Dimensionality Reduction Beyond PCA — Theory
1. WHY — Supervised Separation and Nonlinear Visualization
PCA finds linear directions of maximum variance — it ignores labels entirely. Two common situations where PCA falls short:
-
Supervised classification. Class-discriminative directions may not align with high-variance directions. A feature with large variance can be pure noise across classes, while a low-variance feature perfectly separates them.
-
Nonlinear manifolds. Real data often lies on curved surfaces. PCA can only find flat subspaces, so it distorts local neighborhoods on a manifold.
LDA (Linear Discriminant Analysis) addresses (1) by maximizing between-class separation relative to within-class spread. t-SNE addresses (2) by preserving local pairwise similarities through a nonlinear, non-parametric embedding.
2. WHAT — Notation and Setup
2.1 Common Notation
| Symbol | Type | Meaning |
|---|---|---|
| \(X \in \mathbb{R}^{n \times d}\) | matrix | data matrix, \(n\) samples, \(d\) features |
| \(y_i \in \lbrace1, \ldots, C\rbrace\) | scalar | class label of sample \(i\) |
| \(n_k\) | scalar | number of samples in class \(k\) |
| \(\mu\) | vector | global mean \(\frac{1}{n}\sum_i x_i\) |
| \(\mu_k\) | vector | mean of class \(k\): \(\frac{1}{n_k}\sum_{i: y_i=k} x_i\) |
| \(S_W\) | matrix | within-class scatter matrix |
| \(S_B\) | matrix | between-class scatter matrix |
| \(w \in \mathbb{R}^d\) | vector | projection direction |
2.2 LDA Assumptions
- Labeled data with \(C \geq 2\) known classes.
- Gaussian class-conditional distributions \(x \mid y=k \sim \mathcal{N}(\mu_k, \Sigma_k)\).
- Equal covariance across classes: \(\Sigma_1 = \cdots = \Sigma_C = \Sigma\) (homoscedasticity).
- Non-singular \(S_W\) (requires \(n > d\) or regularization).
2.3 t-SNE Notation
| Symbol | Type | Meaning |
|---|---|---|
| \(p_{j \mid i}\) | scalar | conditional similarity of \(x_j\) to \(x_i\) in high-d |
| \(p_{ij}\) | scalar | symmetrized joint similarity in high-d |
| \(q_{ij}\) | scalar | joint similarity in low-d (Student-t kernel) |
| \(z_i \in \mathbb{R}^m\) | vector | low-d embedding of point \(i\) (typically \(m = 2\)) |
| \(\sigma_i\) | scalar | bandwidth of Gaussian kernel around \(x_i\) |
| \(\text{Perp}\) | scalar | perplexity parameter (effective number of neighbors) |
3. HOW — Fisher's Linear Discriminant Analysis
3.1 Scatter Matrices
Within-class scatter measures how spread out each class is internally:
Between-class scatter measures how separated the class means are:
Total scatter decomposes as \(S_T = S_W + S_B\), where \(S_T = \sum_{i=1}^{n} (x_i - \mu)(x_i - \mu)^\top\).
3.2 Fisher's Criterion
Project all data onto direction \(w\). The projected class mean is \(\tilde{\mu}_k = w^\top \mu_k\) and the projected within-class variance is \(\tilde{s}_k^2 = w^\top S_W^{(k)} w\). Fisher seeks the direction that maximizes class separation relative to within-class spread:
This is a generalized Rayleigh quotient. Taking the gradient and setting to zero:
which is the generalized eigenvalue problem:
Result: The optimal LDA directions are the eigenvectors of \(S_W^{-1} S_B\) corresponding to the largest eigenvalues. At most \(\min(d, C-1)\) directions are non-trivial because \(\text{rank}(S_B) \leq C - 1\).
3.3 Two-Class Special Case
For \(C = 2\), \(S_B = n_1 n_2 / n \cdot (\mu_1 - \mu_2)(\mu_1 - \mu_2)^\top\) has rank 1. The single discriminant direction (up to scale) is:
Result: In the two-class case, LDA reduces to a single closed-form direction — no eigendecomposition needed.
3.4 Regularized LDA
When \(d > n\) or features are highly collinear, \(S_W\) is singular. Add regularization:
for a small \(\alpha > 0\). This is equivalent to assuming a small isotropic noise floor in each class.
4. HOW — t-SNE
4.1 High-Dimensional Similarities
For each pair \((i, j)\), define a conditional probability that \(x_j\) is a neighbor of \(x_i\) under a Gaussian centered at \(x_i\):
Set \(p_{i \mid i} = 0\). The bandwidth \(\sigma_i\) is chosen so that the entropy of the conditional distribution \(P_i\) matches \(\log(\text{Perp})\), where \(\text{Perp}\) is the user-specified perplexity (typically 5–50). Higher perplexity → larger effective neighborhood → more global structure preserved.
Symmetrize: \(p_{ij} = \frac{p_{j \mid i} + p_{i \mid j}}{2n}\).
4.2 Low-Dimensional Similarities (Student-t Kernel)
In the embedding space, use a Student-t distribution with 1 degree of freedom (i.e., a Cauchy kernel):
Why Student-t? In high dimensions, moderate-distance points become nearly equidistant (concentration of measure). In low dimensions, there is less "room" to place moderately distant points — this is the crowding problem.
The heavy tails of the Student-t distribution allow moderately distant high-d neighbors to be placed further apart in low-d without incurring a large cost, while still keeping true neighbors close.
4.3 KL Divergence Objective
t-SNE minimizes:
Because KL divergence is asymmetric: - Large \(p_{ij}\), small \(q_{ij}\) (nearby points mapped far apart): high cost — local structure is preserved. - Small \(p_{ij}\), large \(q_{ij}\) (distant points mapped close): low cost — global distances are not enforced.
This is why t-SNE preserves local neighborhoods but not global geometry.
4.4 Gradient and Attraction/Repulsion Forces
The gradient of \(C\) with respect to \(z_i\) is:
This decomposes into two forces on each embedded point:
- Attraction (when \(p_{ij} > q_{ij}\)): point \(z_j\) pulls \(z_i\) closer — neighbors in high-d that are too far apart in low-d.
- Repulsion (when \(p_{ij} < q_{ij}\)): point \(z_j\) pushes \(z_i\) away — non-neighbors in high-d that are too close in low-d.
Result: t-SNE performs gradient descent on the KL divergence. At convergence, attractions and repulsions balance: nearby high-d points cluster together in low-d, while distant points spread apart.
4.5 Perplexity
Perplexity controls the effective number of neighbors:
- Low perplexity (5–10): focus on very local structure; may fragment clusters.
- High perplexity (30–50): more global context; clusters become rounder.
- Too high perplexity (\(\geq n\)): similarities become nearly uniform; embedding becomes meaningless.
5. Comparison: PCA vs LDA vs t-SNE
| Property | PCA | LDA | t-SNE |
|---|---|---|---|
| Supervision | Unsupervised | Supervised | Unsupervised* |
| Objective | Max variance | Max class separation | Preserve local neighborhoods |
| Linearity | Linear | Linear | Nonlinear |
| Parametric | Yes (\(W_k\)) | Yes (\(W\)) | No (no explicit mapping) |
| Max components | \(\min(n, d)\) | \(C - 1\) | Any (usually 2–3) |
| Preserves | Global variance | Between-class / within-class ratio | Local pairwise distances |
| Invertible | Yes (reconstruction) | Approximately | No |
* t-SNE can use labels for coloring but does not optimize over them.
6. Failure Cases
6.1 LDA Failures
-
Non-Gaussian classes. LDA assumes Gaussian class-conditionals. If classes are multi-modal or have different shapes, the scatter matrices misrepresent the structure.
-
Singular \(S_W\). When \(n < d\) or features are linearly dependent, \(S_W\) is rank- deficient and \(S_W^{-1}\) does not exist. Fix: regularize (\(S_W + \alpha I\)) or use PCA pre-processing to reduce to \(n - 1\) dimensions first.
-
At most \(C - 1\) directions. For binary classification, LDA gives exactly one discriminant direction — insufficient for complex visualization.
-
Homoscedasticity violation. If class covariances differ substantially, the pooled \(S_W\) is a poor summary and LDA may give suboptimal separation.
6.2 t-SNE Failures
-
Stochastic output. Different random initializations yield different embeddings. Conclusions must be robust across runs.
-
Perplexity sensitivity. The visual appearance of clusters depends heavily on perplexity. Always try multiple values (e.g., 5, 15, 30, 50).
-
No global structure. Distances between clusters and relative cluster sizes in a t-SNE plot are not meaningful. Two clusters far apart in the plot may or may not be far apart in the original space.
-
Crowding problem. Without the heavy-tailed kernel, moderate-distance points in high-d crush together in low-d. The Student-t kernel mitigates but does not eliminate this for very high-dimensional data.
-
Non-parametric. t-SNE does not produce a mapping function. New points cannot be projected without re-running the full optimization (or using parametric extensions).
-
Computational cost. Naive t-SNE is \(O(n^2)\) per iteration (pairwise distances). Barnes–Hut approximation reduces this to \(O(n \log n)\).
7. Connections
- PCA: Unsupervised linear baseline; LDA extends PCA with label information; t-SNE provides a nonlinear alternative.
- Information Theory: t-SNE minimizes KL divergence; perplexity is defined via entropy.
- Clustering: t-SNE is often used to visualize cluster structure (but does not itself cluster).
- Autoencoder: Nonlinear parametric dimensionality reduction; can be seen as a learnable t-SNE alternative.
- Geometry of ML: Subspace projection, manifold embedding, and metric preservation are geometric themes shared across PCA, LDA, and t-SNE.
8. References
- Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7(2), 179–188.
- van der Maaten, L., & Hinton, G. (2008). Visualizing data using t-SNE. Journal of Machine Learning Research, 9, 2579–2605.
- Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction (2nd ed.). Springer. Chapter 4.3: Linear Discriminant Analysis.