Topological Representational Similarity Analysis

• tRSA is a method that extends classical RSA by incorporating topological features to provide robust analyses of complex representation structures.
• It employs geo-topological transforms, Vietoris–Rips complexes, and persistent homology to capture both local and global data patterns.
• tRSA enhances model selection and robustness in applications such as fMRI analysis, DNN layer identification, and single-cell developmental studies.

Topological Representational Similarity Analysis (tRSA) is a methodological extension of classical Representational Similarity Analysis (RSA) that incorporates topological features, enabling analyses of neural, biological, and artificial representations to be robust to noise and idiosyncracies while remaining sensitive to both geometric and topological structure. tRSA leverages advances in topological data analysis (TDA), nonlinear transforms of dissimilarity matrices, and bootstrapped statistical estimates to produce summary statistics and distances that adjudicate among candidate models or brain areas by their computational and structural signatures (Lin et al., 2023, Lin, 2024, Easley et al., 2023).

1. Mathematical Foundations of Classical and Topological RSA

Conventional RSA centers on the Representational Dissimilarity Matrix (RDM), where $N$ stimuli or conditions yield response vectors $x_i\in\mathbb{R}^M$, resulting in $D\in\mathbb{R}^{N\times N}$ with entries $D_{ij} = d(x_i,x_j)$, typically computed with Euclidean, Mahalanobis, or correlation-based distances. RSA investigates the representational geometry—how patterns of responses distinguish stimuli across sensors, neurons, brain regions, or models (Lin et al., 2023, Lin, 2024).

tRSA generalizes RSA by augmenting or transforming the RDM to capture not just the geometry (the full metric structure of distances), but also the topology—higher-order and neighborhood structure that is more robust to local noise and individual variation. This transition employs geo-topological transforms and constructs from algebraic topology (Vietoris–Rips complexes, persistent homology), embedding the analysis within the TDA paradigm (Easley et al., 2023).

2. Geo-Topological Transforms and Summary Statistics

A central methodological advance in tRSA is the piecewise-linear, monotonic geo-topological (GT) transform of normalized dissimilarities:

$$GT_{l,u}(r_{ij}) = \begin{cases} 0, & r_{ij} \le l \ \frac{r_{ij} - l}{u-l}, & l < r_{ij} < u \ 1, & r_{ij} \ge u \end{cases}$$

with $r_{ij}$ the normalized RDM entry and thresholds $0\le l < u \le 1$ (quantiles or absolute values). This mapping compresses the influence of small (noise-dominated) and large (idiosyncratic, global) distances, emphasizing the intermediate range that preserves the local manifold structure and neighborhood graph (topological signature). The classical RDM is recovered at $l=0, u=1$, while for $l\approx u$, the transform yields an effective adjacency or connectivity matrix, saturating geometric to topological emphasis (Lin et al., 2023, Lin, 2024).

Applying $GT_{l,u}$ entrywise to the RDM yields the Representational Geo-Topological Matrix (RGTM), a family of summary statistics parameterized by $(l,u)$ that allows the analyst to interpolate between geometric and topological characterization of a representation.

3. Topological Constructs: Vietoris–Rips Complexes and Persistent Homology

Moving beyond metric and graph-based transforms, tRSA harnesses tools from algebraic topology to analyze representational spaces:

• Vietoris–Rips Complex $VR(r)$: For a metric space $(X,d)$ and threshold $r$, $VR(r)$ contains all $k$-simplices for which each pair of vertices are within $r$ of each other.
• Persistent Homology: Tracks the emergence and disappearance (“birth” and “death”) of $k$-dimensional holes as $r$ grows, summarizing this information with Betti numbers $\beta_k(r)$ and persistence diagrams $\mathrm{Pers}_k = \{ (b_i, d_i) \}$ (Easley et al., 2023).

Comparing two representations then involves distances between persistence diagrams (e.g., bottleneck, Wasserstein), or $L^2$ distances between summary Betti curves across the filtration parameter.

A crucial enhancement is the topological bootstrap: resampling the representation and estimating the prevalence score $\pi(\gamma)$ for each homological feature $\gamma$, which quantifies the stability of that feature under resampling and discounts topological "noise" in distance computations (Easley et al., 2023).

4. Statistical Distances and Combined Geometry–Topology Metrics

tRSA enables flexible, parameterized distances between representations by convexly combining geometric and topological components:

$$d_{\mathrm{tRSA}}(A,B) = \alpha\,d_{\mathrm{geom}}(A,B;l,u) + (1-\alpha)\,d_{\mathrm{top}}(A,B)$$

where $d_{\mathrm{geom}}$ is calculated via $L^2$ vectorized discrepancies between RGTMs (or RDMs), and $d_{\mathrm{top}}$ is the distance between topological summaries (matched Betti curves, bottleneck/persistence distances) (Lin et al., 2023). Setting $\alpha=1$ recovers traditional, geometry-only RSA; $\alpha=0$ gives a topology-only analysis.

For persistent-homology-based approaches, prevalence-weighted Wasserstein distances are deployed:

$$W_{p,\pi}(D,D') = \left[ \inf_{\varphi:D\to D'} \sum_{x\in D} \pi_X(x) \,\|x - \varphi(x)\|^p \right]^{1/p}$$

where the prevalence $\pi_X(x)$ down-weights unstable features (Easley et al., 2023). This stabilizes comparison and captures reproducible topological organization.

5. Extensions: Adaptive Dependence, Dynamics, and Single-Cell Analysis

tRSA methodologies extend to frameworks for detecting dependencies and analyzing temporal or single-cell data:

• Adaptive Geo-Topological Dependence Measure (AGTDM): Maximizes distance correlation over the family of GT transforms, enabling robust detection of linear and nonlinear dependence, outperforming classical dCor, HSIC, MIC, and KNN mutual information in synthetic and real data environments (Lin, 2024).
• Procrustes-Aligned Multidimensional Scaling (pMDS): Captures time-evolving neural representations by MDS embedding of RDMs at each timestep, then aligning the resulting trajectories to a common reference using Generalized Procrustes Analysis.
• Temporal Topological Data Analysis (tTDA): Constructs multi-parameter Rips filtrations on high-dimensional time-stamped data, allowing trajectories in both feature and time space (e.g., developmental trajectories in single-cell data).
• Single-Cell Topological Simplicial Analysis (scTSA): Quantifies high-order structure in cell populations using Rips complexes, landmark subsampling, and normalized simplicial complexity, revealing biological developmental stages and population differentiation (Lin, 2024).

6. Applications and Empirical Evaluations

Empirical studies demonstrate tRSA’s utility in model selection, brain-region identification, and biological data analysis:

• fMRI Region Identification: Applications show that compressing RDM extremes with topology-sensitive GT transforms yields region identification accuracy matching or surpassing classical geometry-only RSA, with improved robustness to noise and intersubject differences (Lin et al., 2023, Lin, 2024).
• DNN Layer Identification: Layer-identification accuracy in All-CNN-C models is optimized at intermediate (topology-enhanced) GT settings for low noise; pure geometry is preferable at high noise, reflecting the signal-to-noise adaptation possible with tRSA parameter tuning (Lin et al., 2023, Lin, 2024).
• Single-Cell and Developmental Trajectories: scTSA and tTDA methods accurately recover known biological transitions (e.g., zebrafish gastrulation), outperform conventional clustering by integrating spatial and temporal topology (Lin, 2024).
• Topology in High-Dimensional Data: Prevalence-weighted persistent homology is particularly sensitive to global representational organization in large-scale neuroimaging and is robust to modality or dimensionality mismatches (Easley et al., 2023).

7. Interpretability, Robustness, and Theoretical Implications

tRSA produces a continuum of summary statistics sensitive to the analyst’s choice of geometry–topology emphasis, allowing adaptation to noise structure, sampling variation, and desired theoretical focus. Geo-topological transforms provide honed sensitivity to neighborhood structure over both fine and global scales, while prevalence estimates and adaptive measures improve robustness and interpretability.

Theoretical analyses suggest that discriminability boundaries are primarily encoded in the local topology of the population code; global geometry beyond a certain regime confers diminishing representational distinctiveness. Topological characterization thus formalizes hypotheses regarding invariance to transformations, robustness under resampling, and the computational significance of representational “shapes” (Lin et al., 2023, Lin, 2024, Easley et al., 2023).

---

Key References:

• "The Topology and Geometry of Neural Representations" (Lin et al., 2023)
• "Topological Representational Similarity Analysis in Brains and Beyond" (Lin, 2024)
• "Comparing representations of high-dimensional data with persistent homology: a case study in neuroimaging" (Easley et al., 2023)