Towards a Scalable Reference-Free Evaluation of Generative Models

Abstract

While standard evaluation scores for generative models are mostly reference-based, a reference-dependent assessment of generative models could be generally difficult due to the unavailability of applicable reference datasets. Recently, the reference-free entropy scores, VENDI [1] and RKE [2], have been proposed to evaluate the diversity of generated data. However, estimating these scores from data leads to significant computational costs for large-scale generative models. In this work, we leverage the random Fourier features framework to reduce the computational price and propose the Fourier-based Kernel Entropy Approximation (FKEA) method. We utilize FKEA’s approximated eigenspectrum of the kernel matrix to efficiently estimate the mentioned entropy scores. Furthermore, we show the application of FKEA’s proxy eigenvectors to reveal the method’s identified modes in evaluating the diversity of produced samples. We provide a stochastic implementation of the FKEA assessment algorithm with a complexity $O(n)$ linearly growing with sample size $n$. We extensively evaluate FKEA’s numerical performance in application to standard image, text, and video datasets. Our empirical results indicate the method’s scalability and interpretability applied to large-scale generative models. The codebase is available at https://github.com/aziksh-ospanov/FKEA.

1 Introduction

A quantitative comparison of generative models requires evaluation metrics to measure the quality and diversity of the models’ produced data. Since the introduction of variational autoencoders (VAEs) [3], generative adversarial networks (GANs) [4], and diffusion models [5] that led to impressive empirical results in the last decade, several evaluation scores have been proposed to assess generative models learned by different training methods and architectures. Due to the key role of evaluation criteria in comparing generative models, they have been extensively studied in the literature.

While various statistical methods have been applied to measure the fidelity and variety of a generative model’s produced data, the standard scores commonly perform a reference-based evaluation of generative models, i.e., they quantify the characteristics of generated samples in comparison to a reference distribution. The reference distribution is usually chosen to be either the distribution of samples in the test data partition or a comprehensive dataset containing a significant fraction of real-world sample types, e.g. ImageNet [6] for evaluating image-based generative models.

To provide well-known examples of reference-dependent metrics, note that the distance scores, Fréchet Inception Distance (FID) [7] and Kernel Inception Distance (KID) [8], are explicitly reference-based, measuring the distance between the generative and reference distributions. Similarly, the standard quality/diversity score pairs, Precision/Recall [9, 10] and Density/Coverage [11], perform the evaluation in comparison to a reference dataset. Even the seemingly reference-free Inception Score (IS) [12] can be viewed as implicitly reference-based, since it quantifies the variety and fidelity of data based on the labels and confidence scores assigned by an ImageNet pre-trained neural net, where ImageNet implicitly plays the role of the reference dataset. The reference-based nature of these evaluation scores is desired in many instances including standard image-based generative models, where either a sufficiently large test set or a comprehensive reference dataset such as ImageNet is available for the reference-based evaluation.

On the other hand, a reference-based assessment of generative models may not always be feasible, because the selection of a reference distribution may be challenging in a general learning scenario. For example, in prompt-based generative models where the data are created in response to a user’s input text prompts, the generated data could follow an a priori unknown distribution depending on the specific distribution of the user’s input prompts. Figure 1 illustrates text prompt generated colorful elephant images. Each elephant image corresponds to an unusual color choice, such as luminous yellow or neon pink. Even though reference dataset, such as ImageNet [6] contains elephants, additional elements such as neon colors are not attributed to those images. A proper reference-based evaluation of every user’s generated data would require a distinct reference dataset, which may not be available to the user during the assessment time. Moreover, finding a comprehensive text or video dataset to choose as the reference set would be more difficult compared to image data, because the higher length of text and video samples could significantly contribute to their variety, requiring an inefficiently large reference set to cover all text or video sample types.

Figure 1: Reference-free diversity evaluation of "colorful elephant" samples generated by text-to-image StableDiffusion [13]. FKEA’s top-4 sample clusters and approximated VENDI and RKE are shown.

The discussed challenging scenarios of conducting a reference-based evaluation highlight the need for reference-free assessment methods that remain functional in the absence of a reference dataset. Recently, entropy-based diversity evaluation scores, the $\mathrm{VENDI}$ metric family [1, 14] and $\mathrm{RKE}$ score [2], have been proposed to address the need for reference-free assessment metrics. These scores calculate the entropy of the eigenvalues of a kernel similarity matrix for the generated data. Based on the theoretical results in [2], the evaluation process of these scores can be interpreted as an unsupervised identification of the generative model’s produced sample clusters, followed by the entropy calculation for the frequencies of the detected clusters.

While the $\mathrm{VENDI}$ and $\mathrm{RKE}$ entropy scores provide reference-free assessments of generative models, estimating these scores from generated data could incur significant computational costs. In this work, we show that computing the precise $\mathrm{RKE}$ and $\mathrm{VENDI}$ scores would require at least $\Omega(n^{2})$ and $\Omega(n^{2.376})$ computations for a sample size $n$, respectively. While the randomized projection methods in [15, 1] can reduce the computational costs to $O(n^{2})$ for a general $\mathrm{VENDI}_{\alpha}$ score, the quadratic growth would be a barrier to the method’s application to large $n$ values. Although the computational expenses could be reduced by limiting the sample size $n$, an insufficient sample size would lead to significant error in estimating the entropy scores.

To overcome the challenges of computing the scores, we leverage the random Fourier features (RFFs) framework [16] and develop a scalable entropy-based evaluation method that can be efficiently applied to large sample sizes. Our proposed method, Fourier-based Kernel Entropy Approximation (FKEA), is designed to approximate the kernel covariance matrix using the RFFs drawn from the Fourier transform-inverse of a target shift-invariant kernel. We prove that using a Fourier feature size $r=\mathcal{O}\bigl{(}\frac{\log n}{\epsilon^{2}}\bigr{)}$, FKEA computes the eigenspace of the kernel matrix within an $\epsilon$-bounded error. Furthermore, we demonstrate the application of the eigenvectors of the FKEA’s proxy kernel matrix for identifying the sample clusters used in the reference-free evaluation of entropic diversity.

Finally, we present numerical results of the entropy-based evaluation of standard generative models using the baseline eigendecomposition and our proposed FKEA methods. In our experiments, the baseline spectral decomposition algorithm could not efficiently scale to sample sizes above a few ten thousand. On the other hand, our stochastic implementation of the FKEA method could scalably apply to large sample sizes. Utilizing the standard embeddings of image, text, and video data, we tested the FKEA assessment while computing the sample clusters and their frequencies in application to large-scale datasets and generative models. Here is a summary of our work’s main contributions:

• Characterizing the computational complexity of the kernel entropy scores of generative models,

• Developing the Fourier-based FKEA method to approximate the kernel covariance eigenspace and entropy of generated data,

• Proving guarantees on FKEA’s required size of random Fourier features indicating a complexity logarithmically growing with the dataset size,

• Providing numerical results on FKEA’s reference-free assessment of large-scale image,text, video-based datasets and generative models.

2 Related Work

Evaluation of deep generative models. The assessment of generative models has been widely studied in the literature. The existing scores either quantify a distance between the distributions of real and generated data, as in FID [7] and KID [8] scores, or attempt to measure the quality and diversity of the trained generative models, including the Inception Score [12], quality/diversity metric pairs Precision/Recall [9, 10] and Density/Coverage [11]. The mentioned scores are reference-based, while in this work we focus on reference-free metrics. Also, we note that the evaluation of memorization and novelty has received great attention, and several scores including the authenticity score [17], the feature likelihood divergence [18], and the rarity score [19] have been proposed to quantify the generalizability and novelty of generated samples. Note that the evaluation of novelty and generalization is, by nature, reference-based. On the other hand, our study focuses on the diversity of data which can be evaluated in a reference-free way as discussed in [1, 2].

Role of embedding in quantitative evaluation. Following the discussion in [20], we utilize DinoV2 [21] image embeddings in most of our image experiments, as [20]’s results indicate DinoV2 can yield scores more aligned with the human notion of diversity. As noted in [22], it is possible to utilize other non-ImageNet feature spaces such as CLIP [23] and SwAV [24] as opposed to InceptionV3[25] to further improve metrics such as FID. In this work, we mainly focus on DinoV2 feature space, while we note that other feature spaces are also compatible with entropy-based diversity evaluation.

Diversity assessment for text-based models. To quantify the diversity of text data, the n-gram-based methods are commonly used in the literature. A well-known metric is the BLEU score [26], which is based on the geometric average of n-gram precision scores times the Brevity Penalty. To adapt BLEU score to measure text diversity, [27] proposes the Self-BLEU score, calculating the average BLEU score of various generated samples. To further isolate and measure diversity, N-Gram Diversity scores [28, 29, 30] were proposed and defined by a ratio between the number of unique n-grams and overall number of n-grams in the text. Other prominent metrics include Homogenization (ROUGE-L) [31], FBD [32] and Compression Ratios [33].

Kernel PCA, Spectral Cluttering, and Random Fourier Features. Kernel PCA [34] is a well-studied method of dimensionality reduction that utilizes the eigendecomposition of the kernel matrix, similar to the kernel-based diversity evaluation methods in [1, 2]. The related papers [35, 36] study the connections between kernel PCA and spectral clustering. Also, the analysis of random Fourier features [16] for performing scalable kernel PCA has been studied in [37, 38, 39, 40, 41]. We note that while the mentioned works characterize the complexity of estimating the eigenvectors, our analysis focuses on the complexity of computing the kernel matrix’s eigenvalues via Fourier features, as we primarily seek to quantify the diversity of generated data using the kernel matrix’s eigenvalues.

3 Preliminaries

Consider a generative model $\mathcal{G}$ generating random samples $\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\in\mathbb{R}^{d}$ following the model’s probability distribution $P_{\mathcal{G}}$. In our analysis, we assume the $n$ generated samples are independently drawn from $P_{\mathcal{G}}$. Note that in VAEs [3] and GANs [4], the generative model $\mathcal{G}$ is a deterministic function $G:\mathbb{R}^{r}\rightarrow\mathbb{R}^{d}$ mapping an $r$-dimensional latent random vector $\mathbf{Z}\sim P_{Z}$ from a known distribution $P_{Z}$ to $G(\mathbf{Z})$ distributed according to $P_{\mathcal{G}}$. On the other hand, in diffusion models, $\mathcal{G}$ represents an iterative random process that generates a sample from $P_{\mathcal{G}}$. The goal of a sample-based diversity evaluation of generative model $\mathcal{G}$ is to quantify the variety of its generated data $\mathbf{x}_{1},\ldots,\mathbf{x}_{n}$.

3.1 Kernel Function, Kernel Covariance Matrix, and Matrix-based Rényi Entropy

Following standard definitions, $k:\mathbb{R}^{d}\times\mathbb{R}^{d}\rightarrow\mathbb{R}$ is called a kernel function if for every integer $n\in\mathbb{N}$ and set of inputs $\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\in\mathbb{R}^{d}$, the kernel similarity matrix $K=\bigl{[}k(\mathbf{x}_{i},\mathbf{x}_{j})\bigr{]}_{n\times n}$ is positive semi-definite. We call a kernel function $k$ normalized if for every input $\mathbf{x}$ we have $k(\mathbf{x},\mathbf{x})=1$. A well-known example of a normalized kernel function is the Gaussian kernel $k_{\text{\rm Gaussian}(\sigma^{2})}$ with bandwidth parameter $\sigma^{2}$ defined as

$$k_{\text{\rm Gaussian}(\sigma^{2})}\bigl{(}\mathbf{x},\mathbf{x}^{\prime}\bigr{)}\,:=\,\exp\Bigl{(}-\frac{\bigl{\|}\mathbf{x}-\mathbf{x}^{\prime}\bigr{\|}^{2}_{2}}{2\sigma^{2}}\Bigr{)}$$

For every kernel function $k$, there exists a feature map $\phi:\mathbb{R}^{d}\rightarrow\mathbb{R}^{m}$ such that $k(\mathbf{x},\mathbf{x}^{\prime})=\langle\phi(\mathbf{x}),\phi(\mathbf{x}^{\prime})\rangle$ is the inner product of the $m$-dimensional feature maps $\phi(\mathbf{x})$ and $\phi(\mathbf{x}^{\prime})$. Given a kernel $k$ with feature map $\phi$, we define the kernel covariance matrix ${C}_{X}\in\mathbb{R}^{m\times m}$ of a distribution $P_{X}$ as

$${C}_{X}\,:=\,\mathbb{E}_{\mathbf{X}\sim P_{X}}\Bigl{[}\phi(\mathbf{X})\phi(\mathbf{X})^{\top}\Bigr{]}\,=\,\int p_{X}(\mathbf{x})\phi(\mathbf{x})\phi(\mathbf{x})^{\top}\mathrm{d}\mathbf{x}$$

The above matrix $C_{X}$ is positive semi-definite with non-negative values. Furthermore, assuming a normalized kernel $k$, it can be seen that the eigenvalues of $C_{X}$ will add up to $1$ (i.e., it has unit trace $\mathrm{Tr}(C_{X})=1$), providing a probability model. Therefore, one can consider the entropy of $C_{X}$’s eigenvalues as a quantification of the diversity of distribution $P_{X}$ based on the kernel similarity score $k$. Here, we review the general family of Rényi entropy used to define $\mathrm{VENDI}$ and $\mathrm{RKE}$ scores.

Definition 1.

For a positive semi-definite matrix $C_{X}\in\mathbb{R}^{m\times m}$ with eigenvalues $\lambda_{1},\ldots,\lambda_{m}$, the order-$\alpha$ Rényi entropy $H_{\alpha}(C_{X})$ for $\alpha>0$ is defined as

$$H_{\alpha}(C_{X})\,:=\,\frac{1}{1-\alpha}\log\Bigl{(}\,\sum_{i=1}^{m}\lambda_{i}^{\alpha}\,\Bigr{)}$$

To estimate the entropy scores from finite empirical samples $\mathbf{x}_{1},\ldots,\mathbf{x}_{n}$, we consider the empirical kernel covariance matrix $\widehat{C}_{X}$ defined as $\widehat{C}_{X}\,:=\,\frac{1}{n}\sum_{i=1}^{n}\phi\bigl{(}\mathbf{x}_{i}\bigr{)}\phi\bigl{(}\mathbf{x}_{i}\bigr{)}^{\top}$. This matrix provides an empirical estimation of the population kernel covariance matrix ${C}_{X}$.

It can be seen that the $m\times m$ empirical matrix $\widehat{C}_{X}$ and normalized kernel matrix $\frac{1}{n}K=\frac{1}{n}\bigl{[}k(\mathbf{x}_{i},\mathbf{x_{j}})\bigr{]}_{n\times n}$ share the same non-zero eigenvalues. Therefore, to compute the matrix-based entropy of the empirical covariance matrix $\widehat{C}_{X}$, one can equivalently compute the entropy of the eigenvalues of the kernel similarity matrix $K$. This approach results in the definition of the $\mathrm{VENDI}$ and $\mathrm{RKE}$ diversity scores: [1] defines the family of $\mathrm{VENDI}$ scores as

$$\mathrm{VENDI}_{\alpha}(\mathbf{x}_{1},\ldots,\mathbf{x}_{n})\,:=\,\exp\Bigl{(}H_{\alpha}\bigl{(}\frac{1}{n}K\bigr{)}\Bigr{)}\,=\,\Bigl{(}\sum_{i=1}^{n}\lambda^{\alpha}_{i}\Bigr{)}^{\frac{1}{1-\alpha}},$$

where $\lambda_{1},\ldots,\lambda_{n}$ denote the eigenvalues of the kernel matrix $\frac{1}{n}K$. Also, [2] proposes the $\mathrm{RKE}$ score, which is the special order-2 Renyi entropy, $\mathrm{RKE}(\mathbf{x}_{1},\ldots,\mathbf{x}_{n})=\exp(H_{2}(\frac{1}{n}K))$. To compute $\mathrm{RKE}$ without computing the eigenvalues, [2] points out the RKE score reduces to the Frobenius norm $\|\cdot\|_{F}$ of the kernel matrix as follows:

$$\mathrm{RKE}(\mathbf{x}_{1},\ldots,\mathbf{x}_{n})\,=\,\Bigl{\|}\frac{1}{n}K\Bigr{\|}^{-2}_{F}\,=\,\Bigl{(}\,\frac{1}{n^{2}}\sum_{i=1}^{n}\sum_{j=1}^{n}k\bigl{(}\mathbf{x}_{i},\mathbf{x}_{j}\bigr{)}^{2}\,\Bigr{)}^{-1}$$

3.2 Shift-Invariant Kernels and Random Fourier Features

A kernel function $k$ is called shift-invariant, if there exists a function $\kappa:\mathbb{R}^{d}\rightarrow\mathbb{R}$ such that $k(\mathbf{x},\mathbf{x}^{\prime})=\kappa(\mathbf{x}-\mathbf{x}^{\prime})$ for every $\mathbf{x},\mathbf{x}^{\prime}\in\mathbb{R}^{d}$. Bochner’s theorem proves that a function $\kappa:\mathbb{R}^{d}\rightarrow\mathbb{R}$ will lead to a shift-invariant kernel similarity score $\kappa(x-\mathbf{x}^{\prime})$ between $\mathbf{x},\mathbf{x}^{\prime}$ if and only if its Fourier transform $\widehat{\kappa}:\mathbb{R}^{d}\rightarrow\mathbb{R}$ is non-negative everywhere (i.e, $\widehat{\kappa}(\bm{\omega})\geq 0$ for every $\bm{\omega}$). Note that the Fourier transform $\widehat{\kappa}$ is defined as

$$\widehat{\kappa}(\bm{\omega})\,:=\,\frac{1}{(2\pi)^{d}}\int{\kappa}(\mathbf{x})\exp\bigl{(}-i\bm{\omega}^{\top}\mathbf{x}\bigr{)}\mathrm{d}\mathbf{x}$$

Specifically, Bochner’s theorem shows the Fourier transform $\widehat{\kappa}$ of a normalized shift-invariant kernel $k(\mathbf{x},\mathbf{x}^{\prime})=\kappa(\mathbf{x}-\mathbf{x}^{\prime})$, where $\kappa(0)=1$, will be a probability density function (PDF). The framework of random Fourier features (RFFs) [16] utilizes independent samples drawn from PDF $\widehat{\kappa}$ to approximate the kernel function. Here, given independent samples $\bm{\omega}_{1},\ldots,\bm{\omega}_{r}\sim\widehat{\kappa}$, we form the following proxy feature map $\widetilde{\phi}_{r}:\mathbb{R}^{d}\rightarrow\mathbb{R}^{2r}$

$$\widetilde{\phi}_{r}(\mathbf{x})\,=\,\frac{1}{\sqrt{r}}\Bigl{[}\cos(\bm{\omega}_{1}^{\top}\mathbf{x}),\sin(\bm{\omega}_{1}^{\top}\mathbf{x}),\ldots,\cos(\bm{\omega}_{r}^{\top}\mathbf{x}),\sin(\bm{\omega}_{r}^{\top}\mathbf{x})\Bigr{]}.$$ (1)

As demonstrated in [16, 42], the $2r$-dimensional proxy map $\widetilde{\phi}_{r}$ can approximate the kernel function as $k(\mathbf{x},\mathbf{x}^{\prime})=\mathbb{E}_{\bm{\omega}\sim\widehat{\kappa}}\Bigl{[}\cos(\bm{\omega}^{\top}\mathbf{x})\cos(\bm{\omega}^{\top}\mathbf{x}^{\prime})+\sin(\bm{\omega}^{\top}\mathbf{x})\sin(\bm{\omega}^{\top}\mathbf{x}^{\prime})\Bigr{]}\approx\widetilde{\phi}_{r}(\mathbf{x})^{\top}\widetilde{\phi}_{r}(\mathbf{x}^{\prime})$.

4 Computational Complexity of VENDI & RKE Scores

As discussed, computing $\mathrm{RKE}$ and general $\mathrm{VENDI}_{\alpha}$ scores requires computing the order-$\alpha$ entropy of kernel matrix $\frac{1}{n}K$. Using the standard definition of $\alpha$-norm $\|\mathbf{v}\|_{\alpha}=\bigl{(}\sum_{i=1}^{n}|v_{i}|^{\alpha}\bigr{)}^{1/\alpha}$, we observe that the computation of $\mathrm{VENDI}_{\alpha}$ score is equivalent to computing the $\alpha$-norm $\|\bm{\lambda}\|_{\alpha}$ of the $n$-dimensional eigenvalue vector $\bm{\lambda}=[\lambda_{1},\ldots,\lambda_{n}]$ where $\lambda_{1}\geq\cdots\geq\lambda_{n}$ are the sorted eigenvalues of the normalized kernel matrix $\frac{1}{n}K$.

In the following theorem, we prove that except order $\alpha=2$, which is the $\mathrm{RKE}$ score, computing any other $\mathrm{VENDI}_{\alpha}$ score is at least as expensive as computing the product of two $n\times n$ matrices. Therefore, the theorem suggests that the computational complexity of every member of the VENDI family is lower-bounded by $O(n^{2.367})$ which is the least known cost of multiplying $n\times n$ matrices.

In the theorem, we suppose $\mathcal{B}$ is any fixed set of “basis” functions. A circuit $\mathcal{C}$ is a directed acyclic graph each of whose internal nodes is labeled by a gate coming from a set $\mathcal{B}$. A subset of gates are designated as outputs of $\mathcal{C}$. A circuit with $n$ source nodes and $m$ outputs computes a function from $\mathbb{R}^{n}$ to $\mathbb{R}^{m}$ by evaluating the gate at each internal gate in topological order. The size of a circuit is the number of gates. Also, $\nabla\mathcal{B}$ is the basis consisting of the gradients of all functions in $\mathcal{B}$. We will provide the proof of the theorems in the Appendix.

Theorem 1.

If $\mathrm{VENDI}_{\alpha}(K)$ for $\alpha\neq 2$ is computable by a circuit $\mathcal{C}$ of size $s(n)$ over basis $\mathcal{B}$, then $n\times n$ matrices can be multiplied by a circuit $\mathcal{C}$ of size $O(s(n))$ over basis $\nabla\mathcal{B}|\cup\{+,\times\}$.

Remark 1.

The smallest known circuits for multiplying $n\times n$ matrices have size $\Theta(n^{\omega})$, where $\omega\approx 2.367$. Despite tremendous research efforts only minor improvements have been obtained in recent years. There is evidence that $\omega$ is bounded away from 2 for certain classes of circuits [43, 44]. In contrast, $\mathcal{S}_{2}$ is computable in quadratic time $\Theta(n^{2})$ in the basis $B=\{\times,+,\log\}$.

The above discussion indicates that except the $\mathrm{RKE}(\mathbf{x}_{1},\ldots,\mathbf{x}_{n})$, i.e. order-2 Renyi entropy, whose computational complexity is quadratically growing with sample size $\Theta(n^{2})$, the other members of the VENDI family $\mathrm{VENDI}_{\alpha}$ would have a super-quadratic complexity on the order of $\mathcal{O}(n^{2.367})$. In practice, the computation of $\mathrm{VENDI}_{\alpha}$ scores is performed by the eigendecomposition of the $n\times n$ kernel matrix that requires $O(n^{3})$ computations for precise computation and $O(n^{2}M)$ computations using a randomized projection onto an $M$-dimensional space [15, 1].

5 A Scalable Fourier-based Method for Computing Kernel Entropy Scores

As we showed earlier, the complexity of computing $\mathrm{RKE}$ and $\mathrm{VENDI}$ scores are at least quadratically growing with the sample size $n$. The super-linear growth of the scores’ complexity with sample size $n$ can hinder their application to large-scale datasets and generative models with potentially hundreds of sample types. In such cases, a proper entropy estimation should be performed over potentially hundreds of thousands of data, where the quadratic complexity of the scores would be a significant barrier toward their accurate estimation.

Here, we conisider a shift-invariant kernel matrix $k(\mathbf{x},\mathbf{x}^{\prime})=\kappa(\mathbf{x}-\mathbf{x}^{\prime})$ where $\kappa(\mathbf{0})=1$ and propose applying the random Fourier features (RFF) framework [16] to perform an efficient approximation of the RKE and VENDI scores. To do this, we utilize the Fourier transform $\widehat{\kappa}$ that, according to Bochner’s theorem, is a valid PDF, and we independently generate $\bm{\omega}_{1},\ldots,\bm{\omega}_{r}\stackrel{{\scriptstyle\text{\rm iid}}}{{\sim}}\widehat{\kappa}$. Note that in the case of the Gaussian kernel $k_{\text{\rm Gaussian}(\sigma^{2})}$, the corresponding PDF will be an isotropic Gaussian $\mathcal{N}(\mathbf{0},\frac{1}{\sigma^{2}}I_{d})$ with zero mean and covariance matrix $\frac{1}{\sigma^{2}}I_{d}$. Then, we consider the RFF proxy feature map $\widetilde{\phi}_{r}:\mathbb{R}^{d}\rightarrow\mathbb{R}^{2r}$ as defined in (1) and define the proxy kernel covariance matrix $\widetilde{C}_{X,r}\in\mathbb{R}^{2r\times 2r}$:

$$\widetilde{C}_{X,r}\,=\,\frac{1}{n}\sum_{i=1}^{n}\,\widetilde{\phi}_{r}\bigl{(}\mathbf{x}_{i}\bigr{)}\,\widetilde{\phi}_{r}\bigl{(}\mathbf{x}_{i}\bigr{)}^{\top}$$ (2)

Note that the $2r\times 2r$ matrix $\widehat{C}_{X,r}$ has the same non-zero eigenvalues as the $n\times n$ RFF proxy kernel matrix $\frac{1}{n}\widetilde{K}_{r}$, and therefore can be utilized to approximate the eigenvalues of the original $n\times n$ kernel matrix $\frac{1}{n}K$. Therefore, we propose the Fourier-based Kernel Entropy Approximation (FKEA) method to approximate the $\mathrm{RKE}$ and $\mathrm{VENDI}_{\alpha}$ scores as follows:

$$\mathrm{FKEA}\text{-}\mathrm{RKE}\bigl{(}\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\bigr{)}\, =\,\exp\bigl{(}H_{2}(\widetilde{C}_{X,r})\bigr{)}\,=\,\bigl{\|}\widetilde{C}_{X,r}\bigr{\|}_{F}^{-2},$$ (3)

$$\mathrm{FKEA}\text{-}\mathrm{VENDI}_{\alpha}\bigl{(}\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\bigr{)}\, =\,\exp\bigl{(}H_{\alpha}(\widehat{C}_{X,r})\bigr{)}\,=\,\Bigl{(}\,\sum_{i=1}^{2r}\,{\widetilde{\lambda}_{r,i}}^{\alpha}\,\Bigr{)}^{\frac{1}{1-\alpha}}$$ (4)

Note that in the above, $\widetilde{\lambda}^{\alpha}_{r,i}$ denotes the $i$th eigenvalue of the $2r\times 2r$ matrix $\widehat{C}_{X,r}$. We remark that the computation of both $\mathrm{FKEA}\text{-}\mathrm{RKE}$ and $\mathrm{FKEA}\text{-}\mathrm{VENDI}_{\alpha}$ can be done by a stochastic algorithm which computes the proxy covariance matrix (2) by summing the sample-based $2r\times 2r$ matrix terms, and then computing the resulting matrix’s Frobenius norm for $\mathrm{RKE}$ score or all the $2r$ matrix’s eigenvalues for a general $\mathrm{VENDI}_{\alpha}$ with $\alpha\neq 2$.

Therefore, to show the FKEA method’s scalability, we need to bound the required RFF size $2r$ for an accurate approximation of the original $n\times n$ kernel matrix. The following theorem proves that the needed feature size will be $\mathcal{O}\bigl{(}\frac{\log n}{\epsilon^{2}}\bigr{)}$ for an $\epsilon$-accurate approximations of the matrix’s eigenspace.

Theorem 2.

Consider a shift-invariant kernel $k(\mathbf{x},\mathbf{x}^{\prime})=\kappa(\mathbf{x}-\mathbf{x}^{\prime})$ where $\kappa(\mathbf{0})=1$. Suppose $\bm{\omega}_{1},\ldots,\bm{\omega}_{r}\sim\widehat{\kappa}$ are independently drawn from PDF $\widehat{\kappa}$. Let $\lambda_{1}\geq\ldots\geq\lambda_{n}$ be the sorted eigenvalues of the normalized kernel matrix $\frac{1}{n}K=\frac{1}{n}\bigl{[}k(\mathbf{x}_{i},\mathbf{x}_{j})\bigr{]}_{n\times n}$. Also, consider the eigenvalues of $\widetilde{\lambda}_{1}\geq\ldots\geq\widetilde{\lambda}_{2r}$ of random matrix $\widetilde{C}_{X,r}$ with their corresponding eigenvectors $\widetilde{\mathbf{v}}_{1},\ldots,\widetilde{\mathbf{v}}_{2r}$. Let $\,\widetilde{\lambda}_{j}=0$ for every $j>2r$. Then, for every $\delta>0$, the following holds with probability at least $1-\delta$:

$$\sqrt{\sum_{i=1}^{n}\bigl{(}\widetilde{\lambda}_{i}-\lambda_{i}\bigr{)}^{2}}\,\leq\,\sqrt{\frac{8\log(n/2\delta)}{r}}\qquad\text{\rm and}\qquad\sqrt{\sum_{i=1}^{n}\Bigl{\|}\frac{1}{n}K\widehat{\mathbf{v}}_{i}-\lambda_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}^{2}_{2}}\,\leq\,\sqrt{\frac{32\log(n/2\delta)}{r}},$$

where $\widehat{\mathbf{v}}_{i}:=\sum_{j=1}^{r}\sin\bigl{(}\widetilde{\mathbf{v}}_{2j}^{\top}\mathbf{x}_{i}\bigr{)}\widetilde{\mathbf{v}}_{2j}+\cos\bigl{(}\widetilde{\mathbf{v}}_{2j-1}^{\top}\mathbf{x}_{i}\bigr{)}\widetilde{\mathbf{v}}_{2j-1}$ is the $i$th proxy eigenvector for $\frac{1}{n}K$.

Corollary 1.

In the setting of Theorem 2, the following approximation guarantees hold for $\mathrm{RKE}$ and $\mathrm{VENDI}_{\alpha}$ scores

• For every $\mathrm{VENDI}_{\alpha}$ with $\alpha\geq 2$, including $\mathrm{RKE}$ for $\alpha=2$, the following dimension-independent bound holds with probability at least $1-\delta$:

  $$\Bigl{|}\mathrm{FKEA}\text{-}\mathrm{VENDI}_{\alpha}\bigl{(}\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\bigr{)}^{\frac{1-\alpha}{\alpha}}\,-\,\mathrm{VENDI}_{\alpha}\bigl{(}\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\bigr{)}^{\frac{1-\alpha}{\alpha}}\Bigr{|}\,\leq\,\sqrt{\frac{8\log(n/2\delta)}{r}}$$

• For every $\mathrm{VENDI}_{\alpha}$ with $1\leq\alpha<2$, assuming a finite dimension for the kernel feature map $\mathrm{dim}(\phi)=m$, the following bound holds with probability at least $1-\delta$:

  $$\Bigl{|}\mathrm{FKEA}\text{-}\mathrm{VENDI}_{\alpha}\bigl{(}\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\bigr{)}^{\frac{1-\alpha}{\alpha}}\,-\,\mathrm{VENDI}_{\alpha}\bigl{(}\mathbf{x}_{1},\ldots,\mathbf{x}_{n}\bigr{)}^{\frac{1-\alpha}{\alpha}}\Bigr{|}\,\leq\,m^{\frac{1}{\alpha}-\frac{1}{2}}\sqrt{\frac{8\log(n/2\delta)}{r}}$$

Remark 2.

According to the theoretical results in [45], the top-$t$ eigenvectors of kernel covariance matrix $C_{X}$ will correspond to the mean of the modes of a mixture distribution with $t$ well-separable modes. Theorem 2 shows for every $1\leq i\leq 2r$, FKEA provides the proxy score function $\widetilde{u}_{i}:\mathbb{R}^{d}\rightarrow\mathbb{R}$ that assigns a likelihood score for an input $\mathbf{x}$ to belong to the $i$th identified mode:

$$\widetilde{u}_{i}(\mathbf{x})\,=\,\sum_{j=1}^{r}\sin\bigl{(}\widetilde{\mathbf{v}}_{2j}^{\top}\mathbf{x}\bigr{)}\widetilde{\mathbf{v}}_{2j,i}+\cos\bigl{(}\widetilde{\mathbf{v}}_{2j-1}^{\top}\mathbf{x}\bigr{)}\widetilde{\mathbf{v}}_{2j-1,i}$$ (5)

Therefore, one can compute the above FKEA-based score for each of the $2r$ eigenvectors over a sample set, and use the samples with the highest scores according to every $\widetilde{u}_{i}$ to characterize the $i$ sample cluster captured by the FKEA method.

Figure 2: RFF-based identified clusters used in FKEA Evaluation in single-colored MNIST [46] dataset with pixel embedding, Fourier feature dimension $2r=4000$ and bandwidth $\sigma=7$. The graphs indicate increase in FKEA RKE/VENDI diversity metrics with increasing number of labels.

6 Numerical Results

We evaluated the FKEA method on several image, text, and video datasets to assess its performance in quantifying diversity in different domains. In the experiments, we computed the empirical covariance matrix of $2r$-dimensional Fourier features with a Gaussian kernel with bandwidth parameter $\sigma$ tuned for each dataset, and then applied FKEA approximation for the $\mathrm{VENDI}_{1}$, $\mathrm{VENDI}_{1.5}$, and the $\mathrm{RKE}$ (same as $\mathrm{VENDI}_{2}$) scores. Experiments were conducted on 8 RTX3090 GPUs. We interpreted the modes identified by FKEA entropy-based diversity evaluation using the eigenvectors of the proxy covariance matrix as discussed in Remark 2. For each eigenvector, we presented the training data with maximum eigenfunction values corresponding to the eigenvector.

6.1 Experimental Results on Image Data

To investigate the FKEA method’s diversity evaluation in settings where we know the ground-truth clusters and their quantity, we simulated an experiment on the colored MNIST [46] data with the images of 10 colored digits as shown in Figure 2. We evaluated the FKEA approximations of the diversity scores while including samples from $t$ digits for $t\in\{1,\ldots,10\}$. The plots in Figure 2 indicate the increasing trend of the scores and FKEA’s tight approximations of the scores. Also, we show the top-25 training samples with the highest scores according to the top-10 FKEA eigenvectors, showing the method captured the ground-truth clusters.

We conducted an experiment on the ImageNet data to monitor the scores’ behavior evaluated for 50k samples from an increasing number of ImageNet labels. Figure 3 shows the increasing trend of the scores as well as the top-9 samples representing the top-4 identified clusters used for the entropy calculation. Figures 3, 5 and 4 illustrate the top FKEA modes with corresponding parameters. The modes represent a specific characteristic common between the top images. Table 1 evaluates the diversity scores of various ImageNet GAN models using the FKEA method applied to VENDI-1 and RKE, with potential extension to the entire VENDI family. The comparison includes baseline diversity metrics such as Inception Score [12], FID [7], Improved Precision/Recall [10], and Density/Coverage [11]. Also, Figure 7 presents the FKEA approximated entropy scores with different truncation factors in StyleGAN3 [47] on 30k generated data for each truncation factor, showing the increasing diversity scores with the truncation factor.

We defer discussing the results on generated synthetic images to the Appendix.

Figure 3: RFF-based identified clusters used in FKEA Evaluation in ImageNet[6] dataset with DinoV2 embedding, Fourier feature dimension $2r=16k$ and Gaussian Kernel bandwidth $\sigma=25$. The graphs indicate increase in FKEA diversity metrics with increasing number of labels per 50k samples.

Figure 4: RFF-based identified clusters used in FKEA Evaluation in FFHQ[48] dataset with DinoV2 embedding, Fourier feature dimension $2r=16k$ and Gaussian Kernel bandwidth $\sigma=20$.

Figure 5: RFF-based identified clusters used in FKEA Evaluation in AFHQ[49] dataset with DinoV2 embedding, Fourier feature dimension $2r=16k$ and Gaussian Kernel bandwidth $\sigma=20$.

Figure 6: RFF-based identified clusters used in FKEA Evaluation in MS COCO [50] dataset with DinoV2 embedding, Fourier feature dimension $2r=16k$ and Gaussian Kernel bandwidth $\sigma=22$.

Figure 7: FKEA metrics behavior under different truncation factor $\psi$ of StyleGAN3 [47] generated FFHQ samples.

Method	IS $\uparrow$	FID $\downarrow$	Precision $\uparrow$	Recall $\uparrow$	Density $\uparrow$	Coverage $\uparrow$	FKEA VENDI-1 $\uparrow$	FKEA RKE $\uparrow$
Dataset (100k)	-	-	-	-	-	-	9176.9	996.7
ADM [51]	542.6	11.12	0.78	0.79	0.88	0.89	8360.3	633.4
ADMG [51]	659.3	5.63	0.87	0.84	0.80	0.85	8524.2	811.5
ADMG-ADMU [51]	701.6	4.78	0.90	0.73	1.20	0.96	8577.6	839.8
BigGAN [52]	696.4	7.91	0.81	0.44	0.99	0.57	7120.5	492.4
DiT-XL-2 [53]	743.2	3.56	0.92	0.84	1.16	0.97	8626.5	855.8
GigaGAN [54]	678.8	4.29	0.89	0.74	0.74	0.70	8432.5	671.6
LDM [55]	734.4	4.75	0.93	0.76	1.04	0.93	8573.7	811.9
Mask-GIT [56]	717.4	5.66	0.91	0.72	1.01	0.82	8557.4	759.5
RQ-Transformer [57]	558.3	9.57	0.80	0.76	0.77	0.59	8078.4	512.1
StyleGAN-XL[58]	675.4	4.34	0.89	0.74	1.18	0.96	8171.9	703.5

Table 1: Evaluated scores for ImageNet generative models. The Gaussian kernel bandwidth parameter chosen for RKE, VENDI, FKEA-VENDI and FKEA-RKE is $\sigma=25$ and Fourier features dimension $2r=16k$. The scores were obtained by running the GitHub of [20] on pre-generated 50k samples.

6.2 Experimental Results on Text and Video Data

To perform experiments on the text data with known clustering ground-truth, we generated 500,000 paragraphs using GPT-3.5 [59] about 100 randomly selected countries (5k samples per country). In the experiments, the text embedding used was text-embedding-3-large [60, 61, 59]. We evaluated the diversity scores over data subsets of size 50k with different numbers of mentioned countries. Figure 8 shows the growing trend of the diversity scores when including more countries. The figure also shows the countries mentioned in the top-6 modes provided by FKEA-based principal eigenvectors, which shows the RFF-based clustering of countries correlates with their continent and geographical location. We discuss the numerical results on non-synthetic text datasets, CNN/Dailymail [62][63], CMU Movie Corpus [64], in the Appendix.

In Table 2 we extend FKEA to non-synthetic Wikipedia dataset. To visualise the results, we used YAKE [65] algorithm to extract and present the identified unigram and bigram keywords. Identified mode 1 correlates most with historical figures/events/places. Mode 2 clusters smaller villages and rural regions together. Mode 3 is exclusively about people in sports, such as athletes and referees. Mode 4 visualises various music bands and albums. Lastly, mode 5 presents the articles about sports events, such as football leagues.

Mode #1	Mode #2	Mode #3	Mode #4	Mode #5
Grosse Pointe	Ishkli	Gerson Garca	Girugamesh (album)	2009 WPSL season
Mark Scharf	Khazora	Valentin Mogilny	Japonesque (album)	2012 Milwaukee…
Alexander McKee	Sis, Azerbaijan	Gerald Lehner (referee)	Documentaly	2020 San Antonio FC…
Clay Huffman	Zasun	Dmitri Nezhelev	EX-Girl	2020 HFX Wanderers FC…
Ravenna, Ohio	Zaravat	Grigori Ivanov	Indie 2000	2020 Sporting Kansas City…
C. M. Eddy Jr.	Bogat	Leonidas Morakis	Triangle (Perfume album)	FC Tucson
Hornell, New York	Yakkakhona	Jos Luis Alonso Ber…	Waste Management (album)	2008 K League
Larchmont, New York	Yava, Tajikistan	Giovanni Gasperini	Fush Yu Mang	201112 New Zealand Football…
Robert Hague	Ikizyak	Mohamed Chab	Fantastipo (song)	200809 Melbourne Victory FC…
General Hershy Bar	Khushikat	Louis Darmanin	Xtort	2012 Pittsburgh Power season
Keywords
London	populated places	players category	music video	American football
American History	Maplandia.com Category	Association football	album	players Category
University Press	municipality	FIFA World	studio album	Football League
United States	village	World Cup	Records albums	League
World War	Osh Region	Summer Olympics	Singles Chart	League Soccer

Table 2: Top 5 Wikipedia Dataset Modes with corresponding eigenvalues with text-embedding-3-large embeddings and bandwidth $\sigma=1.0$

For video data experiments, we considered two standard video datasets, UCF101 [66] and Kinetics-400 [67]. Following the video evaluation literature [68, 69], we used the I3D pre-trained model [70] as embedding, which maps a video sample to a 1024-dimensional vector. As shown in Figure 9, increasing the number of video classes of test samples led to an increase in the FEKA approximated diversity metrics. Also, while the samples identified for the first identified cluster look broad, the next modes seemed more specific. We discuss the results of the Kinetics dataset in the Appendix.

Figure 8: FKEA diversity metrics with the increasing number of countries in the synthetic dataset.

Mode #1		Mode #2		Mode #3		Mode #4		Mode #5		Mode #6
Burkina Faso	34%	Argentina	77%	Azerbaijan	100%	Cambodia	94%	Belarus	100%	Bolivia	97%
Benin	23%	Chile	23%			Afghanistan	6%			Ecuador	3%
Chad	22%
Burundi	13%
Cameroon	8%

Table 3: Top 5 synthetic countries dataset modes with text-embedding-3-large embedding, Fourier features dim $2r=8000$ and Gaussian Kernel bandwidth $\sigma=0.9$. The table summarises the mentions of each country in the top 100 paragraphs identified for the eigenvectors corresponding to each mode.

(a) Mode #1

(b) Mode #2

(c) Mode #3

(d) Mode #4

Figure 9: RFF-based identified clusters used in FKEA evaluation in UCF101 dataset with I3D embedding. The graphs indicate an increase in FKEA diversity metrics with more classes.

7 Limitations

Incompatibility with non shift-invariant kernels. Our analysis targets a shift-invariant kernel, which does not apply to a general kernel function, such as polynomial kernels. In practice, many ML algorithms rely on simpler kernels that may not have the shift-invariant property. Due to to specifics of FKEA framework, we cannot directly extend the work to such kernels. We leave the framework’s extension to other kernel functions for future studies.

Reliance on Embeddings. FKEA clustering and diversity assessment metrics rely on the quality of the underlying embedding space. Depending on the training and pre-training datasets, the semantic clustering properties may change. We leave in-depth study of embedding space behavior for future research.

8 Conclusion

In this work, we proposed the Fourier-based FKEA method to efficiently approximate the kernel-based entropy scores $\mathrm{VENDI}_{\alpha}$ and $\mathrm{RKE}$ scores. The application of FKEA results in a scalable reference-free evaluation of generative models, which could be utilized in applications where no reference data is available for evaluation. A future direction to our work is to study the sample complexity of the matrix-based entropy scores and the FKEA’s approximation under high-dimensional kernel feature maps, e.g. the Gaussian kernel. Also, analyzing the role of feature embedding in the method’s application to text and video data would be interesting for future exploration.

Appendix A Proofs

A.1 Proof of Theorem 1

The proof of Theorem 1 combines third ingredients. The first is the relation between the circuit size of a function $C$ and of its partial derivatives $\nabla C=(\del C/\del x_{1},\dots,\del C/\del x_{n})$.

Lemma 1.

The function $\nabla C$ has a circuit over basis $\nabla B\cup\{+,\times\}$ whose size is within a constant factor of the size of $C$.

Lemma 1 is a feature of the backpropagation algorithm [71, 72]. This is a linear-time algorithm for constructing a circuit for $\nabla C$ given the circuit $C$ as input. In contrast, the forward propagation algorithm allows efficient computation of a single (partial) derivative even for circuits with multivalued outputs, giving the second ingredient:

Lemma 2.

Let $C$ be a circuit over basis $B$ and $t$ be an input to $C$. There exists a circuit that computes the derivative $\del g/\del t$ for every gate $g$ of $C$ over basis $\nabla B\cup\{+,\times\}$ whose size is within a constant factor of the size of $C$.

The last ingredient is the following identity. For a scalar function $f$ over the complex numbers and matrix $X$ diagonalizable as $U\Lambda U^{T}$, we define $f(X)$ to be the function $Uf(\Lambda)U^{T}$ where $f$ is applied entry-wise to the diagonal matrix $\Lambda$.

Lemma 3.

For every $f$ that is analytic over an open domain $\Omega$ containing all sufficiently large complex numbers and every matrix $X$ whose spectrum is contained in $\Omega$, $\nabla\mathrm{Tr}(f(X))=f^{\prime}(X)$.

We first illustrate the proof in the special case when $\norm{X}$ is within the radius of convergence of $f$. Namely, $f(x)$ is represented by the absolutely convergent series $\sum f^{(k)}(0)x^{k}/k!$ for all $\abs{x}\leq\rho$. Then $f(X)=\sum f^{(k)}(0)X^{k}/k!$ assuming $\norm{X}\leq\rho$. By linearity (and using the fact that derivatives preserve radius of convergence) it is sufficient to show that

$$\nabla\mathrm{Tr}X^{k}=\frac{dX^{k}}{dX},$$ (6)

which can be verified by explicit calculation: Both sides equal $kX^{k-1}$. This is sufficient to establish Theorem 1 for all integer $\alpha>2$.

Proof of Lemma 3.

The Cauchy integral formula for matrices yields the representation

$$f(X)=\frac{1}{2\pi i}\int_{C}f(z)(zI-X)^{-1}dz,$$

for any closed curve $C$ whose interior contains the spectrum of $X$. As $(zI-X)^{-1}$ is continuous along $C$, we can write

$$\nabla\mathrm{Tr}f(X)=\frac{1}{2\pi i}\int_{C}f(z)\nabla\mathrm{Tr}(zI-X)^{-1}dz.$$ (7)

Choosing $C$ to be a circle of radius $\rho$ larger than the spectral norm of $X$, for all $z$ of magnitude $\rho$ we have the identity

$$(zI-X)^{-1}=z^{-1}(I-z^{-1}X)^{-1}=z^{-1}\sum_{k=0}^{\infty}z^{-k}X^{k}$$

As the series $\sum z^{-k}\nabla\mathrm{Tr}X^{k}=\sum kz^{-k}X^{k-1}$ converges absolutely in spectral norm, using (6) we obtain the identity $\nabla\mathrm{Tr}(zI-X)^{-1}=d(zI-X)^{-1}/dX$, namely the lemma holds for the function $f(X)=(zI-X)^{-1}$. Plugging into (7) and exchanging the order of integration and derivation proves the lemma. ∎

Proof of Theorem 1.

Assume $\mathrm{Tr}\rho^{\alpha}$ (resp., $-\mathrm{Tr}\rho\log\rho$) has circuit size $s(d)$. By Lemma 1 and Lemma 3, $\nabla\mathrm{Tr}\rho^{\alpha}=\alpha\rho^{\alpha-1}$ (resp., $-\nabla\mathrm{Tr}\rho\log\rho=\log\rho-1/\ln 2$) has circuit size $O(s(d))$. For every symmetric matrix $X$ and sufficiently small $t$, the matrix $\rho=I+tX$ is positive semi-definite. By Lemma 2 the $R^{d^{2}}$-valued function $\partial^{2}\rho/\partial t^{2}$ has circuit size $O(s^{d})$. The value of this function at $t=0$ is $\alpha(\alpha-1)(\alpha-2)X^{2}$ (resp., $X^{2}$), namely the square of the input matrix $X$ up to constant. Finally, computing the product $AB$ reduces to squaring the symmetric matrix

$$\begin{pmatrix}&A^{T}&B\\ A&&\\ B^{T}&&\end{pmatrix}.\hfill\qed$$

A.2 Proof of Theorem 2

Assuming that the shift-invariant kernel $k(\mathbf{x},\mathbf{x}^{\prime})=\kappa(\mathbf{x}-\mathbf{x}^{\prime})$ is normalized (i.e. $\kappa(\mathbf{0})=1$), then the Fourier transform $\widehat{\kappa}$ is a valid PDF according to Bochner’s theorem and also an even function because $\kappa$ takes real values. Then, we have

$$k\bigl{(}\mathbf{x},\mathbf{x}^{\prime}\bigr{)}\, =\,\kappa_{\sigma}\bigl{(}\mathbf{x}-\mathbf{x}^{\prime}\bigr{)}$$

$$\stackrel{{\scriptstyle(a)}}{{=}}\,\int\widehat{\kappa_{\sigma}}(\bm{\omega})\exp\bigl{(}i\bm{\omega}^{\top}(\mathbf{x}-\mathbf{x}^{\prime})\bigr{)}\mathrm{d}\bm{\omega}$$

$$\stackrel{{\scriptstyle(b)}}{{=}}\,\int\widehat{\kappa_{\sigma}}(\bm{\omega})\cos\bigl{(}\bm{\omega}^{\top}(\mathbf{x}-\mathbf{x}^{\prime})\bigr{)}\mathrm{d}\bm{\omega}$$

$$\stackrel{{\scriptstyle}}{{=}}\,\mathbb{E}_{\bm{\omega}\sim\widehat{\kappa}}\Bigl{[}\cos\bigl{(}\bm{\omega}^{\top}(\mathbf{x}-\mathbf{x}^{\prime})\bigr{)}\Bigr{]}$$

$$=\,\mathbb{E}_{\bm{\omega}\sim\widehat{\kappa}}\Bigl{[}\cos\bigl{(}\bm{\omega}^{\top}\mathbf{x})\cos\bigl{(}\bm{\omega}^{\top}\mathbf{x}^{\prime})+\sin\bigl{(}\bm{\omega}^{\top}\mathbf{x})\sin\bigl{(}\bm{\omega}^{\top}\mathbf{x}^{\prime})\Bigr{]}$$

Here, (a) comes from the synthesis property of the Fourier transform. (b) holds since $\widehat{\kappa_{\sigma}}$ is an even function, resulting in a zero imaginary term in the Fourier synthesis.

Therefore, since $\bigl{|}\cos\bigl{(}\bm{\omega}^{\top}\mathbf{y}\bigr{)}\bigr{|}\leq 1$ for all $\bm{\omega}$ and $\mathbf{y}$, one can apply Hoeffding’s inequality to show for independently drawn $\bm{\omega}_{1},\ldots,\bm{\omega}_{r}\stackrel{{\scriptstyle\mathrm{iid}}}{{\sim}}\widehat{\kappa}$ the following probably correct bound holds:

$$\mathbb{P}\biggl{(}\Bigl{|}\frac{1}{r}\sum_{i=1}^{r}\cos\bigl{(}\bm{\omega}_{i}^{\top}(\mathbf{x}-\mathbf{x}^{\prime})\bigr{)}-\mathbb{E}_{\bm{\omega}\sim\widehat{\kappa}}\Bigl{[}\cos\bigl{(}\bm{\omega}^{\top}(\mathbf{x}-\mathbf{x}^{\prime})\bigr{)}\Bigr{]}\Bigr{|}\geq\epsilon\biggr{)}\leq 2\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}$$

Therefore, as the identity $\cos(a-b)=\cos(a)\cos(b)+\sin(a)\sin(b)$ reveals $\frac{1}{r}\sum_{i=1}^{r}\cos\bigl{(}\bm{\omega}_{i}^{\top}(\mathbf{x}-\mathbf{x}^{\prime})\bigr{)}\,=\,\widetilde{\phi}_{r}(\mathbf{x})^{\top}\widetilde{\phi}_{r}(\mathbf{x}^{\prime})$, the above bound can be rewritten as

$$\mathbb{P}\biggl{(}\Bigl{|}\widetilde{\phi}_{r}(\mathbf{x})^{\top}\widetilde{\phi}_{r}(\mathbf{x}^{\prime})-k(\mathbf{x},\mathbf{x}^{\prime})\Bigr{|}\geq\epsilon\biggr{)}\leq 2\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}.$$

Also, $\widetilde{k}_{r}(\mathbf{x},\mathbf{x}^{\prime})=\widetilde{\phi}_{r}(\mathbf{x})^{\top}\widetilde{\phi}_{r}(\mathbf{x}^{\prime})$ is by definition a normalized kernel, implying that

$$\forall\mathbf{x}\in\mathbb{R}^{d}:\quad\widetilde{\phi}_{r}(\mathbf{x})^{\top}\widetilde{\phi}_{r}(\mathbf{x})-k(\mathbf{x},\mathbf{x})\,=\,0.$$

As a result, one can apply the union bound to combine the above inequalities and show for every sample set $\mathbf{x}_{1},\ldots,\mathbf{x}_{n}$:

$$\mathbb{P}\biggl{(}\max_{1\leq i,j\leq n}\Bigl{(}\widetilde{\phi}_{r}(\mathbf{x}_{i})^{\top}\widetilde{\phi}_{r}(\mathbf{x}_{j})-k_{\text{\rm Gaussian}(\sigma^{2})}(\mathbf{x}_{i},\mathbf{x}_{j})\Bigr{)}^{2}\geq\epsilon^{2}\biggr{)}\leq 2{n\choose 2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}.$$

Considering the normalized kernel matrix $\frac{1}{n}K=\frac{1}{n}\bigl{[}k(\mathbf{x}_{i},\mathbf{x}_{j})\bigr{]}_{1\leq i,j\leq n}$ and the proxy normalized kernel matrix $\frac{1}{n}\widetilde{K}=\frac{1}{n}\bigl{[}\widetilde{\phi}_{r}(\mathbf{x}_{i})^{\top}\widetilde{\phi}_{r}(\mathbf{x}_{j})\bigr{]}_{1\leq i,j\leq n}$, the above inequality implies that

$$\mathbb{P}\Bigl{(}\bigl{\|}\frac{1}{n}\widetilde{K}-\frac{1}{n}K\bigr{\|}^{2}_{F}\,\geq\,n^{2}\frac{\epsilon^{2}}{n^{2}}\Bigr{)}\;\leq\;{n\choose 2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}.$$

$$\Longrightarrow\quad \mathbb{P}\Bigl{(}\bigl{\|}\frac{1}{n}\widetilde{K}-\frac{1}{n}K\bigr{\|}_{F}\,\geq\,\epsilon\Bigr{)}\,<\,\frac{n^{2}}{2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}.$$ (8)

Leveraging the eigenvalue-perturbation bound in [73], we can translate the above bound to the following for the sorted eigenvalues $\lambda_{1}\geq\cdots\geq\lambda_{n}$ of $\frac{1}{n}K$ and the sorted eigenvalues ${\widetilde{\lambda}}_{1}\geq\cdots\geq{\widetilde{\lambda}}_{n}$ of $\frac{1}{n}\widetilde{K}$

$$\sqrt{\sum_{i=1}^{n}(\widetilde{\lambda}_{i}-\lambda_{i})^{2}}\,\leq\,\Bigl{\|}\frac{1}{n}{\widetilde{K}}-\frac{1}{n}{K}\Bigr{\|}_{F}$$

which shows

$$\mathbb{P}\Bigl{(}\sqrt{\sum_{i=1}^{r^{\prime}}(\widetilde{\lambda}_{i}-\lambda_{i})^{2}}\,\geq\,\epsilon\Bigr{)}\,\leq\,\frac{n^{2}}{2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}$$ (9)

Defining $\delta=\frac{n^{2}}{2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}$, i.e., $\epsilon=\sqrt{\frac{8\log\bigl{(}n/2\delta\bigr{)}}{r}}$, leads to

$$\mathbb{P}\Bigl{(}\sqrt{\sum_{i=1}^{n}(\widetilde{\lambda}_{i}-\lambda_{i})^{2}}\,\leq\,\epsilon\Bigr{)}\,\geq\,1-\delta.$$ (10)

Noting that the normalized proxy kernel matrix $\widetilde{K}$ and the proxy kernel covariance matrix $\widetilde{C}_{X}$ share identical non-zero eigenvalues together with the above bound finish the proof of Theorem 2’s first part.

Concerning Theorem 2’s approximation guarantee for the eigenvectors, note that for each eigenvectors $\widehat{\mathbf{v}}_{i}$ of the proxy kernel matrix $\frac{1}{n}\widetilde{K}$, the following holds:

$$\Bigl{\|}\frac{1}{n}{K}\widehat{\mathbf{v}}_{i}-{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}_{2}\, \leq\,\Bigl{\|}\frac{1}{n}{K}\widehat{\mathbf{v}}_{i}-\widetilde{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}_{2}+\Bigl{\|}\widetilde{\lambda}_{i}\widehat{\mathbf{v}}_{i}-{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}_{2}$$

$$=\,\Bigl{\|}\bigl{(}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\bigr{)}\widehat{\mathbf{v}}_{i}\Bigr{\|}_{2}+\bigl{|}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{|}$$

Therefore, applying Young’s inequality shows that

$$\Bigl{\|}\frac{1}{n}{K}\widehat{\mathbf{v}}_{i}-{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}^{2}_{2}\, \leq\,2\Bigl{\|}\bigl{(}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\bigr{)}\widehat{\mathbf{v}}_{i}\Bigr{\|}^{2}_{2}+2\bigl{(}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{)}^{2}$$

$$=\,2\mathrm{Tr}\Bigl{(}\widehat{\mathbf{v}}^{\top}_{i}\bigl{(}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\bigr{)}^{2}\widehat{\mathbf{v}}_{i}\Bigr{)}+2\bigl{(}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{)}^{2}$$

$$=\,2\mathrm{Tr}\Bigl{(}\widehat{\mathbf{v}}_{i}\widehat{\mathbf{v}}^{\top}_{i}\bigl{(}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\bigr{)}^{2}\Bigr{)}+2\bigl{(}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{)}^{2},$$

which implies that

$$\sum_{i=1}^{n}\Bigl{\|}\frac{1}{n}{K}\widehat{\mathbf{v}}_{i}-{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}^{2}_{2}\, \leq 2\mathrm{Tr}\Bigl{(}\bigl{(}\sum_{i=1}^{n}\widehat{\mathbf{v}}_{i}\widehat{\mathbf{v}}^{\top}_{i}\bigr{)}\bigl{(}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\bigr{)}^{2}\Bigr{)}+2\sum_{i=1}^{n}\bigl{(}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{)}^{2}$$

$$=2\mathrm{Tr}\Bigl{(}\bigl{(}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\bigr{)}^{2}\Bigr{)}+2\sum_{i=1}^{n}\bigl{(}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{)}^{2}$$

$$=2\Bigl{\|}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\Bigr{\|}^{2}_{F}+2\sum_{i=1}^{n}\bigl{(}\tilde{\lambda}_{i}-{\lambda}_{i}\bigr{)}^{2}$$

$$\leq 4\Bigl{\|}\frac{1}{n}{K}-\frac{1}{n}\widetilde{K}\Bigr{\|}^{2}_{F}.$$

The above proves that

$$\mathbb{P}\Bigl{(}\sqrt{\sum_{i=1}^{n}\Bigl{\|}\frac{1}{n}{K}\widehat{\mathbf{v}}_{i}-{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}^{2}_{2}}\geq\epsilon\Bigr{)}\, \leq\,\mathbb{P}\Bigl{(}\bigl{\|}\frac{1}{n}\widetilde{K}-\frac{1}{n}K\bigr{\|}_{F}\,\geq\,\frac{\epsilon}{2}\Bigr{)}$$

$$<\,\frac{n^{2}}{2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{8}\Bigr{)}.$$

Therefore, considering the provided definition $\delta=\frac{n^{2}}{2}\exp\Bigl{(}-\frac{r\epsilon^{2}}{2}\Bigr{)}$, i.e., $2\epsilon=\sqrt{\frac{32\log\bigl{(}n/2\delta\bigr{)}}{r}}$, we will have the following which completes the proof:

$$\mathbb{P}\Bigl{(}\sqrt{\sum_{i=1}^{n}\Bigl{\|}\frac{1}{n}{K}\widehat{\mathbf{v}}_{i}-{\lambda}_{i}\widehat{\mathbf{v}}_{i}\Bigr{\|}^{2}_{2}}\leq 2{\epsilon}\Bigr{)}\,\geq\,1-\delta.$$

Appendix B Additional Numerical Results

B.1 Synthetic Image Dataset Modes

In addition to running clustering on real image datasets, we also applied FKEA with varying Gaussian Kernel bandwidth parameter $\sigma$ to generated synthetic datasets. The results are presented for LDM, VDVAE, StyleGAN-XL and InsGen models trained on FFHQ samples.

Figure 10: RFF-based identified clusters used in FKEA Evaluation of LDM [55] generative model in FFHQ with DINOv2 embeddings and bandwidth $\sigma=20$

Figure 11: RFF-based identified clusters used in FKEA Evaluation of VDVAE [74] generative model in FFHQ with DINOv2 embeddings and bandwidth $\sigma=20$

Figure 12: RFF-based identified clusters used in FKEA Evaluation of InsGen [75] generative model in FFHQ with DINOv2 embeddings and bandwidth $\sigma=20$

Figure 13: RFF-based identified clusters used in FKEA Evaluation of StyleGAN-XL[58] generative model in FFHQ with DINOv2 embeddings and bandwidth $\sigma=20$

B.2 Effect of other embeddings on FKEA clustering

Even though DinoV2 is a primary embedding in our experimental settings, we acknowldge the use of other embedding models such as SwAV[24] and CLIP[23]. The resulting clusters differ from original DinoV2 clusters and require separate bandwidth parameter finetuning. In our experiments, SwAV embedding emphasizes object placement, such as animal in grass or white backgrounds, as seen in figure 14. CLIP on the other hand clusters by objects, such as birds/dogs/bugs, as seen in figure 15. These results indicate that FKEA powered by other embeddings will slightly change the clustering features; however, it does not hinder the clustering performance of RFF based clustering with FKEA method.

Figure 14: RFF-based identified clusters used in FKEA Evaluation of SwAV embedding on ImageNet with bandwidth $\sigma=0.8$

Figure 15: RFF-based identified clusters used in FKEA Evaluation of CLIP embedding on ImageNet with bandwidth $\sigma=7.0$

B.3 Text Dataset Modes

To understand the applicability and effectiveness of the FKEA method beyond images, we extended our study to text datasets. We observed that clustering text data poses a more challenging task compared to image data. This increased difficulty arises from the ambiguity in defining clear separability factors within text, a contrast to the more visually distinguishable criteria in images. The process of evaluating text clusters is not straightforward and often varies significantly based on human judgment and perception.

To visualise the results, we use YAKE [65] algorithm to extract the keywords in each text mode and present the identified unigram and bigram keywords. We demonstrate that the results hold for text datasets and identified clusters are meaningful.

Table 4 outlines the largest modes identified within a news dataset analyzed using the FKEA method, with a detailed focus on the content themes of each mode. The most dominant mode is associated with topics related to crime and police activities, indicating a frequent coverage area in the dataset. Mode 2 is closely correlated with President Obama, reflecting a significant focus on political coverage. Mode 3 pertains to dieting, which suggests a presence of health and lifestyle topics. Mode 4 is linked to environmental disasters, highlighting the dataset’s attention to ecological and crisis-related news. Finally, Mode 5 deals with plane crash accidents, underscoring the coverage of major transportation incidents.

Mode #1	Mode #2	Mode #3	Mode #4	Mode #5
police	President Obama	size	people	died
British police	Obama	year	severe weather	family
police officer	Barack Obama	weight	Death toll	plane crash
Police found	President	dress size	heavy rain	mother
family	White House	stone	Environment Agency	plane
found	Obama administration	Slimming World	million people	found
told police	Obama calls	lost	rain	people
court	House	lose weight	flood warnings	children
arrested	United States	diet	people dead	hospital
home	Obama plan	model	people killed	found dead

Table 4: Top 5 CNN/Dailymail 3.0.0 [62][63] Dataset Modes with corresponding eigenvalues with text-embedding-3-large embeddings and bandwidth $\sigma=0.8$.

Table 5 delineates the distribution of genres and production types within a dataset of movie summaries analyzed using the FKEA method. The first mode predominantly covers drama TV shows without focusing on any specific subtopic, indicating a broad categorization within this genre. From mode 2 onwards, the features become more distinct and defined. Mode 2 specifically represents Bollywood movies, with a significant emphasis on the Romance genre. Mode 3 is dedicated to clustering comedy shows. Mode 4 is exclusively associated with cartoons, evidenced by keywords such as "Tom & Jerry". Lastly, mode 5 clusters together detective and crime fiction shows.

Mode #1	Mode #2	Mode #3	Mode #4	Mode #5
The House on Tele..	Anand	Bring Your Smile Along	Chhota Bheem…	Walk a Crooked Mile
Seems Like Old Times	I Love You	The Girl Most Likely	Duck Amuck	Assignment to Kill
Shadows and Fog	Toh Baat Pakki	Hips, Hips, Hooray!	Hare-Abian Nights	The Crime of the Century
Obsession	Abodh	Lady Be Good	Porky’s Five and Ten	Murder at Glen Athol
Milk Money	Khulay Aasman…	The Courtship of Eddie’s…	Sock-a-Doodle-Do	Guns
Very Bad Things	Kasthuri Maan	You Live and Learn	Buccaneer Bunny	Because of the Cats
Blame It on the Bell…	Chhaya	Dames	Hare Lift	The House of Hate
The Miracle Man	Yeh Dillagi	Painting the Clouds…	Scrap Happy Daffy	The Ace of Scotland Yard
The Sleeping Tiger	Deva	Pin Up Girl	Hic-cup Pup	The World Gone Mad
The Scapegoat	Bhalobasa Bha…	Too Young to Kiss	The Goofy Gophers	Firepower
Keywords
mystery	hindi film	musical	animation	crime
noir	romance	theme songs	Tom & Jerry	murder
kidnap	love	city	Spike	detective
crime	marriage	romance	adventure	investigation
police	daughter		comedy	killer

Table 5: Top 5 CMU Movie Summary Corpus [64] Dataset Modes with corresponding eigenvalues with text-embedding-3-large embeddings and bandwidth $\sigma=0.8$.

B.4 Video Dataset Modes

In this section, we present additional experiments on the Kinetics-400[67] video dataset. This dataset comprises 400 human action categories, each with a minimum of 400 video clips depicting the action. Similar to the video evaluation metrics, we used the I3D pre-trained model[70] which maps each video to a 1024-vector feature. Figure 16, the first mode captured broader concepts while the other models focused on specific ones. Also, the plots indicate that increasing the number of classes from 40 to 400 results in an increase in the FKEA metrics.

(a) Mode #1

(b) Mode #2

(c) Mode #3

(d) Mode #4

(e) Mode #5

(f) Mode #6

Figure 16: RFF-based identified clusters used in FKEA Evaluation in Kinetics-400 dataset with I3D embeddings. Plots indicate that increasing the number of classes from 40 to 400 results in an increase in the FKEA metrics.