A Characteristic Function Approach to Deep Implicit Generative Modeling

Abdul Fatir Ansari, Jonathan Scarlett, Harold Soh

Abstract

Implicit Generative Models (IGMs) such as GANs have emerged as effective data-driven models for generating samples, particularly images. In this paper, we formulate the problem of learning an IGM as minimizing the expected distance between characteristic functions. Specifically, we minimize the distance between characteristic functions of the real and generated data distributions under a suitably-chosen weighting distribution. This distance metric, which we term as the characteristic function distance (CFD), can be (approximately) computed with linear time-complexity in the number of samples, in contrast with the quadratic-time Maximum Mean Discrepancy (MMD). By replacing the discrepancy measure in the critic of a GAN with the CFD, we obtain a model that is simple to implement and stable to train. The proposed metric enjoys desirable theoretical properties including continuity and differentiability with respect to generator parameters, and continuity in the weak topology. We further propose a variation of the CFD in which the weighting distribution parameters are also optimized during training; this obviates the need for manual tuning, and leads to an improvement in test power relative to CFD. We demonstrate experimentally that our proposed method outperforms WGAN and MMD-GAN variants on a variety of unsupervised image generation benchmarks.

Introduction

Training Generative Adversarial Networks (GANs) is notoriously difficult due to issues like vanishing gradients and mode collapse. While Integral Probability Metrics (IPMs) such as the Wasserstein distance (WGAN) and Maximum Mean Discrepancy (MMD) have improved stability, they come with trade-offs. MMD, for instance, has a computational cost that grows quadratically ($O(n^2)$) with the number of samples, making it expensive for large-scale training. This paper introduces the Characteristic Function Distance (CFD), a linear-time alternative that maintains theoretical robustness while offering superior performance.

Methodology

The core innovation lies in comparing distributions using their Characteristic Functions (CFs)—the Fourier transform of their probability density functions. Since a CF uniquely identifies a distribution, the weighted $L_2$ distance between the CFs of real and generated data serves as an effective loss function.

Linear Complexity and Optimized Weighting

Unlike MMD, the Empirical CFD (ECFD) requires comparing samples against a set of $k$ random "frequencies," resulting in $O(nk)$ complexity. A key advancement is the Optimized CFD (OCF-GAN), where the parameters of the weighting distribution are optimized during training. This allows the discriminator to focus on the frequencies where the distributions differ most, significantly increasing statistical power.

Test power comparison between optimized (OECFD) and standard (ECFD) metrics as dimensionality increases.

Experimental Results

The authors evaluated OCF-GAN on synthetic data and standard benchmarks including MNIST, CIFAR-10, STL-10, and CelebA.

Performance and Convergence

On the CelebA dataset, OCF-GAN-GP achieved a Fréchet Inception Distance (FID) of 9.39, outperforming WGAN-GP (10.03) and MMD-GAN (18.48). Beyond final quality, the optimized weighting scheme leads to significantly faster convergence compared to fixed-weighting baselines.

Mean Absolute Error (MAE) convergence across iterations, demonstrating the rapid training of the OCF-GAN-GP variant.

Discussion and Conclusion

The CFD approach offers a "frequency domain advantage," capturing structural differences that spatial-domain metrics might miss. Theoretical proofs confirm that the CFD is continuous and differentiable, providing smooth gradients even when distributions are initially far apart. By shifting the focus to characteristic functions, this work provides a mathematically elegant and computationally efficient framework for training implicit generative models. The OCF-GAN variant demonstrates that "learning where to look" in the frequency domain is a powerful strategy for improving both the speed and quality of generative modeling.

Summary updated

Abstract

Implicit Generative Models (IGMs) such as GANs have emerged as effective data-driven models for generating samples, particularly images. In this paper, we formulate the problem of learning an IGM as minimizing the expected distance between characteristic functions. Specifically, we minimize the distance between characteristic functions of the real and generated data distributions under a suitably-chosen weighting distribution. This distance metric, which we term as the characteristic function distance (CFD), can be (approximately) computed with linear time-complexity in the number of samples, in contrast with the quadratic-time Maximum Mean Discrepancy (MMD). By replacing the discrepancy measure in the critic of a GAN with the CFD, we obtain a model that is simple to implement and stable to train. The proposed metric enjoys desirable theoretical properties including continuity and differentiability with respect to generator parameters, and continuity in the weak topology. We further propose a variation of the CFD in which the weighting distribution parameters are also optimized during training; this obviates the need for manual tuning, and leads to an improvement in test power relative to CFD. We demonstrate experimentally that our proposed method outperforms WGAN and MMD-GAN variants on a variety of unsupervised image generation benchmarks.
BibTeX
@inproceedings{ansari2020characteristic,
  author = {Abdul Fatir Ansari and Jonathan Scarlett and Harold Soh},
  title = {{A Characteristic Function Approach to Deep Implicit Generative Modeling}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2020},
  url = {https://openreview.net/forum?id=BAxbfxjkDe9},
}
Figure