frac{e(|G|+N)}{N}\right)^N \le\left(4\frac{R^2N}{\epsilon^2}\right)^N.

Next, if we remove the option of points with infinity as a death time, we can note that our space DN,R\mathcal{D}^{N,R} is included in diagram spaces constructed by other authors such as by Mileyko et al. [27] and Perea et al. [30]. The conditions imposed by these authors essentially require the number of points that are far away from the diagonal to be finite. They define Dp\mathcal{D}_p spaces in analogy with LpL^p function spaces as below:

Mileyko et al.: Dp={X∈D:dp(X∅,X)<∞ for finite p∈[1,∞)}\mathcal{D}_p = \{X \in\mathcal{D}: d_p(X_{\emptyset}, X) < \infty\text{ for finite } p \in[1,\infty)\}

Perea et al.: D∞={X∈D:∣Wϵ(X)∣<∞ for any ϵ>0}\mathcal{D}_{\infty} = \{X \in\mathcal{D}: |W^{\epsilon}(X)| < \infty\text{ for any } \epsilon> 0\}

Authors in [30] have shown the following inclusion of spaces: D1⊂D2…⊂D∞\mathcal{D}_1 \subset\mathcal{D}_2 \ldots\subset\mathcal{D}_{\infty}. Since we explicitly have a multiplicity bound NN, if we excluded points with infinite death times then would get that DN,R⊂Dp\mathcal{D}^{N,R} \subset\mathcal{D}_p for any pp.

Remark 3.2. Authors in [27] use a slightly different definition for dpd_p, where they fix the l∞l^{\infty}-norm on the R2\mathbb{R}^2 plane instead of letting it vary with pp. Note this morally doesn’t cause issues since for any p<∞p < \infty we have ∥x−y∥∞≤∥x−y∥p≤2∥x−y∥∞\lVert x-y\rVert_{\infty} \le\lVert x-y\rVert_p \le2\lVert x-y\rVert_{\infty}.

In [27], the space Dp\mathcal{D}_p has been shown to be complete and separable. Again, if we excluded the points with infinite death we could use inclusion of DN,R\mathcal{D}^{N,R} as a compact subset in Dp\mathcal{D}_p to show that Dp\mathcal{D}_p is also compact and separable. Since we wish to have up to NN points with infinite death we will instead directly prove completeness and separability.

Finally, we can show that our space is complete and separable.

Theorem 3.3 (Completeness and Separability). For any choice of p∈[1,∞]p \in[1,\infty], the finite paradigm restricted space (DN,R,dp)(\mathcal{D}^{N,R}, d_p) is complete and separable.

Proof. (DN,R,dp)(\mathcal{D}^{N,R}, d_p) is complete because it is compact (Theorem 3.1). We can prove separability through the bounds on the covering numbers. We can construct a countable dense subset by sequentially including sets of persistence diagrams at the centers of balls that cover at radius ϵk=1/k\epsilon_k = 1/k for k=N+k = \mathbb{N}^{+}.

Asymptotic Guarantees

Having set up our metric space of predictors (DN,R,dp)(\mathcal{D}^{N,R}, d_p) over which the local averaging kernel estimator can be defined, we will now establish the mean squared consistency of our estimator, as well as an explicit bound on the rate of convergence. We use the following regression set-up:

Definition 4.1 (Regression Setup). Suppose (M,d)(M,d) is a metric space and let P\mathbb{P} be a Borel probability measure on (M,d)(M,d) with compact support S∗⊂MS^* \subset M. Let {(Xi,Yi)}i=1n\{(X_i,Y_i)\}_{i=1}^n and (X,Y)(X,Y) be i.i.d. random elements taking values in M×RM \times\mathbb{R} with common distribution, satisfying the regression model Y=m(X)+ϵY=m(X)+\epsilon, where

  • m:M→Rm:M\to\mathbb{R}, m(X)=E(Y∣X)m(X)=\mathbb{E}(Y\mid X), is continuous;

  • ϵ\epsilon is random noise such that E(ϵ∣X)=0\mathbb{E}(\epsilon\mid X)=0 and E(ϵ2∣X)=σ2<∞\mathbb{E}(\epsilon^2\mid X)=\sigma^2<\infty.

Remark 4.1. Note that the assumption of P\mathbb{P} having compact support may appear restrictive; however, the ambient space of interest (DN,R,dp)(\mathcal{D}^{N,R},d_p) is itself compact. Furthermore, continuity of mm is a mild strengthening of the Besicovitch condition, which is known to be essentially necessary for the consistency of local averaging methods (see [21, 11, 16]). Thus, these additional assumptions on the classical regression set up are natural in our context.

We can construct an estimator m^\hat{m} for the regression function mm via a local averaging approach

m^n(X)=∑i=1nWi(X)Yi\hat{m}_n(X)=\sum_{i=1}^{n}W_i(X)Y_i

where the weight Wi(X)W_i(X) is a measurable function of (X,X1,…,Xn)(X,X_1,\ldots,X_n). To prove consistency, we want a generalization of Stone’s theorem [37] which provides a set of conditions on the weighting functions which guarantee consistency in mean squared error. Instead of working with the already generalized work in [21] or [1], we will work with the Euclidean proof in [22] which is cleaner and more directly applicable under our assumptions and has not been generalized this way before. As the proof is very analogous, we have provided it in the Appendix B.

Theorem 4.1 (Generalized Stone’s Theorem). Under the regression set up in Definition 4.1, suppose the weight functions {Wi(X)}i=1n\{W_i(X)\}_{i=1}^n of the local averaging estimator satisfy the following conditions:

  1. There is a constant D≥1D\geq1 such that for all nn:

P[∑i=1n∣Wi(X)∣≤D]=1\mathbb{P}\left[\sum_{i=1}^{n}\lvert W_i(X)\rvert\leq D\right]=1
  1. For all δ>0\delta>0:

lim⁡n→∞E[∑i=1n∣Wi(X)∣⋅1{d(Xi,X)>δ}]=0\lim_{n\to\infty}\mathbb{E}\left[\sum_{i=1}^{n}\lvert W_i(X)\rvert\cdot\mathbf{1}_{\{d(X_i,X)>\delta\}}\right]=0

3.

lim⁡n→∞∑i=1nWi(X)=1in probability.\lim_{n\to\infty}\sum_{i=1}^{n}W_i(X)=1\quad\text{in probability}.

4.

lim⁡n→∞E[∑i=1nWi(X)2]=0\lim_{n\to\infty}\mathbb{E}\left[\sum_{i=1}^{n}W_i(X)^2\right]=0

Then m^n(X)\hat{m}_n(X) is consistent in MSE, that is:

E(∣m^n(X)−m(X)∣2)→0asn→∞\mathbb{E}\left(\lvert\hat{m}_n(X)-m(X)\rvert^2\right)\to0\quad\text{as}\quad n\to\infty

The kernel estimator is a type of local averaging estimator and so we can define it in terms of weight functions as below:

Definition 4.2 (Regular Kernel Function). A function K:R→[0,∞]K : \mathbb{R} \to[0,\infty] which satisfies b⋅1[−r,r](x)≤K(x)≤1[−R,R](x)b \cdot\mathbf{1}_{[-r,r]}(x) \le K(x) \le\mathbf{1}_{[-R,R]}(x) for some 0<r≤R0 < r \le R is referred to as a regular kernel function.

Definition 4.3 (Kernel Regression Estimator). The kernel regression estimator m^n(X)=∑i=1nWi(X,X1,…,Xn)Yi\widehat{m}_n(X) = \sum_{i=1}^{n} W_i(X,X_1,\ldots,X_n)Y_i is defined with the weighting functions:

Wi(X,X1,…,Xn)=K(d(Xi,X)hn)∑j=1nK(d(Xj,X)hn)1AW_i(X,X_1,\ldots,X_n) = \frac{K\left(\frac{d(X_i,X)}{h_n}\right)} {\sum_{j=1}^{n}K\left(\frac{d(X_j,X)}{h_n}\right)} \mathbf{1}_{A}

where A={ω∈Mn+1:∑j=1nK(d(Xj,X)hn)≠0}A = \left\{\omega\in M^{n+1} : \sum_{j=1}^{n}K\left(\frac{d(X_j,X)}{h_n}\right) \ne0\right\} for some choice of regular kernel function and choice of bandwidth parameter hnh_n.

It remains to verify that the kernel weights satisfy the conditions of Theorem 4.1. This proof is a direct generalization of the proof for the Euclidean case in [22] (cf. Theorem 5.1), hence it is relegated to the Appendix B.

Theorem 4.2 (Consistency). Consider the kernel estimator in Definition 4.3 under the regression set up in Definition 4.1. Suppose (hn)(h_n) is a sequence of positive numbers where hn→0h_n \to0 as n→∞n \to\infty such that

C(S∗,rhn/2)n→0 as n→∞\frac{C(S^{*},rh_n/2)}{n} \to0 \text{ as } n \to\infty

where rr is the inner support radius of the chosen regular kernel. Then we have m^n(X)→m(X)\widehat{m}_n(X) \to m(X) as n→∞n \to\infty in the L2L^2-norm. Or equivalently:

E(∣m^n(X)−m(X)∣2)→0as n→∞\mathbb{E}\left(\left|\widehat{m}_n(X)-m(X)\right|^2\right) \to0 \quad\text{as } n \to\infty

Remark 4.2. There always exists a sequence (hn)(h_n) satisfying the bandwidth condition of Theorem 4.2. Since S∗S^{*} is compact, its covering number C(S∗,ϵ)C(S^{*},\epsilon) is finite for every ϵ>0\epsilon> 0. Hence, one may choose hn→0h_n \to0 sufficiently slowly so that

C(S∗,rhn/2)n→0.\frac{C(S^{*},rh_n/2)}{n} \to0.

For example, one may choose (hn)(h_n) such that

C(S∗,rhn/2)≤n1/2C(S^{*},rh_n/2) \le n^{1/2}

for all sufficiently large nn. Such a sequence exists because the covering number is finite at every positive radius, and the bandwidth can be chosen to decrease arbitrarily slowly.

Note that Theorem 4.2 can be directly applied to our metric space of interest (DN,R,dp)(\mathcal{D}^{N,R},d_p) given Remark 4.1. Hence, it remains to derive a bound on the rate of convergence for our estimator in line with our second objective. To achieve this, we adapt and extend the Euclidean proof from [22], together with the explicit covering number bound established in Theorem 3.2.

Theorem 4.3 (Explicit Bound on Rate of Convergence). Consider the kernel estimator in Definition 4.3 under the regression set up in Definition 4.1 where (M,d)=(DN,R,dp)(M,d) = (\mathcal{D}^{N,R},d_p). Suppose:

  1. For some C>0C > 0, mm is Lipschitz continuous:

∣m(x)−m(y)∣≤C⋅dp(x,y)|m(x)-m(y)| \le C \cdot d_p(x,y)
  1. The kernel function K(x)K(x) is the naive kernel

K(x)=1[−1,1](x)K(x)=\mathbf{1}_{[-1,1]}(x)
  1. The bandwidth sequence is chosen as:

hn=n−12N+2h_n=n^{-\frac{1}{2N+2}}

Then for large enough nn:

E[∣m^n(X)−m(X)∣2]≤a⋅n−22N+2\mathbb{E}\left[|\widehat{m}_n(X)-m(X)|^2\right]\le a\cdot n^{-\frac{2}{2N+2}}

where aa is some constant independent of nn. Thus, we obtain an upper bound of order O(n−2/(2N+2))O(n^{-2/(2N+2)}) on the rate of convergence of the kernel estimator over our restricted diagram space (DN,R,dp)(D^{N,R},d_p).

Proof. We begin by noting that:

E[(m^n(X)−m(X))2]=∫S∗E[(m^n(x)−m(x))2] dP=∫S∗E[E[(m^n(x)−m(x))2∣X1,…,Xn]] dP\begin{aligned} \mathbb{E}\left[(\widehat{m}_n(X)-m(X))^2\right]=\int_{S_*}\mathbb{E}\left[(\widehat{m}_n(x)-m(x))^2\right]\,d\mathbb{P} \\ &=\int_{S_*}\mathbb{E}\left[\mathbb{E}\left[(\widehat{m}_n(x)-m(x))^2\mid X_1,\ldots,X_n\right]\right]\,d\mathbb{P} \end{aligned}

where the integral is being taken the new input XX. We will first bound the inner expectation term. Note that every random variable ZZ with a finite second moment satisfies the identity:

E[(Z−a)2]=E[(Z−E(Z))2]+(E[Z]−a)2\mathbb{E}\left[(Z-a)^2\right]=\mathbb{E}\left[(Z-\mathbb{E}(Z))^2\right]+(\mathbb{E}[Z]-a)^2

where aa is a constant. We can recover the conditional bias and variance decomposition by substituting Z=m^n(x)Z=\widehat{m}_n(x) and a=m(x)a=m(x) with the expectation being taken conditionally over the random sample X1,…,XnX_1,\ldots,X_n:

E[∣m^n(x)−m(x)∣2∣X1,…,Xn]=E[∣m^n(x)−m^C(x)∣2∣X1,…,Xn]+∣m^C(x)−m(x)∣2\mathbb{E}\left[|\widehat{m}_n(x)-m(x)|^2\mid X_1,\ldots,X_n\right] =\mathbb{E}\left[|\widehat{m}_n(x)-\widehat{m}_C(x)|^2\mid X_1,\ldots,X_n\right]+|\widehat{m}_C(x)-m(x)|^2

where

m^C(x)=E[m^n(x)∣X1,…,Xn]=∑i=1nm(Xi)1{Xi∈B(x,hn)}∑i=1n1{Xi∈B(x,hn)}1A\widehat{m}_C(x)=\mathbb{E}\left[\widehat{m}_n(x)\mid X_1,\ldots,X_n\right] =\frac{\sum_{i=1}^{n}m(X_i)\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}\mathbf{1}_A

Now we can control each of the conditional bias and variance terms separately. For the first conditional bias term we have:

E[∣m^n(x)−m^C(x)∣2∣X1,…,Xn]=E[(∑i=1n(Yi−m(Xi))1{Xi∈B(x,hn)}∑i=1n1{Xi∈B(x,hn)}1A)2|X1,…,Xn]=∑i=1nE[ϵi2∣Xi]1{Xi∈B(x,hn)}(∑i=1n1{Xi∈B(x,hn)})21Asince for a≠b, E(ϵaϵb∣Xa,Xb)=0≤σ2∑i=1n1{Xi∈B(x,hn)}(∑i=1n1{Xi∈B(x,hn)})21A=σ21A∑i=1n1{Xi∈B(x,hn)}\begin{aligned} \mathbb{E}\left[|\widehat{m}_n(x)-\widehat{m}_C(x)|^2\mid X_1,\ldots,X_n\right] \\ &=\mathbb{E}\left[\left(\frac{\sum_{i=1}^{n}(Y_i-m(X_i))\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}\mathbf{1}_A\right)^2\middle|X_1,\ldots,X_n\right] \\ &=\frac{\sum_{i=1}^{n}\mathbb{E}\left[\epsilon_i^2\mid X_i\right]\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\left(\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}\right)^2}\mathbf{1}_A \qquad\text{since for } a\ne b,\ \mathbb{E}(\epsilon_a\epsilon_b\mid X_a,X_b)=0 \\ &\le\sigma^2\frac{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\left(\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}\right)^2}\mathbf{1}_A \\ &=\sigma^2\frac{\mathbf{1}_A}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}} \end{aligned}

Then the second term is controlled since:

(m^C(x)−m(x))2=(∑i=1nm(Xi)1{Xi∈B(x,hn)}∑i=1n1{Xi∈B(x,hn)}1A−[m(x)1A+m(x)1Ac])2≤(∑i=1n∣m(Xi)−m(x)∣1{Xi∈B(x,hn)}∑i=1n1{Xi∈B(x,hn)}1A+∣m(x)∣1Ac)2≤(∑i=1n(Chn)1{Xi∈B(x,hn)}∑i=1n1{Xi∈B(x,hn)}1A+∣m(x)∣1Ac)2since ∣m(Xi)−m(x)∣≤Cdp(Xi,x)≤Chn≤C2hn21A+m(x)21Ac where the cross terms above vanish since 1A1Ac=0≤C2hn2+m(x)21Ac.\begin{aligned} (\hat{m}_{C}(x)-m(x))^{2} &=\left(\frac{\sum_{i=1}^{n}m(X_i)\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}\mathbf{1}_{A}-[m(x)\mathbf{1}_{A}+m(x)\mathbf{1}_{A^c}]\right)^{2}\\ &\leq\left(\frac{\sum_{i=1}^{n}|m(X_i)-m(x)|\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}\mathbf{1}_{A}+|m(x)|\mathbf{1}_{A^c}\right)^{2}\\ &\leq\left(\frac{\sum_{i=1}^{n}(Ch_n)\mathbf{1}_{\{X_i\in B(x,h_n)\}}}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}\mathbf{1}_{A}+|m(x)|\mathbf{1}_{A^c}\right)^{2}\\ &\hspace{7cm}\text{since }|m(X_i)-m(x)|\leq C d_p(X_i,x)\leq Ch_n\\ &\leq C^{2}h_n^{2}\mathbf{1}_{A}+m(x)^{2}\mathbf{1}_{A^c}\text{ where the cross terms above vanish since }\mathbf{1}_{A}\mathbf{1}_{A^c}=0\\ &\leq C^{2}h_n^{2}+m(x)^{2}\mathbf{1}_{A^c}. \end{aligned}

These two terms can then be used to control the L2L^{2}-norm as follows:

E[∣m^n(X)−m(X)∣2]=∫S∗E[E[∣m^n(x)−m^C(x)∣2∣X1,…,Xn]]dP+∫S∗E[∣m^C(x)−m(x)∣2]dP≤σ2∫S∗E[1A∑i=1n1{Xi∈B(x,hn)}]dP+∫S∗[C2hn2+m(x)2E[1Ac]]dP.\begin{aligned} \mathbb{E}\left[|\hat{m}_{n}(X)-m(X)|^{2}\right] &=\int_{S^{*}}\mathbb{E}\left[\mathbb{E}\left[|\hat{m}_{n}(x)-\hat{m}_{C}(x)|^{2}\mid X_{1},\ldots,X_{n}\right]\right]\mathrm{d}\mathbb{P}\\ &\quad+\int_{S^{*}}\mathbb{E}\left[|\hat{m}_{C}(x)-m(x)|^{2}\right]\mathrm{d}\mathbb{P}\\ &\leq\sigma^{2}\int_{S^{*}}\mathbb{E}\left[\frac{\mathbf{1}_{A}}{\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}}}\right]\mathrm{d}\mathbb{P} +\int_{S^{*}}\left[C^{2}h_n^{2}+m(x)^{2}\mathbb{E}[\mathbf{1}_{A^c}]\right]\mathrm{d}\mathbb{P}. \end{aligned}

Noting that ∑i=1n1{Xi∈B(x,hn)}\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(x,h_n)\}} is a binomial random variable as a sum of i.i.d bernoulli random variables, we can apply Lemma A.2 to upper bound the first integral as:

E[∣m^n(X)−m(X)∣2]≤2σ2n+1∫S∗1P(B(x,hn))dP+C2hn2+∫S∗(1−P(B(x,hn)))nm(x)2dP.\mathbb{E}\left[|\hat{m}_{n}(X)-m(X)|^{2}\right] \leq\frac{2\sigma^{2}}{n+1}\int_{S^{*}}\frac{1}{\mathbb{P}(B(x,h_n))}\mathrm{d}\mathbb{P} +C^{2}h_n^{2} +\int_{S^{*}}(1-\mathbb{P}(B(x,h_n)))^{n}m(x)^{2}\mathrm{d}\mathbb{P}.

Since mm is continuous on a compact support S∗S^{*}, it attains the extremum value B=sup⁡x∈S∗∣m(x)∣2B=\sup_{x\in S^{*}}|m(x)|^{2}. Using this we can control the second integral as follows:

∫S∗(1−P(B(x,hn)))nm(x)2dP≤B∫S∗(1−P(B(x,hn)))ndP≤B∫S∗exp⁡(−n⋅P(B(x,hn)))dPsince ∀x∈[0,1]:(1−x)n≤e−nx=B∫S∗n⋅P(B(x,hn))n⋅P(B(x,hn))exp⁡(−n⋅P(B(x,hn)))dP≤Bn∫S∗1P(B(x,hn))dPsince ∀x∈[0,1]:nxe−nx≤1.\begin{aligned} \int_{S^{*}}(1-\mathbb{P}(B(x,h_n)))^{n}m(x)^{2}\mathrm{d}\mathbb{P} &\leq B\int_{S^{*}}(1-\mathbb{P}(B(x,h_n)))^{n}\mathrm{d}\mathbb{P}\\ &\leq B\int_{S^{*}}\exp\left(-n\cdot\mathbb{P}(B(x,h_n))\right)\mathrm{d}\mathbb{P} \qquad\text{since }\forall x\in[0,1]:(1-x)^{n}\leq e^{-nx}\\ &=B\int_{S^{*}}\frac{n\cdot\mathbb{P}(B(x,h_n))}{n\cdot\mathbb{P}(B(x,h_n))} \exp\left(-n\cdot\mathbb{P}(B(x,h_n))\right)\mathrm{d}\mathbb{P}\\ &\leq\frac{B}{n}\int_{S^{*}}\frac{1}{\mathbb{P}(B(x,h_n))}\mathrm{d}\mathbb{P} \qquad\text{since }\forall x\in[0,1]:nxe^{-nx}\leq1. \end{aligned}

Hence, we get:

E[∣m^n(X)−m(X)∣2]≤(2σ2n+1+Bn)∫S∗1P(B(x,hn))dP+C2hn2.\mathbb{E}\left[|\hat{m}_{n}(X)-m(X)|^{2}\right] \leq\left(\frac{2\sigma^{2}}{n+1}+\frac{B}{n}\right)\int_{S^{*}}\frac{1}{\mathbb{P}(B(x,h_n))}\mathrm{d}\mathbb{P}+C^{2}h_n^{2}.

Now applying the small ball probability control Lemma A.1 and simplifying we get the following inequality in terms of hnh_n:

E[∣m^n(X)−m(X)∣2]≤(2σ2+Bn)C(S∗,hn2)+C2hn2.\mathbb{E}\left[|\hat{m}_{n}(X)-m(X)|^{2}\right] \leq\left(\frac{2\sigma^{2}+B}{n}\right)C\left(S^{*},\frac{h_n}{2}\right)+C^{2}h_n^{2}.

Now we use the explicit upper bound on the covering number function from Theorem 3.2 to construct an explicit hnh_n sequence. Substituting in the bound we get:

E[∣m^n(X)−m(X)∣2]≤(2σ2+sup⁡x∈S∗∣m(x)∣2n)(16R2Nhn2)N+C2hn2=(16R2N)N(2σ2+sup⁡x∈S∗∣m(x)∣2)(1nhn2N)+C2hn2≤a⋅1nhn2N+C2hn2\begin{aligned} \mathbb{E}\left[\left|\hat{m}_n(X)-m(X)\right|^2\right]\leq\left(\frac{2\sigma^2+\sup_{x\in S^*}|m(x)|^2}{n}\right)\left(\frac{16R^2N}{h_n^2}\right)^N+C^2h_n^2 \\ &= (16R^2N)^N\left(2\sigma^2+\sup_{x\in S^*}|m(x)|^2\right)\left(\frac{1}{nh_n^{2N}}\right)+C^2h_n^2 \\ &\leq a\cdot\frac{1}{nh_n^{2N}}+C^2h_n^2 \end{aligned}

where aa is some constant independent of nn. Now we can choose our sequence of hnh_n to be given by:

hn=n−12N+2h_n=n^{-\frac{1}{2N+2}}

which we have picked so that we get the following simplification after substituting hnh_n into the upper bound:

E[∣m^n(X)−m(X)∣2]≤a⋅1nhn2N+C2hn2=a⋅n−1⋅n2N2N+2+C2⋅n−22N+2=a′⋅n−22N+2for large n\begin{aligned} \mathbb{E}\left[\left|\hat{m}_n(X)-m(X)\right|^2\right]\leq a\cdot\frac{1}{nh_n^{2N}}+C^2h_n^2 \\ &= a\cdot n^{-1}\cdot n^{\frac{2N}{2N+2}}+C^2\cdot n^{-\frac{2}{2N+2}} \\ &= a'\cdot n^{-\frac{2}{2N+2}}\quad\text{for large }n \end{aligned}

where a′a' is just some constant independent of nn.

Now there is an interesting point of comparison to be made with the Euclidean case where the optimal rate of convergence has upper bounded of order O(n−2/(d+2))O(n^{-2/(d+2)}) [22] for Lipschitz continuous regression functions in ambient dimension Rd\mathbb{R}^d. Thus, the ambient dimension parameter dd is effectively replaced by 2N2N in our setting where NN is the bound on the number of off-diagonal points in the persistence diagrams. We can interpret the value 2N2N as NN points living in 2-dimensional space (with birth-death coordinates). Intuitively, letting each birth and death coordinate of NN points be a free variable, we end up in R2N\mathbb{R}^{2N}. Hence, in this context, we can see how the bound on the total multiplicity of a set of persistence diagrams fulfills a dimensionality criteria.

Experimental Results

Finally, we will test the performance of our estimator.

Implementation Outline

Given the sample data {(xi,yi)}i=1N\{(x_i,y_i)\}_{i=1}^N and a new input xN+1x_{N+1}, we have implemented the following algorithm in Python:

Step 1: For each sample input xix_i and the new input xN+1x_{N+1}, we construct the corresponding set of persistence diagram, resulting in a set of persistence diagrams {d1,…,dN,dN+1}\{d_1,\ldots,d_N,d_{N+1}\}. We have used the Ripser library [4] which constructs a computationally efficient Vietoris-Rips simplicial complex.

Step 2: Compute the distances between the new input persistence diagram dN+1d_{N+1} and the sample diagrams {d1,…,dN}\{d_1,\ldots,d_N\} using the metric dpd_p; where we take p=2p=2 throughout. Distances are computed using the GUDHI library [32] to ensure computational efficiency. Note that essential classes (infinite death time) are retained and matched only to one another by default. The distances for the H0H_0 and H1H_1 persistence diagrams are evaluated separately and then summed, yielding a collection of distances {r1,…,rN}\{r_1,\ldots,r_N\}. Step 3: Using the distances, we calculate a set of weights for each input sample {w1,…,wN}\{w_1,\ldots,w_N\} using Definition 4.3. We’ve chosen the Gaussian Kernel Kh(z)=12πexp⁡(−z22h2)K_h(z) = \frac{1}{\sqrt{2\pi}}\exp\left(-\frac{z^2}{2h^2}\right) and chosen the bandwidth parameter hh via leave-one-out cross-validation. Note that we have not chosen a value for RR in Definition 4.3 so we never run into the case where the sum of kernel weights is 0.

Step 4: Create the regression estimate by the weighted average of target values {yi}i=1N\{y_i\}_{i=1}^{N} using the previously calculated weights.

The Python code used to implement the method, along with all datasets used in the experiments have been made publicly available at kernel-regression-persistence-diagrams.git

Interpolation Parameter Experiment

We begin with demonstrating our method on a toy dataset.

Dataset

We illustrate our method on a synthetic toy dataset designed to interpolate continuously between two simple geometric structures. Fix 20 equidistant angles θ1,…,θ20∈[0,2π]\theta_1,\ldots,\theta_{20} \in[0,2\pi]. Let Ui(x),Ui(y)∼Uniform⁡(−0.2,0.2)U_i^{(x)}, U_i^{(y)} \sim\operatorname{Uniform}(-0.2, 0.2) be independent noise variables for each point and each coordinate. For a given interpolation parameter t∈[0,1]t \in[0,1], we generate a point cloud {(xi,yi)}i=120⊂R2\{(x_i,y_i)\}_{i=1}^{20} \subset\mathbb{R}^2 according to

xi=cos⁡(θi)+Ui(x),yi=(1−t)cos⁡(θi)+tsin⁡(θi)+Ui(y).x_i = \cos(\theta_i) + U_i^{(x)}, \qquad y_i = (1-t)\cos(\theta_i) + t\sin(\theta_i) + U_i^{(y)}.

When t=0t = 0, the point cloud concentrates around a noisy circle, while for t=1t = 1 it lies approximately along a noisy line. Intermediate values of tt yield a smooth geometric interpolation between these two regimes. The regression task is to predict the parameter tt from a noisy point cloud realization. Figure 5 illustrates representative samples for several values of tt.

Sample point clouds, persistence diagrams, and persistence images for the interpolation parameter experiment

Figure 5. Sample point clouds from the Interpolation Parameter Experiment (t=0t = 0, 0.50.5, and 11) and their associated full persistence diagrams PD and persistence images PI.

Methods

To assess the benefits of operating directly in the space of persistence diagrams, we compare our kernel regression method described in § 5.1 with a commonly used vectorization approach based on persistence images. In both cases, we employ the same kernel regression framework. The difference lies solely in the representation of the persistence diagrams: either as elements of persistence diagram space equipped with the Wasserstein distance, or as finite-dimensional vectors in Euclidean space via persistence images. Figure 5 includes the persistence diagrams and persistence images associated with the sample point clouds for each value of tt.

Persistence images provide a stable vectorization of persistence diagrams by convolving each persistence point with a Gaussian kernel and discretization the resulting surface on a fixed grid. We follow the construction of [12] and refer to that work for full methodological details. As noted in [12], the method is robust to the choice of resolution and Gaussian variance parameters. Hence, we’ve chosen suitable parameters and fixed the Gaussian variance at 0.1 and use a 50×5050 \times50 grid resolution throughout. No weighting function is applied, as we aim to evaluate the estimator without assuming higher-persistence features are more important. All persistence image summaries are computed using GUDHI [32].

Results

Predictive performance is evaluated using the coefficient of determination R2R^2 between the predicted and true parameter values. Let ntrainn_{\mathrm{train}} and ntestn_{\mathrm{test}} denote the number of training and testing samples respectively. We fix ntest=50n_{\mathrm{test}} = 50 and vary ntrainn_{\mathrm{train}} to study performance under differing levels of data availability. For each value of ntrainn_{\mathrm{train}}, the experiment is repeated 10 times with independent training and test sets, and we report the mean and standard deviation of the resulting R2R^2 scores.

Figure 6 summarizes the results by plotting the mean R2R^2 as a function of ntrainn_{\mathrm{train}} for each method, with error bars indicating one standard deviation across the 10 repetitions. Across all training sample sizes, the proposed kernel regression method on persistence diagrams attains higher mean R2R^2 values than the persistence image based approach, with the performance gap most pronounced in low-sample regimes. As ntrainn_{\mathrm{train}} increases, the performance of both methods improves, though the persistence diagram based estimator remains consistently superior. The results indicate that preserving the geometry of persistence diagram space can yield improved predictive performance, particularly when training data are limited. Exact numerical values corresponding to the curves in Figure 6 are reported in the Table 1.

Mean predictive performance as a function of training sample size

Figure 6. Mean predictive performance (R2R^2) as a function of training sample size ntrainn_{\mathrm{train}} for persistence diagram (PD) and persistence image (PI) based methods. Points denote empirical means across repetitions. Error bars indicate ±1\pm1 standard deviation.

ntrainn_{\mathrm{train}}PD_kernelPI_kernel
100.8435 (0.0901)0.5283 (0.1354)
200.9221 (0.0246)0.7219 (0.0684)
300.9277 (0.0233)0.7166 (0.1675)
400.9368 (0.0167)0.7058 (0.1545)
500.9329 (0.0101)0.7652 (0.0351)
600.9331 (0.0145)0.7798 (0.0745)
700.9247 (0.0317)0.7970 (0.0353)
800.9260 (0.0238)0.7716 (0.0290)
900.9501 (0.0146)0.8007 (0.0724)
1000.9458 (0.0129)0.7931 (0.0956)

Table 1. Interpolation parameter experiment: mean R2R^2 with standard deviation (in parentheses) over 10 repetitions.

Curvature Parameter Experiment

Next, we evaluate our method on a regression task introduced in [7], which aims to predict the curvature κ\kappa of an underlying surface with constant curvature from a point cloud sampled from a unit disk embedded in that surface. This task is motivated by applications in geometry processing and shape analysis, where curvature encodes fundamental local geometric information. An extensive computational study of this task is conducted in [39]. We replicate their experimental setup, but reduce both the number of training samples and the number of points sampled per point cloud. This modification allows us to assess performance in a low-data regime, where differences between methods are most apparent.

Dataset

Following the procedure of [39], each point cloud is generated by sampling from a unit disk DκD_{\kappa} embedded in a surface with constant curvature κ∈[−2,2]\kappa\in[-2,2]. For 50 evenly spaced curvature values, we generate point clouds consisting of 50 points, where the sampling density is proportional to the surface area measure. This yields 50 training samples. Test data are constructed analogously using 50 curvature values sampled uniformly at random from [−2,2][-2,2]. We refer the reader to [39] for full details of the data generation procedure and the associated code.

Methods

We evaluate the same kernel regression framework described in § 5.1, using persistence diagram representations of the point clouds (denoted as PD_kernel). For comparison, all baseline methods from [39] are included. These are comprised of TDA-based machine learning approaches after vectorization (ml_on_ph0, ml_on_ph1, ml_on_ph0_longest, ml_on_ph1_longest, ph0_ml, ph1_ml, ml_on_dis_mat) and neural network approaches (nn_shallow, nn_deep, point_net). All methods are evaluated using the original implementations provided by the authors. We refer the reader to [39] for full methodological details.

Results

Predictive performance is evaluated using the mean squared error (MSE), following the same evaluation protocol as [39]. Table 2 summarizes the results. The proposed kernel regression method on persistence diagrams attains the lowest MSE among the methods considered.

MethodMSE
ml_on_ph00.3498
ml_on_ph10.6315
ml_on_ph0_longest0.2912
ml_on_ph1_longest0.6189
ph0_ml0.2669
ph1_ml0.5247
ml_on_dis_mat0.3916
nn_shallow1.1812
nn_deep2.7978
point_net37.1275
PD_kernel0.2258

Table 2. Curvature experiment results: Mean Square Error (MSE) comparison between all methods.

Reynolds Number Experiment

Finally, we introduce a new experiment to investigate the applicability of the proposed method to regression problems involving complex spatial data. Specifically, we consider the prediction of the Reynolds number of a simulated fluid flow from its velocity field. The Reynolds number characterizes the relative importance of inertial and viscous forces in a fluid, with low Reynolds numbers generally associated with smooth, orderly flow patterns, while high Reynolds numbers give rise to more complex, irregular spatial structures and turbulent behaviour. This provides an interesting setting for persistence diagram regression, as the topological features of the velocity field may encode information about these underlying differences in flow structure.

Dataset

We consider the 2D Unsteady Fluid Simulation Ensemble for Machine Learning introduced in [23]. The dataset contains 8000 simulations of incompressible flow on a periodic square domain (a torus), each stored as a sequence of velocity field snapshots (u,v)(u, v) on a 512×512512 \times512 grid at times t∈[0,10]t \in[0, 10]. The simulations differ in their Reynolds number Re∈[1,4096]\mathrm{Re} \in[1, 4096], which controls the turbulence and spatial complexity of the flow. Representative flows for different Reynolds numbers at different snapshot times are shown in Figure 8. See [31] for details on how the fluid flow simulations were created. The regression task is to predict ln⁡Re\ln\mathrm{Re} of a simulation from a snapshot of the fluid velocity field at a fixed point in time for t∈{1,2.5,5,7.5,9}t \in\{1, 2.5, 5, 7.5, 9\}.

Sample vorticity fields from fluid dynamics dataset and associated persistence diagrams

Figure 8. Sample vorticity fields from fluid dynamics dataset [23] and their associated persistence diagrams, at snapshot times t=1t = 1 and t=9t = 9. Different simulations are shown (with each row) with varying Reynolds numbers.

Methods

We use the same kernel regression procedure of § 5.1 with a few changes to adapt to this setting. Since the snapshots are vector fields on a grid, rather than point clouds, Step 1 differs from the other experiments. For each snapshot of the velocity field we compute the vorticity scalar field ω=∂xv−∂yu\omega= \partial_x v - \partial_y u and the persistence diagram of its sublevel-set filtration (Figure 7). Steps 2–4 in the procedure are unchanged.

Data pipeline from velocity field to vorticity field to persistence diagram

Figure 7. Data pipeline for converting a fluid simulation snapshot at a fixed time point to a persistence diagram. The velocity field (left) gives the vorticity field (middle) via its curl; and the sublevel-set filtration of the vorticity yields the persistence diagram (right).

The diagrams are computed with the cubical complex module of GUDHI [32]. The 512×512512 \times512 grid of vorticity values is passed as the filtration values on the vertices of a cubical complex, in which each higher-dimensional cell takes the maximum value of its vertices, so the sublevel set {ω≤s}\{\omega\le s\} is the union of the cells with filtration value at most ss. Because the simulation domain is periodic, we use the periodic cubical complex with periodic boundary conditions in both directions, so the underlying space is a torus rather than a square. Representative vorticity fields and persistence diagrams at different Reynolds numbers and snapshot times are shown in Figure 8.

Results

We use 400 simulations spread evenly over the whole range of Reynolds numbers, and the same simulations at every time point t∈{1,2.5,5,7.5,9}t \in\{1, 2.5, 5, 7.5, 9\}. Performance is measured by the out-of-sample R2R^2 on ln⁡Re⁡\ln\operatorname{Re}. For each of 5 random train/test splits, we hold out 100 simulations for testing and train on nested subsets of the remaining simulations of size ntrain∈{25,50,100,200,300}n_{\mathrm{train}} \in\{25, 50, 100, 200, 300\}.

Figure 9 plots the mean R2R^2 as a function of ntrainn_{\mathrm{train}} for each snapshot time, with error bars indicating one standard deviation across the 5 splits, and the exact values are reported in Table 3. The mean R2R^2 is higher at earlier snapshot times for every training size, demonstrating that our kernel method for persistence diagrams has strong predictive power for inferring the Reynold’s number at time t=1t = 1.

ntrainn_{\mathrm{train}}t=1t = 1t=2.5t = 2.5t=5t = 5t=7.5t = 7.5t=9t = 9
250.7620 (0.0319)0.5295 (0.0232)0.2811 (0.0590)0.1401 (0.0886)0.0979 (0.0589)
500.7643 (0.0411)0.5175 (0.0710)0.2927 (0.0502)0.1632 (0.0720)0.1204 (0.0708)
1000.7756 (0.0382)0.5543 (0.0704)0.3113 (0.0980)0.2004 (0.0856)0.1715 (0.0844)
2000.8011 (0.0275)0.5685 (0.0689)0.3317 (0.0893)0.2230 (0.0619)0.1918 (0.0597)
3000.8083 (0.0243)0.5856 (0.0563)0.3451 (0.0669)0.2498 (0.0515)0.2162 (0.0552)

Table 3. Reynolds Number Experiment: mean R2R^2 with standard deviation (in parentheses) over 5 random train/test splits, for each snapshot time tt.

Mean predictive performance as a function of training sample size for persistence diagram kernel regression

Figure 9. Mean predictive performance (R2R^2) as a function of training sample size ntrainn_{\mathrm{train}} for persistence diagram kernel regression of ln⁡Re\ln\mathrm{Re} from a single vorticity snapshot, shown for snapshot times t∈{1,2.5,5,7.5,9}t \in\{1, 2.5, 5, 7.5, 9\}. Points denote empirical means across 5 random train/test splits with 100 test simulations each. Error bars indicate ±1\pm1 standard deviation.

For snapshots of the fluid simulations at later times, the method does not perform as well, and the R2R^2 value drops significantly. On further examination, we see that this is due to the fluid settling in to an almost-steady state at later times with similar vortex structures in each simulation at different Reynold’s numbers. See Figure 8 to compare the vorticity fields at t=1t = 1 and t=9t = 9.

The similarity at say t=9t = 9 makes the corresponding persistence diagrams harder to distinguish, but that difficulty also holds to some extent for visual inspection of the vortex field. Looking at the vortex fields in Figure 8, it is hard to see much difference for example between the Re⁡=158.2\operatorname{Re} = 158.2 and Re⁡=682.5\operatorname{Re} = 682.5 simulations at t=9t = 9. It is no surprise then that our kernel estimation of the Reynold’s number loses accuracy at later times. This is reflected in Figure 10, which plots the predicted against the true ln⁡Re⁡\ln\operatorname{Re} for a single split with ntrain=300n_{\mathrm{train}} = 300 at t=1t = 1, 5 and 9. The predicted values deviate much more from the true values at higher Reynolds numbers, and this deviation is larger at the later time points (R2=0.814R^2 = 0.814, 0.242 and 0.128 respectively).

Predicted versus true $\ln \mathrm{Re}$ on a single train/test split for persistence diagram kernel regression at snapshot times $t = 1$, $5$ and $9$

Figure 10. Predicted versus true ln⁡Re\ln\mathrm{Re} on a single train/test split (ntrain=300n_{\mathrm{train}} = 300, ntest=100n_{\mathrm{test}} = 100) for persistence diagram kernel regression at snapshot times t=1t = 1, 55 and 99. Each point is one test simulation, and the dotted line y=xy = x marks perfect prediction. The R2R^2 of the split is shown in each panel.

Technical Lemmas

The lemma below allows us to bound probabilities with covering numbers of our space. The original proof can be found in [1].

Lemma A.1 (Small Ball Probability Control). Let (M,B(M),P)(M,\mathcal{B}(M),\mathbb{P}) be a probability space and let S∗⊂MS^{*} \subset M be compact. Consider the ball B(x,r)⊂MB(x,r) \subset M for some fixed x∈Mx \in M and r>0r > 0. Then:

∫S∗1P(B(x,r)) dP≤C(S∗,r2)\int_{S^{*}} \frac{1}{\mathbb{P}(B(x,r))}\,\mathrm{d}\mathbb{P} \le C\left(S^{*},\frac{r}{2}\right)

The proof for the lemma below can be found in [22].

Lemma A.2 (Binomial inverse expectation bound). Let B(n,p)B(n,p) be a binomial random variable with parameters nn and pp. Then:

(i)

E[11+B(n,p)]≤1(n+1)p\mathbb{E}\left[\frac{1}{1+B(n,p)}\right] \le\frac{1}{(n+1)p}

(ii)

E[1B(n,p)1{B(n,p)>0}]≤2(n+1)p\mathbb{E}\left[\frac{1}{B(n,p)}\mathbf{1}_{\{B(n,p)>0\}}\right] \le\frac{2}{(n+1)p}

Additional Proofs

Proof of Theorem 4.1.

Proof.

∥m^n(X)−m(X)∥L22=E[∣m^n(X)−m(X)∣2]=E[(∑i=1nWi(X)Yi−m(X))2]\left\|\widehat{m}_{n}(X)-m(X)\right\|_{L^{2}}^{2} = \mathbb{E}\left[\left|\widehat{m}_{n}(X)-m(X)\right|^{2}\right] = \mathbb{E}\left[\left(\sum_{i=1}^{n} W_i(X)Y_i-m(X)\right)^{2}\right]

We can write the above expression as:

E[(∑i=1nWi(X)Yi−∑i=1nWi(X)m(Xi)+∑i=1nWi(X)m(Xi)−∑i=1nWi(X)m(X)+∑i=1nWi(X)m(X)−m(X))2]\mathbb{E}\left[\left(\sum_{i=1}^{n} W_i(X)Y_i-\sum_{i=1}^{n}W_i(X)m(X_i)+\sum_{i=1}^{n}W_i(X)m(X_i)-\sum_{i=1}^{n}W_i(X)m(X)+\sum_{i=1}^{n}W_i(X)m(X)-m(X)\right)^{2}\right]

Using the inequality (a+b+c)2≤3a2+3b2+3c2(a+b+c)^{2} \le3a^{2}+3b^{2}+3c^{2}, we split the expectation as follows:

E[(m^n(X)−m(X))2]≤3E[(∑i=1nWi(X)(Yi−m(Xi)))2]+3E[(∑i=1nWi(X)(m(Xi)−m(X)))2]+3E[(∑i=1nWi(X)−1)2m(X)2]=:3In+3Jn+3Ln\begin{aligned} \mathbb{E}\left[(\widehat{m}_{n}(X)-m(X))^{2}\right] \le3\mathbb{E}\left[\left(\sum_{i=1}^{n}W_i(X)(Y_i-m(X_i))\right)^{2}\right] +3\mathbb{E}\left[\left(\sum_{i=1}^{n}W_i(X)(m(X_i)-m(X))\right)^{2}\right] \\ &+3\mathbb{E}\left[\left(\sum_{i=1}^{n} W_i(X)-1\right)^2m(X)^2\right] =:3I_n+3J_n+3L_n \end{aligned}

We will show that each of In,Jn,Ln→0I_n,J_n,L_n \to0 as n→∞n \to\infty, which is the desired result.

  1. First we show Ln→0L_n \to0 as n→∞n \to\infty. We know by condition 1, that almost surely for all nn:

(∑i=1nWi(X)−1)2m(X)2≤(D+1)2∣m(X)∣2\left(\sum_{i=1}^{n} W_i(X)-1\right)^2m(X)^2 \le(D+1)^2|m(X)|^2

We only need to consider mm on the compact support S∗S^*. Since mm is continuous on S∗S^*, we know ∣m(X)∣≤Bm|m(X)| \le B_m for some Bm≥0B_m \ge0. Hence, we attain a uniform bound by an integrable function:

(∑i=1nWi(X)−1)2m(X)2≤(D+1)2Bm2\left(\sum_{i=1}^{n} W_i(X)-1\right)^2m(X)^2 \le(D+1)^2B_m^2

Finally, since condition 3 gives us the following convergence in probability:

(∑i=1nWi(X)−1)2m(X)2→P0as n→∞\left(\sum_{i=1}^{n} W_i(X)-1\right)^2m(X)^2 \xrightarrow{\mathbb{P}} 0 \quad\text{as } n \to\infty

we can apply the Vitali convergence theorem to obtain the desired result:

Ln=E[(∑i=1nWi(X)−1)2m(X)2]→0L_n=\mathbb{E}\left[\left(\sum_{i=1}^{n} W_i(X)-1\right)^2m(X)^2\right]\to0
  1. Next we work on In→0I_n \to0 as n→∞n \to\infty

In=E[(∑i=1nWi(X)(Yi−m(Xi)))2]=∑i=1n∑j=1nE[Wi(X)Wj(X)(Yi−m(Xi))(Yj−m(Xj))]\begin{aligned} I_n=\mathbb{E}\left[\left(\sum_{i=1}^{n} W_i(X)(Y_i-m(X_i))\right)^2\right] \\ &=\sum_{i=1}^{n}\sum_{j=1}^{n}\mathbb{E}\left[W_i(X)W_j(X)(Y_i-m(X_i))(Y_j-m(X_j))\right] \end{aligned}

But the cross-terms disappear since for i≠ji\ne j:

E[Wi(X)⋅Wj(X)⋅(Yi−m(Xi))⋅(Yj−m(Xj))]=E[E[(Yj−m(Xj))∣X,X1,…,Xn,Yi]Wi(X)Wj(X)(Yi−m(Xi))]\begin{aligned} \mathbb{E}\left[W_i(X)\cdot W_j(X)\cdot(Y_i-m(X_i))\cdot(Y_j-m(X_j))\right] \\ &=\mathbb{E}\left[\mathbb{E}\left[(Y_j-m(X_j))\mid X,X_1,\ldots,X_n,Y_i\right]W_i(X)W_j(X)(Y_i-m(X_i))\right] \end{aligned}

by law of total expectation and independence

=0sinceE[(Yj−m(Xj))∣X,X1,…,Xn,Yi]=E(Yj∣Xj)−m(Xj)=0.=0 \quad\text{since}\quad\mathbb{E}\left[(Y_j-m(X_j))\mid X,X_1,\ldots,X_n,Y_i\right]=\mathbb{E}(Y_j\mid X_j)-m(X_j)=0.

So it remains to control:

In=E[∑i=1nWi(X)2(Yi−m(Xi))2]=E[∑i=1nWi(X)2ϵi2]I_n=\mathbb{E}\left[\sum_{i=1}^{n}W_i(X)^2(Y_i-m(X_i))^2\right]=\mathbb{E}\left[\sum_{i=1}^{n}W_i(X)^2\epsilon_i^2\right]

But conditioning on X,X1,…,XnX, X_1, \ldots, X_n gives:

E[∑i=1nWi(X)2ϵi2]=E[∑i=1nE(Wi(X)2ϵi2∣X,X1,…,Xn)]=E[∑i=1nWi(X)2E(ϵi2∣Xi)]≤σ2E[∑i=1nWi(X)2]→n→∞0\begin{aligned} \mathbb{E}\left[\sum_{i=1}^{n} W_i(X)^2\epsilon_i^2\right] &= \mathbb{E}\left[\sum_{i=1}^{n}\mathbb{E}\left(W_i(X)^2\epsilon_i^2 \mid X,X_1,\ldots,X_n\right)\right] \\ &= \mathbb{E}\left[\sum_{i=1}^{n} W_i(X)^2\mathbb{E}\left(\epsilon_i^2 \mid X_i\right)\right] \\ &\leq\sigma^2\mathbb{E}\left[\sum_{i=1}^{n}W_i(X)^2\right]\xrightarrow{n\to\infty}0 \end{aligned}

where the inequality follows from the conditional second-moment assumption in the regression set up, and the convergence follows from condition 4. Hence In→0I_n \to0 as required.

  1. Now it remains to show Jn→0J_n \to0

Jn≤E[(∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣)2]=E[(∑i=1n∣Wi(X)∣12⋅∣Wi(X)∣12∣m(Xi)−m(X)∣)2]\begin{aligned} J_n \leq\mathbb{E}\left[\left(\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|\right)^2\right] \\ &= \mathbb{E}\left[\left(\sum_{i=1}^{n}|W_i(X)|^{\frac{1}{2}}\cdot|W_i(X)|^{\frac{1}{2}}|m(X_i)-m(X)|\right)^2\right] \end{aligned}

By the Cauchy-Schwarz inequality and applying condition 1:

≤E[(∑i=1n∣Wi(X)∣)(∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣2)]≤D⋅E[∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣2]\begin{aligned} \leq\mathbb{E}\left[\left(\sum_{i=1}^{n}|W_i(X)|\right)\left(\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|^2\right)\right] \\ &\leq D\cdot\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|^2\right] \end{aligned}

We now split the expectation depending on whether XiX_i lies inside or outside the ball B(X,δ)B(X,\delta):

D⋅E[∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣2⋅1{d(Xi,X)>δ}]+D⋅E[∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣2⋅1{d(Xi,X)≤δ}]\begin{aligned} &D\cdot\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|^2\cdot\mathbf{1}_{\{d(X_i,X)>\delta\}}\right] \\ &\quad+D\cdot\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|^2\cdot\mathbf{1}_{\{d(X_i,X)\leq\delta\}}\right] \end{aligned}

The first expectation goes to zero by application of condition 2 and since ∣m(X)∣≤b|m(X)|\leq b:

D⋅E[∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣2⋅1{d(Xi,X)>δ}]≤4b2D⋅E[∑i=1n∣Wi(X)∣⋅1{d(Xi,X)>δ}]→n→∞0D\cdot\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|^2\cdot\mathbf{1}_{\{d(X_i,X)>\delta\}}\right] \leq4b^2D\cdot\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot\mathbf{1}_{\{d(X_i,X)>\delta\}}\right]\xrightarrow{n\to\infty}0

The second expression (for values inside the ball B(X,δ)B(X,\delta)) can be controlled because:

D⋅E[∑i=1n∣Wi(X)∣⋅∣m(Xi)−m(X)∣2⋅1{d(Xi,X)≤δ}]≤D2sup⁡{p1,p2∈S∗:d(p1,p2)≤δ}{∣m(p1)−m(p2)∣2}\begin{aligned} D\cdot\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot|m(X_i)-m(X)|^2\cdot\mathbf{1}_{\{d(X_i,X)\leq\delta\}}\right] \\ &\le D^2 \sup_{\{p_1,p_2\in S^*:d(p_1,p_2)\le\delta\}}\left\{|m(p_1)-m(p_2)|^2\right\} \end{aligned}

And since mm is continuous on a compact set, we know it is uniformly continuous. Thus:

sup⁡{p1,p2∈S∗:d(p1,p2)≤δ}{∣m(p1)−m(p2)∣2}→0as δ→0\sup_{\{p_1,p_2\in S^*:d(p_1,p_2)\le\delta\}}\left\{|m(p_1)-m(p_2)|^2\right\}\to0 \quad\text{as }\delta\to0

And so Jn→n→∞0J_n \xrightarrow{n\to\infty} 0 as required.

Thus, we have shown that In,Jn,Ln→0I_n,J_n,L_n\to0, completing the proof.

Proof for Theorem 4.2

PROOF. We just need to verify that the weights for the kernel estimator satisfy Stone’s conditions. Note that we use the following notation for scaling the kernel function:

Kh(x)=K(xh)K_h(x)=K\left(\frac{x}{h}\right)
  1. Since our weights are normalised in the definition this condition becomes trivial since for any nn and hnh_n:

∑i=1n∣Wi(X)∣=∑i=1nKhn(d(Xi,X))∑j=1nKhn(d(Xj,X))1A=1A≤1\sum_{i=1}^{n}|W_i(X)|=\frac{\sum_{i=1}^{n}K_{h_n}(d(X_i,X))}{\sum_{j=1}^{n}K_{h_n}(d(X_j,X))}\mathbb{1}_A=\mathbb{1}_A\le1
  1. By our condition on the kernel function we know:

0≤K(d(Xi,X)hn)≤1{d(Xi,X)hn∈[0,R]}=1{d(Xi,X)∈[0,Rhn]}0\le K\left(\frac{d(X_i,X)}{h_n}\right)\le1_{\left\{\frac{d(X_i,X)}{h_n}\in[0,R]\right\}}=\mathbb{1}_{\{d(X_i,X)\in[0,Rh_n]\}}

But since hn→n→∞0h_n\xrightarrow{n\to\infty}0, for any δ>0\delta>0 we can choose nn large enough so Rhn<δRh_n<\delta so:

1{d(Xi,X)>δ}=1⟹1{d(Xi,X)∈[0,Rhn]}=0\mathbb{1}_{\{d(X_i,X)>\delta\}}=1\Longrightarrow\mathbb{1}_{\{d(X_i,X)\in[0,Rh_n]\}}=0
⟹Khn(d(Xi,X))=0\Longrightarrow K_{h_n}(d(X_i,X))=0
⟹∣Wi(X)∣1{d(Xi,X)>δ}→n→∞0\Longrightarrow|W_i(X)|\mathbb{1}_{\{d(X_i,X)>\delta\}}\xrightarrow{n\to\infty}0

Finally by dominated convergence theorem we obtain the desired result:

lim⁡n→∞E[∑i=1n∣Wi(X)∣⋅1{d(Xi,X)>δ}]=E[lim⁡n→∞∑i=1n∣Wi(X)∣⋅1{d(Xi,X)>δ}]=0\begin{aligned} \lim_{n\to\infty}\mathbb{E}\left[\sum_{i=1}^{n}|W_i(X)|\cdot\mathbb{1}_{\{d(X_i,X)>\delta\}}\right] =\mathbb{E}\left[\lim_{n\to\infty}\sum_{i=1}^{n}|W_i(X)|\cdot\mathbb{1}_{\{d(X_i,X)>\delta\}}\right] \\ &=0 \end{aligned}
  1. Since our weights are normalised, they always sum to 1 as long as the sum of kernel values is non-zero. We will show that the probability of the kernel sum being 0 is also 0 in the limit:

P(∑i=1n∣Wi(X)∣≠1)=P(∑i=1nKhn(d(Xi,X))=0)≤P(∑i=1n1{Xi∈B(X,rhn)}=0)\begin{aligned} \mathbb{P}\left(\sum_{i=1}^{n}|W_i(X)|\ne1\right)=\mathbb{P}\left(\sum_{i=1}^{n}K_{h_n}(d(X_i,X))=0\right) \\ &\le\mathbb{P}\left(\sum_{i=1}^{n}\mathbb{1}_{\{X_i\in B(X,rh_n)\}}=0\right) \end{aligned}

For any fixed X=xX=x, we are demanding that there are no samples XiX_i in B(x,rhn)B(x,rh_n). Since the sample observations are i.i.d:

P(∑i=1n1{Xi∈B(X,rhn)}=0)=∫S∗(1−P(B(x,rhn)))n dP≤∫S∗exp⁡(−n⋅P(B(x,rhn))) dP\begin{aligned} \mathbb{P}\left(\sum_{i=1}^{n}\mathbf{1}_{\{X_i\in B(X,rh_n)\}}=0\right)=\int_{S_*}\left(1-\mathbb{P}(B(x,rh_n))\right)^n\,\mathrm{d}\mathbb{P} \\ &\leq\int_{S_*}\exp\left(-n\cdot\mathbb{P}(B(x,rh_n))\right)\,\mathrm{d}\mathbb{P} \end{aligned}
since ∀x∈[0,1]:(1−x)n≤e−nx=∫S∗n⋅P(B(x,rhn))n⋅P(B(x,rhn))exp⁡(−n⋅P(B(x,rhn))) dP≤∫S∗1n⋅P(B(x,rhn)) dP\begin{aligned} \text{since }\forall x\in[0,1]:(1-x)^n\leq e^{-nx} \\ &=\int_{S_*}\frac{n\cdot\mathbb{P}(B(x,rh_n))}{n\cdot\mathbb{P}(B(x,rh_n))}\exp\left(-n\cdot\mathbb{P}(B(x,rh_n))\right)\,\mathrm{d}\mathbb{P} \\ &\leq\int_{S_*}\frac{1}{n\cdot\mathbb{P}(B(x,rh_n))}\,\mathrm{d}\mathbb{P} \end{aligned}
since ∀x∈[0,1]:nxe−nx≤1\text{since }\forall x\in[0,1]:nxe^{-nx}\leq1

Finally, we can apply the small ball probability control Lemma A.1 and obtain:

1n∫S∗1P(B(x,rhn)) dP≤C(S∗,rhn2)n→n→∞0\frac{1}{n}\int_{S_*}\frac{1}{\mathbb{P}(B(x,rh_n))}\,\mathrm{d}\mathbb{P}\leq\frac{C\left(S_*,\frac{rh_n}{2}\right)}{n}\xrightarrow[n\to\infty]{}0
  1. Since K(x)≤1K(x)\leq1 we have:

∑i=1nWi(X)2=∑i=1n(Kh(d(Xi,X))2⋅1A2)(∑j=1nKh(d(Xj,X)))2≤1A∑i=1nKh(d(Xi,X))\sum_{i=1}^{n}W_i(X)^2=\frac{\sum_{i=1}^{n}\left(K_h(d(X_i,X))^2\cdot\mathbf{1}_A^2\right)}{\left(\sum_{j=1}^{n}K_h(d(X_j,X))\right)^2}\leq\frac{\mathbf{1}_A}{\sum_{i=1}^{n}K_h(d(X_i,X))}

Since the observations are i.i.d for a fixed X=xX=x, let Zn(x)=∑i=1n1Xi∈B(x,rhn)Z_n(x)=\sum_{i=1}^{n}\mathbf{1}_{X_i\in B(x,rh_n)} be a binomial random variable where pn(x)=P(B(x,rhn))p_n(x)=\mathbb{P}(B(x,rh_n)) is the probability of success. When Zn(X)>0Z_n(X)>0 we have:

∑i=1nWi(X)2≤1bZn(X)\sum_{i=1}^{n}W_i(X)^2\leq\frac{1}{bZ_n(X)}

since b⋅1[−r,r](x)≤K(x)b\cdot\mathbf{1}_{[-r,r]}(x)\leq K(x). When Zn(X)=0Z_n(X)=0:

∑i=1nWi(X)2≤(∑i=1nWi(X))2≤1\sum_{i=1}^{n}W_i(X)^2\leq\left(\sum_{i=1}^{n}W_i(X)\right)^2\leq1

since the weights are non-negative and sum to at most 1. Combining these two cases gives:

∑i=1nWi(X)2≤1{Zn(X)=0}+1b1{Zn(X)>0}Zn(X)\sum_{i=1}^{n}W_i(X)^2\leq\mathbf{1}_{\{Z_n(X)=0\}}+\frac{1}{b}\frac{\mathbf{1}_{\{Z_n(X)>0\}}}{Z_n(X)}

Taking the conditional expectation given X=xX=x and integrating over the distribution of XX for the expectation, we obtain:

E[∑i=1nWi(X)2]≤∫S∗[P(Zn(x)=0)+1bE(1{Zn(x)>0}Zn(x))]dP=∫S∗[(1−pn(x))n+1bE(1{Zn(x)>0}Zn(x))]dP\begin{aligned} \mathbb{E}\left[\sum_{i=1}^{n}W_i(X)^2\right]\leq\int_{S_*}\left[\mathbb{P}(Z_n(x)=0)+\frac{1}{b}\mathbb{E}\left(\frac{\mathbf{1}_{\{Z_n(x)>0\}}}{Z_n(x)}\right)\right]\mathrm{d}\mathbb{P} \\ &=\int_{S_*}\left[(1-p_n(x))^n+\frac{1}{b}\mathbb{E}\left(\frac{\mathbf{1}_{\{Z_n(x)>0\}}}{Z_n(x)}\right)\right]\mathrm{d}\mathbb{P} \end{aligned}
≤∫S∗[1npn(x)+2b(n+1)pn(x)] dPby xref‘LemmaA.2‘=(1n+2b(n+1))∫S∗1pn(x) dP\begin{aligned} &\le\int_{S^*}\left[\frac{1}{n p_n(x)}+\frac{2}{b(n+1)p_n(x)}\right]\,\mathrm{d}\mathbb{P} \quad\text{by } {xref}`Lemma A.2` \\ &= \left(\frac{1}{n}+\frac{2}{b(n+1)}\right)\int_{S^*}\frac{1}{p_n(x)}\,\mathrm{d}\mathbb{P} \end{aligned}

Again applying the small ball probability control Lemma A.1 we obtain the desired result:

(1n+2b(n+1))∫S∗1pn(x) dP≤(1+2b)C(S∗,rhn2)n→n→∞0\left(\frac{1}{n}+\frac{2}{b(n+1)}\right)\int_{S^*}\frac{1}{p_n(x)}\,\mathrm{d}\mathbb{P} \le\left(1+\frac{2}{b}\right)\frac{C\left(S^*,\frac{r h_n}{2}\right)}{n}\xrightarrow[n\to\infty]{}0

References

  1. [1]ABRAHAM, C., BIAU, G. and CADRE, B. (2006). On the Kernel Rule for Function Classification. Annals of the Institute of Statistical Mathematics 58 619–633. https://doi.org/10.1007/s10463-006-0032-1
  2. [2]ADAMS, H., EMERSON, T., KIRBY, M., NEVILLE, R., PETERSON, C., SHIPMAN, P., CHEPUSHTANOVA, S., HANSON, E., MOTTA, F. and ZIEGELMEIER, L. (2017). Persistence images: A stable vector representation of persistent homology. Journal of Machine Learning Research 18 1–35.
  3. [3]AKTAS MEHMET E, F. A. E. AKBAS ESRA (2019). Persistence homology of networks: methods and applications. Applied Network Science. https://doi.org/10.1007/s41109-019-0179-3
  4. [4]BAUER, U. (2021). Ripser: efficient computation of Vietoris-Rips persistence barcodes. J. Appl. Comput. Topol. 5 391–423. https://doi.org/10.1007/s41468-021-00071-5 MR4298669
  5. [5]BIAU, G., CEROU, F. and GUYADER, A. (2010). Rates of Convergence of the Functional k-Nearest Neighbor Estimate. IEEE Transactions on Information Theory 56 2034–2040. https://doi.org/10.1109/TIT.2010.2040857
  6. [6]BUBENIK, P. (2015). Statistical topological data analysis using persistence landscapes. Journal of Machine Learning Research 16 77–102.
  7. [7]BUBENIK, P., HULL, M., PATEL, D. and WHITTLE, B. (2020). Persistent homology detects curvature. Inverse Problems 36 025008. https://doi.org/10.1088/1361-6420/ab4ac0
  8. [8]BUBENIK, P. and VERGILI, T. (2018). Topological spaces of persistence modules and their properties. Journal of Applied and Computational Topology 2 233–269. https://doi.org/10.1007/s41468-018-0022-4
  9. [9]CARRIÈRE, M., CUTURI, M. and OUDOT, S. (2017). Sliced Wasserstein kernel for persistence diagrams. In International Conference on Machine Learning 664–673. PMLR.arxiv.org/abs/1706.03358
  10. [10]CARRIÈRE, M. and BAUER, U. (2019). On the Metric Distortion of Embedding Persistence Diagrams into Separable Hilbert Spaces. LIPIcs, Volume 129, SoCG 2019 129 21:1–21:15. https://doi.org/10.4230/LIPIcs.SOCG.2019.21
  11. [11]CÉROU, F. and GUYADER, A. (2006). Nearest neighbor classification in infinite dimension. ESAIM: Probability and Statistics 10 340–355. https://doi.org/10.1051/ps:2006014 MR2274925
  12. [12]CHEPUSHTANOVA, S., EMERSON, T., HANSON, E. M., KIRBY, M., MOTTA, F. C., NEVILLE, R., PETERSON, C., SHIPMAN, P. D. and ZIEGELMEIER, L. (2015). Persistence Images: An Alternative Persistent Homology Representation. Corr abs/1507.06217.
  13. [13]COHEN-STEINER, D., EDELSBRUNNER, H. and HARER, J. (2007). Stability of persistence diagrams. Discrete & Computational Geometry 37 103–120.
  14. [14]COHEN-STEINER, D., EDELSBRUNNER, H., HARER, J. and MILEYKO, Y. (2010). Lipschitz functions have Lp-stable persistence. Foundations of Computational Mathematics 10 127–139.
  15. [15]CRAWLEY-BOEVEY, W. (2015). Decomposition of pointwise finite-dimensional persistence modules. Journal of Algebra and Its Applications 14 1550066. https://doi.org/10.1142/S0219498815500668
  16. [16]DABO-NIANG, S. and RHOMARI, N. (2003). Estimation non paramétrique de la régression avec variable explicative dans un espace métrique. Comptes Rendus Mathematique 336 75-80. https://doi.org/10.1016/S1631-073X(02)00012-2
  17. [17]DEVROYE, L. P. and WAGNER, T. J. (1980). Distribution-Free Consistency Results on Nonparametric Discrimination and Regression Function Estimation. The Annals of Statistics 8 231 – 239. https://doi.org/10.1214/aos/1176344949
  18. [18]EDELSBRUNNER, H. and HARER, J. (2010). Computational Topology: An Introduction. https://doi.org/10.1007/978-3-540-33259-6_7
  19. [19]FASY, B. T., LECCI, F., RINALDO, A., WASSERMAN, L., BALAKRISHNAN, S. and SINGH, A. (2014). Confidence sets for persistence diagrams. The Annals of Statistics 42 2301–2339.
  20. [20]Ferraty, F. and Vieu, P. (2006). Nonparametric Functional Data Analysis: Theory and Practice 51. https://doi.org/10.1007/0-387-36620-2
  21. [21]Forzani, L., Fraiman, R. and Llоp, P. (2012). Consistent Nonparametric Regression for Functional Data Under the Stone–Besicovitch Conditions. IEEE Transactions on Information Theory 58 6697–6708. https://doi.org/10.1109/TIT.2012.2209628
  22. [22]Györfi, L., Kohler, M., Krzyżak, A. and Walk, H. (2002). A Distribution-Free Theory of Nonparametric Regression. Springer series in statistics. Springer.DOI
  23. [23]Jakob, J., Gross, M. and Günther, T. (2021). A Fluid Flow Data Set for Machine Learning and its Application to Neural Flow Map Interpolation. IEEE Transactions on Visualization and Computer Graphics 27 1279–1289. https://doi.org/10.1109/TVCG.2020.3028947
  24. [24]Kpotufe, S. (2009). Fast, smooth and adaptive regression in metric spaces. In Advances in Neural Information Processing Systems 22 1024–1032.
  25. [25]Kulkarni, S. R. and Posner, S. E. (1995). Rates of convergence of nearest neighbor estimation under arbitrary sampling. IEEE Transactions on Information Theory 41 1028–1039. https://doi.org/10.1109/18.391248
  26. [26]Kwitt, R., Huber, S., Niethammer, M., Lin, W. and Bauer, U. (2015). Statistical topological data analysis – A kernel perspective. In Advances in Neural Information Processing Systems 28 3070–3078.
  27. [27]Mileyko, Y., Mukherjee, S. and Harer, J. (2011). Probability measures on the space of persistence diagrams. Inverse Problems 27 124007. https://doi.org/10.1088/0266-5611/27/12/124007
  28. [28]Nadaraya, E. A. (1964). On Estimating Regression. Theory of Probability & Its Applications 9 141–142. https://doi.org/10.1137/1109020
  29. [29]Obayashi, I., Hiraoka, Y. and Kimura, M. (2018). Persistence diagrams with linear machine learning models. Journal of Applied and Computational Topology.arxiv.org/abs/1706.10082
  30. [30]Perea, J., Munch, E. and Khasawneh, F. (2022). Approximating Continuous Functions on Persistence Diagrams Using Template Functions. Foundations of Computational Mathematics. https://doi.org/10.48550/arXiv.1902.07190
  31. [31]Popinet, S. (2004). Free Computational Fluid Dynamics. ClusterWorld 2.
  32. [32]Project, T. G. (2023). GUDHI – Geometry Understanding in Higher Dimensions. https://gudhi.inria.fr/.
  33. [33]Pun, C. S., Xia, K. and Lee, S. (2022). Persistent-Homology-based Machine Learning and its Applications – A Survey. Artificial Intelligence Review. https://doi.org/10.48550/arXiv.1811.00252
  34. [34]Reininghaus, J., Huber, S., Bauer, U. and Kwitt, R. (2015). A stable multi-scale kernel for topological machine learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 4741–4748.arxiv.org/abs/1412.6821
  35. [35]Robinson, A. and Turner, K. (2017). Hypothesis Testing for Topological Data Analysis. Journal of Applied and Computational Topology 1 241–261. https://doi.org/10.1007/s41468-017-0008-7
  36. [36]Spiegelman, C. and Sacks, J. (1980). Consistent Window Estimation in Nonparametric Regression. The Annals of Statistics 8 240–246. https://doi.org/10.1214/aos/1176344950
  37. [37]Stone, C. J. (1977). Consistent Nonparametric Regression. The Annals of Statistics 5 595 – 620. https://doi.org/10.1214/aos/1176343886
  38. [38]Stone, C. J. (1982). Optimal global rates of convergence for nonparametric regression. The Annals of Statistics 10 1040–1053.DOI
  39. [39]Turkeš, R., Montúfar, G. and Otter, N. (2023). On the effectiveness of persistent homology. https://arxiv.org/abs/2206.10551.
  40. [40]Turner, K., Mileyko, Y., Mukherjee, S. and Harer, J. (2012). Frechet Means for Distributions of Persistence Diagrams. Discrete and Computational Geometry 52. https://doi.org/10.1007/s00454-014-9604-7
  41. [41]Turner, K. and Spreemann, G. (2019). Same But Different: Distance Correlations Between Topological Summaries. arXiv preprint arXiv:1903.01051. https://doi.org/10.48550/arXiv.1903.01051
  42. [42]Watson, G. S. (1964). Smooth regression analysis. Sankhyā: The Indian Journal of Statistics, Series A 359–372.

Paper details

vlnc:W_3zrpwwxj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents