Denoising deterministic networks using iterative Fourier transforms

We detail a novel Fourier-based approach (IterativeFT) for identifying deterministic network structure in the presence of both edge pruning and Gaussian noise. This technique involves the iterative execution of forward and inverse 2D discrete Fourier transforms on a target network adjacency matrix. The denoising ability of the method is achieved via the application of a sparsification operation to both the real and frequency domain representations of the adjacency matrix with algorithm convergence achieved when the real domain sparsity pattern stabilizes. To demonstrate the effectiveness of the approach, we apply it to noisy versions of several deterministic models including Kautz, lattice, tree and bipartite networks. For contrast, we also evaluate preferential attachment networks to illustrate the behavior on stochastic graphs. We compare the performance of IterativeFT against simple real domain and frequency domain thresholding, reduced rank reconstruction and locally adaptive network sparsification. Relative to the comparison network denoising approaches, the proposed IterativeFT method provides the best overall performance for lattice and Kuatz networks with competitive performance on tree and bipartite networks. Importantly, the InterativeFT technique is effective at both filtering noisy edges and recovering true edges that are missing from the observed network.

💡 Research Summary

The paper introduces a novel denoising technique called IterativeFT that targets adjacency matrices of networks corrupted by both edge pruning and additive Gaussian noise. The core idea is to alternate forward and inverse two‑dimensional discrete Fourier transforms (DFT) with simple mean‑based thresholding operations applied separately in the real (spatial) domain and the frequency domain. In each iteration the algorithm first zeroes out all entries whose absolute value falls below the mean absolute value of the current matrix, then computes the DFT, zeroes out all frequency components whose magnitude is below the mean magnitude, and finally applies the inverse DFT to obtain a new real‑domain matrix. The process repeats until the sparsity pattern in the real domain stabilizes, which the authors argue is guaranteed by the Fourier uncertainty principle: sparsity in one domain inevitably reduces sparsity in the other, leading to convergence at a non‑trivial fixed point rather than the trivial all‑zero solution.

To evaluate the method, the authors generate five benchmark networks: four deterministic models (Kautz, 10×10 lattice, a 3‑ary tree, and a full bipartite graph, each with 108 vertices) and one stochastic model (preferential attachment). For each model they randomly prune a fixed proportion of edges (default 25 %) and add i.i.d. Gaussian noise with configurable standard deviation (default 0.25). They sweep pruning ratios from 0.05 to 0.95 and noise levels from 0 to 1 in steps of 0.05, repeating each configuration ten times. Four baseline denoising approaches are compared: (1) simple real‑domain mean thresholding, (2) simple frequency‑domain mean thresholding, (3) low‑rank reconstruction using a rank‑3 SVD, and (4) locally adaptive network sparsification (LANS).

Performance is measured using two complementary metrics. The F1 score treats any non‑zero entry in the denoised matrix as an edge and compares it to the ground‑truth adjacency, emphasizing correct edge classification. Mean squared error (MSE) quantifies the overall deviation between the denoised and true matrices, capturing both false positives and false negatives even when the network is extremely sparse. Results show that IterativeFT consistently outperforms all baselines on the deterministic Kautz and lattice graphs in both F1 and MSE across the full range of pruning and noise levels. On the tree graph it achieves the lowest MSE and the second‑best F1 (behind LANS at high pruning). For the full bipartite graph, simple real‑domain thresholding yields the highest F1, while low‑rank reconstruction gives the best MSE; IterativeFT remains competitive on F1 but lags on MSE. As expected, on the stochastic preferential‑attachment graph IterativeFT performs poorly, with LANS dominating both metrics.

The authors acknowledge several limitations. The experimental suite is confined to five network types of a single size, and baseline methods are used with fixed hyper‑parameters (e.g., rank‑3 SVD, fixed significance level for LANS), potentially under‑representing their optimal performance. Moreover, IterativeFT can introduce spurious edges, especially in structured matrices such as Kautz, where artificial edge bands appear in the corners of the adjacency matrix. This artifact likely stems from the global nature of the Fourier transform, which can over‑emphasize periodic patterns.

Future work is outlined along four main directions. First, hybrid or ensemble strategies that combine IterativeFT with local sparsification (e.g., LANS) could mitigate artifact generation. Second, a broader simulation campaign covering diverse topologies, varying network sizes, and alternative noise models (e.g., discrete errors, structural perturbations) would test robustness. Third, expanding the set of competing denoising techniques—including non‑linear low‑rank methods, graph‑signal processing filters, and probabilistic edge‑prediction models—would provide a more comprehensive benchmark. Fourth, a deeper theoretical analysis of the convergence behavior under the Fourier uncertainty principle could clarify why deterministic structures are recovered while stochastic ones are not. Finally, applying the method to real‑world weighted networks (biological, social, infrastructural) would demonstrate practical utility.

In summary, IterativeFT offers a simple, hyper‑parameter‑free pipeline that leverages alternating Fourier domain sparsification to recover deterministic network structure from noisy, partially observed data. It excels on regular graphs but struggles with random graphs and may produce artificial connections, highlighting both its promise and the need for further refinement.