Unfolded Laplacian Spectral Embedding (ULSE)

• Unfolded Laplacian Spectral Embedding (ULSE) is a framework that creates stable, dynamic node embeddings by unfolding time-evolving Laplacians through singular value decomposition.
• It separates anchor and dynamic embeddings, capturing both invariant and temporal features with rigorous spectral geometry and stability guarantees.
• ULSE integrates dual convex optimization and root-Laplacian operations to minimize distortion, improve clustering, and establish conductance bounds across dynamic networks.

Unfolded Laplacian Spectral Embedding (ULSE) is a theoretically-grounded framework for dynamic network representation that utilizes spectral properties of normalized graph Laplacians and their extensions. ULSE provides principled embeddings for time-evolving graphs, rigorously linking spectral geometry, stability guarantees, and conductance bounds. It generalizes static Laplacian spectral embeddings and “unfolds” temporal graph information through singular value decomposition (SVD) of concatenated Laplacian matrices, yielding both anchor and dynamic embeddings with strong stability properties under stochastic network models (Ezoe et al., 18 Aug 2025). This construct is further illuminated via dual convex optimization and root-Laplacian operators in related work (Shakeri et al., 2019, Choudhury, 2023), which collectively form the modern mathematical basis for ULSE.

1. Mathematical Preliminaries and Laplacian Embedding

ULSE operates on a sequence of $T$ graphs $G^{(1)}, \dots, G^{(T)}$ with common node set, represented by adjacency matrices $A^{(t)} \in \{0,1\}^{n\times n}$ and degree matrices $D^{(t)} = \mathrm{diag}(d_1^{(t)}, ..., d_n^{(t)})$, with individual degrees $d_i^{(t)} = \sum_j A_{ij}^{(t)}$ at each time. The normalized Laplacian for each snapshot is defined as

$$L^{(t)} = D^{(t)\,-\tfrac12} (D^{(t)} - A^{(t)}) D^{(t)\,-\tfrac12} \in \mathbb{R}^{n\times n}.$$

Classical static spectral embedding selects the $k$ smallest nontrivial eigenpairs, embedding node $i$ as the $i$th row of $$Y^{(t)} = [u_1^{(t)} \mid \dots \mid u_k^{(t)}] \in \mathbb{R}^{n\times k}$$.

The root Laplacian operator $S = L^{1/2}$, defined via spectral decomposition, is also utilized to moderate sensitivity to larger eigenvalues, reducing embedding distortion and improving clusterability (Choudhury, 2023). If $L = U \Lambda U^T$ with non-negative eigenvalues, then $S = U \Lambda^{1/2} U^T$.

2. Theoretical Formulation of ULSE

ULSE extends the Unfolded Adjacency Spectral Embedding (UASE) from adjacency to normalized Laplacians. Its core procedure involves:

• Unfolding and SVD: Stack Laplacians $L^{(t)}$ for all $t$ to form the unfolded matrix $L \in \mathbb{R}^{n \times nT}$, then perform a rank-$(d+1)$ SVD:

$L = U \Sigma V^\top + \cdots,$

with $U \in \mathbb{R}^{n\times(d+1)}$, singular values $\Sigma$, and $V \in \mathbb{R}^{nT\times(d+1)}$, partitioned into $T$ blocks $V^{(t)}$.

• Anchor and Dynamic Embeddings: The anchor (time-invariant) embedding is $Z = U \Sigma^{1/2}$; dynamic embeddings at time $t$ are

$$\widetilde Y^{(t)} = V^{(t)} \Sigma^{1/2}, \qquad Y^{(t)} = \widetilde Y^{(t)} - Z.$$

This decomposition isolates time-invariant structure and corrects for longitudinal drift.

• Complexity: The main computational cost is truncated SVD on size $n \times nT$, approximately $O(n^2 T d)$.

An alternative normalization strategy (ULSE-n2) employs partially aggregated degree normalization: $L^{(t)} = D^{(1:T)-1/2} A^{(t)} D^{(t)-1/2}$, with $D^{(1:T)} = \sum_{s=1}^T D^{(s)}$.

3. Stability Analysis in Dynamic Stochastic Models

ULSE is rigorously analyzed under dynamic Stochastic Block Models (SBMs) with inhomogeneous edge probabilities $P^{(t)}$. Key stability results include:

• Cross-Sectional Stability: Within each snapshot, nodes with identical $P^{(t)}_{i:}$ receive identical embeddings.
• Longitudinal Stability: Across time, if $P^{(t)}_{i:} = P^{(u)}_{i:}$ for $i$, then $Y^{(t)}_{i:} = Y^{(u)}_{i:}$.

The central theorem states for $d = K - 1$ ($K$ is the SBM community count), there exists orthogonal alignment $W \in O(d+1)$ so that

$$\max_{i,t} \|Y^{(t)}_{i:} - \widetilde Y^{(t)}_{i:} W\| = \mathcal{O}((n\rho)^{-1/2}) \quad \text{a.s.}$$

where $\rho$ controls the expected edge density.

Proof methods adapt spectral norm bounds to normalized Laplacians, analyze singular-vector subspaces, and apply Davis–Kahan or Yu–Wang–Samworth perturbation results. These yield nontrivial guarantees of stable spectral embeddings under noise.

4. Duality, Multiplex Extensions, and Root-Laplacian Approaches

ULSE can be framed as a dual convex optimization problem for multiplex networks (Shakeri et al., 2019). For two-layer multiplex graphs with interlayer weights $w$ under a budget constraint, maximizing the second-smallest eigenvalue leads to embedding collapse/unfold transitions:

• Subcritical Budget: For coupling $c < c^*$, all layer nodes embed at $\pm h$ on the real line, dimension $d=1$.
• Supercritical Budget: For $c > c^*$, the embedding unfolds into higher dimensions, $d = m$, the multiplicity of $\lambda_2(L^*)$.

The ULSE coordinates in such cases are extracted from the Fiedler (second), third, etc., eigenvectors of the total Laplacian. Explicit calculation and example workflows are presented for small $n$.

Root-Laplacian spectral embedding (also termed ULSE in (Choudhury, 2023)) replaces the quadratic Dirichlet energy in Laplacian eigenmaps with its square root, leading to lower spectral distortion and improved theoretical properties:

• ULSE shares eigenvectors with classical Laplacian eigenmaps, but square roots the spectrum, reducing sensitivity to outliers and improving stability under perturbations.
• Large-scale deviations are penalized less stringently, enhancing robustness to noisy or anomalous edges.

5. Cheeger-Style Inequalities and Conductance Bounds

ULSE provides new bounds between embedding singular values and dynamic graph conductance:

$$\frac{\sqrt{\max\{\sigma_k^2 - \min_t \sigma_k(L^{-t})^2, 0\}}}{2} \leq \phi_k(G) \leq \mathrm{poly}(k)\sqrt{\sigma_k}$$

where $\phi_k(G)$ is the $k$-way conductance (max over snapshots), $\sigma_k$ is the $k$-th smallest singular value of the unfolded Laplacian, and $L^{-t}$ omits snapshot $t$. These are established via Weyl’s inequality and static Cheeger relations: $\phi_k(G^{(t)}) \approx \sqrt{\lambda_k(L^{(t)})}$.

This systematic linkage connects spectral embedding quality with graph bottleneck phenomena, informing downstream clustering or partitioning analyses.

6. Empirical Performance and Comparative Analysis

Extensive synthetic and real-data experiments establish the empirical efficacy of ULSE:

• Synthetic SBMs: For $K=3$ merging communities, ULSE (both n1 and n2 forms) accurately tracks latent community transitions, where conventional methods (supra-Laplacians, deep learning baselines) fail cross-sectional and longitudinal stability.
• Real-World Datasets: ULSE variants are evaluated on:
  • Brain connectome networks ($n \approx 100$)
  • School interaction graphs ($n \approx 200$)
  • S&P-500 stock-correlation networks ($n \approx 400$)
  • Comparisons involve nine baseline methods: OMNI, UASE, supra-Laplacian methods (TemporalCut-N/S), node2vec, JODIE, DyRep, TGN, DyGFormer.
• Clustering Results: $k$-means applied to ULSE embeddings yields top-tier accuracy (ACC), normalized mutual information (NMI), adjusted Rand index (ARI), and $\mathrm{F}_1$ scores. ULSE-n1 is best or near-best on all datasets, often surpassing UASE and established deep-learning alternatives. The close performance of ULSE-n2 validates both normalization choices.

Qualitative visualizations (t-SNE) confirm sharp, interpretable clustering, aligned with theoretical stability properties.

7. Algorithmic Implementation and Practical Considerations

The primary ULSE (n1) workflow is summarized below:

Input: {A^{(t)}}_{t=1}^T (adjacency matrices), embed dim d Output: anchor Z ∈ R^{n×(d+1)}, dynamics {Y^{(t)}}_{t=1}^T 1) For each t, compute D^{(t)} = diag(degrees of A^{(t)}) 2) Form L = [ D^{-1/2} (D−A) D^{-1/2} ]_{t=1..T} ∈ R^{n×nT} 3) Compute rank-(d+1) SVD: L ≈ U Σ V^T 4) Partition V into blocks V^{(t)} ∈ R^{n×(d+1)} 5) Z ← U Σ^{1/2} 6) For t=1..T, Y^{(t)} ← V^{(t)} Σ^{1/2} − Z Complexity: O(n^2 T d) (dominated by truncated SVD)

Correct handling of normalization, orthogonality, and singular-vector alignment is essential for ensuring ULSE’s stability properties. SVD robustness and efficient implementation are critical for scalability to large dynamic graphs.

8. Summary and Research Directions

Unfolded Laplacian Spectral Embedding consolidates spectral graph theory, random graph models, and convex optimization into a unified approach for dynamic network representation. Its novel combination of anchor/dynamic embeddings, SVD-based “memory”, dynamic Cheeger-style bounds, and strong empirical performance creates a principled basis for interpretable, stable temporal embeddings (Ezoe et al., 18 Aug 2025, Shakeri et al., 2019, Choudhury, 2023).

Potential future directions involve extending ULSE to multi-layer or multiplex architectures, leveraging root-Laplacian approaches for geometric deep learning, and exploring further connections to manifold learning and graph signal processing frameworks. This suggests broad applicability in time-series graph analytics, neuroscience, financial networks, and multi-modal relational data structures.