Pairwise Relative Shift (PARS)

• Pairwise Relative Shift (PARS) is a mathematically precise framework that analyzes pairwise differences to infer global structure across domains such as graph clustering and sequence alignment.
• It uses adaptive shifts, spectral methods, and projected power iterations to optimize pairwise relationships, ensuring balanced partitions and accurate alignment even under noise.
• PARS enhances interpretability in diverse applications by decomposing global shifts into actionable, fine-grained contributions, aiding diagnostics in text, EEG, and distributional analyses.

Pairwise Relative Shift (PARS) encompasses a family of mathematically precise frameworks, methodologies, and algorithms for analyzing and operationalizing pairwise differences or shifts between entities. Across domains such as graph-based clustering, sequence alignment, self-supervised learning, and empirical distribution comparisons, PARS formalizes how local, pairwise relations yield global structure in either discrete or continuous spaces. Central to PARS is the explicit characterization of pairwise relationships, either as shifts in similarity, alignment differences, or distributional contributions.

1. Mathematical Foundations and Formal Definitions

PARS quantifies the difference, similarity, or shift between all pairs in a set of entities—nodes in a graph, tokens in a sequence, or features in empirical distributions. The specific formalism depends on context, but always reduces to aggregating pairwise relationships to infer global structure.

a) Graph Clustering: Adaptive Pairwise Shift

Given a symmetric similarity matrix $S=[s_{ij}]\in\mathbb{R}^{n\times n}$ on $n$ objects $O=\{1,\dots,n\}$, and a $K$-way partition $c:O\to\{1,\dots,K\}$, PARS replaces the usual MinCut objective with a "shifted" version:

• The shift $\Delta_{ij}$ is adaptively computed for each pair to enforce balance:

$$\Delta_{ij} = \frac{1}{n}\sum_{p=1}^n s_{ip} + \frac{1}{n}\sum_{p=1}^n s_{pj} - \frac{1}{n^2}\sum_{p,q=1}^n s_{pq}$$

• The shifted similarity $S'_{ij} = s_{ij} - \Delta_{ij}$, yielding the objective:

$$R^{PARS}(c;S') = - \sum_{k=1}^K\sum_{i,j\in O_k} S'_{ij}$$

This adaptive regularization encourages balanced partitions and mitigates the tendency toward trivial singleton clusters typical of classical MinCut, by penalizing cluster size and supporting negative similarities (Chehreghani, 2021).

b) Joint Alignment: Modulo-Difference Recovery

In the joint alignment setting, observations $y_{ij} = (x_i - x_j) \bmod m$ for unknown $n$0 are used to recover the global vector $n$1:

• The PARS objective is the maximum-likelihood estimator for assignment consistent with all observed pairwise relative shifts.
• Optimization is recast as a quadratic program over indicator variables, leveraging block matrices and projected power iterations (Chen et al., 2016).

c) Distributional Differences: Decomposition of Shift

Given distributions $n$2 over $n$3 features with weights $n$4, PARS expresses the total shift in a weighted average as:

$n$5

This decomposes any global functional shift (frequency, sentiment, entropy) into per-feature pairwise contributions, enabling fine-grained diagnostics and visualization (Gallagher et al., 2020).

d) Self-supervised Time-series Learning: Relative Temporal Shift

In the context of time series such as EEG, PARS defines the pretext loss for self-supervised learning as the mean-squared error in predicting normalized pairwise temporal shifts:

$n$6

where $n$7 are masked positions of sampled windows, forcing encoders to capture long-range temporal dependencies (Sandino et al., 14 Nov 2025).

2. Algorithmic Frameworks and Optimization

PARS methodologies feature tailored optimization techniques that exploit the structure of pairwise relations.

a) Graph Clustering via Shifted Min-Cut

The shifted cost function is non-convex and NP-hard (Chehreghani, 2021). Efficient local search is enabled by:

• Greedy reassignment: Move objects to clusters minimizing $n$8, tracked via incrementally maintained intra-cluster sums.
• Computational cost: $n$9 per sweep; avoids eigendecomposition required by spectral clustering.
• Frank-Wolfe style convergence: Achieves $O=\{1,\dots,n\}$0 error decay after $O=\{1,\dots,n\}$1 iterations.

b) Spectral and Projected Power Methods for Joint Alignment

A two-stage optimization procedure is employed for discrete circular alignment (Chen et al., 2016):

1. Spectral initialization: Find a low-rank (rank-$O=\{1,\dots,n\}$2) approximation, project onto the product of simplices to yield an initial guess.
2. Projected power iterations: Iteratively update by linear map followed by projection onto the discrete feasible set. The error contracts geometrically, guaranteeing exact recovery under appropriate random-correlation models.

c) Temporal Shift Pretraining in EEG

• Encoder: Transformer architecture with linear patch embeddings and masked positional tokens.
• Decoder: Cross-attention module using pairwise concatenated latent vectors to regress $O=\{1,\dots,n\}$3.
• Loss: Mean-squared error over the antisymmetric pairwise shift matrix; hyperparameters such as number of patches ($O=\{1,\dots,n\}$4), masking ratio ($O=\{1,\dots,n\}$5), and learning schedule are tuned for optimal signal composition encoding.

3. Applications Across Domains

PARS enables principled solutions and improved performance in diverse machine learning and data analysis scenarios.

Domain	PARS Instantiation	Core Output/Benefit
Graph clustering	Shifted similarity (PARS)	Balanced, robust partitioning; negative edge weighting
Joint discrete alignment	Modulo-difference inference	Exact label recovery under noise, efficient $O=\{1,\dots,n\}$6 steps
EEG self-supervised pretraining	Pairwise shift regression	Encodes long-range dependencies, boosts label efficiency
Text comparison	Word shift decomposition	Feature-level contributions to sentiment, entropy, etc.

• In clustering, experimental results across UCI and document datasets show that PARS outperforms K-means, spectral clustering, GMM, and other baselines by significant margins in Adjusted Mutual Information, Rand Index, and V-Measure (Chehreghani, 2021).
• In EEG representation learning, PARS-pretrained models achieve superior balanced accuracy and Kappa in transfer tasks over MAE and position prediction baselines, and excel in low-label settings (Sandino et al., 14 Nov 2025).
• In text comparison, generalized word shift graphs enable interpretable, per-feature decomposition of differences in sentiment, topic, or information-theoretic measures (Gallagher et al., 2020).

4. Connections to Related Frameworks

PARS structurally subsumes or connects to several existing paradigms:

• Correlation Clustering: The shifted MinCut with negative similarities is equivalent—up to constants—to correlation clustering with signed weights; their optima coincide. PARS thus provides an algorithmic bridge between classic cut-based partitioning and sign-sensitive graph partitioning (Chehreghani, 2021).
• Masked Position Prediction/MP3: In sequence models, PARS extends position prediction by focusing on learning all $O=\{1,\dots,n\}$7 pairwise relations, enhancing global context modeling compared to one-way position pretext tasks (Sandino et al., 14 Nov 2025).
• Weighted Feature Contributions: In text and distributional analysis, PARS generalizes frequency- and score-based change decompositions to arbitrary weighted functionals, enabling unified visualization of distributional shift (Gallagher et al., 2020).

5. Theoretical Guarantees and Computational Properties

PARS-derived frameworks exhibit strong theoretical properties:

• Clustering: Local search for PARS is proven to converge with $O=\{1,\dots,n\}$8 error, substantially faster than generic nonconvex solvers ($O=\{1,\dots,n\}$9). Incremental updates reduce per-move complexity to $K$0 (Chehreghani, 2021).
• Alignment: Under random corruption or general noise, spectral-plus-projected-power methods guarantee geometric contraction of misclassification, with exact recovery in $K$1 steps given sufficient noise separation (Chen et al., 2016).
• Complexity Considerations: While enumeration over all pairs is $K$2 or worse, algorithmic designs (batching, efficient mat-vec via FFT, approximate decoders) maintain tractability for practical $K$3, $K$4, $K$5.

6. Limitations, Extensions, and Practical Considerations

PARS approaches are subject to domain-dependent constraints and offer avenues for extension:

• Balancing Local and Global Learning: In EEG and similar temporal domains, PARS captures global order but may complement, rather than substitute for, reconstruction-based or local-context learning objectives (Sandino et al., 14 Nov 2025).
• Memory Scalability: The quadratic scaling of pairwise predictions can stress computational resources, motivating the development of efficient sampling and decoding schemes.
• Domain Extension: Adapting PARS to multimodal, long-duration, or frequency-domain signals, and merging with attention-based or contrastive feature learning, constitutes an active direction.
• Visualization and Interpretability: In text analysis, the sum decomposition offers both diagnosis (e.g., detecting artifacts or measurement drift) and mechanistic interpretation, but requires careful treatment of low-frequency features, reference selection, and smoothing.

7. Empirical Evaluation and Benchmark Performance

Empirical validation across multiple papers demonstrates the efficacy of PARS:

• Clustering: On UCI datasets (Breast, Ecoli, Pima), PARS achieves or matches state-of-the-art clustering quality without requiring kernel parameter tuning or eigenvector computation (Chehreghani, 2021).
• EEG Decoding: In low-data sleep staging and transfer learning benchmarks, PARS yields top balanced accuracy and Kappa/AUROC in three of four tasks, with improvements robust to random seed and subject-level variation (Sandino et al., 14 Nov 2025).
• Textual Shift Explanation: In comparative text analysis, PARS-derived word shift graphs recover the dominant contributors to informational or affective shift in interpretable form, aiding substantive analyses in computational social science (Gallagher et al., 2020).

In summary, Pairwise Relative Shift methodologies formalize and operationalize the use of pairwise relationships as building blocks for a range of inferential, representational, and interpretive tasks, yielding both rigorous theory and practical performance across modalities.