What is Local Tangent Space Alignment?
Local Tangent Space Alignment (LTSA) is a nonlinear dimensionality-reduction method that estimates tangent coordinates in overlapping sample neighborhoods and finds global coordinates that best align those local affine charts.
Quick Facts
| Created | Introduced by Zhenyue Zhang and Hongyuan Zha in 2004 |
|---|---|
| Specification | Official Specification |
How It Works
Estimate a tangent chart in every neighborhood
For each sample, LTSA gathers nearby points, subtracts the neighborhood mean, and applies a local SVD or PCA. The leading d left or right singular directions define a tangent basis, while projected neighbor coordinates form a local chart. Translation and rotation inside each chart are arbitrary and cannot be interpreted independently.
The original LTSA paper frames these charts as local approximations to a parameterized manifold observed with noise. The approximation requires neighborhoods small enough for linearization but large and well-distributed enough to estimate rank d.
Align overlapping charts through one global eigenspace
LTSA removes the constant direction and projects each neighborhood onto the complement of its estimated tangent-coordinate span. It scatters these local alignment errors into a global symmetric matrix, then takes the smallest nontrivial eigenvectors as global coordinates. Overlap makes the charts constrain one another.
This is why LTSA is close to LLE computationally but different conceptually. LLE freezes barycentric reconstruction weights; LTSA freezes local tangent-coordinate subspaces. Sign, basis rotation, and nearly repeated global eigenvalues still make individual output axes non-identifiable.
Validate local dimension, overlap, and global consistency
Too few neighbors produce noisy or rank-deficient tangent estimates. Too many cross curvature, branches, boundaries, density transitions, or nearby folds. A wrong output dimension changes every local chart and can create a plausible but invalid global alignment.
Current scikit-learn documentation exposes LTSA as method="ltsa". Audit local singular-value gaps, graph components, overlap residuals, eigengaps, Trustworthiness and Continuity, resampling stability, and held-out utility against PCA, Standard LLE, and Hessian Eigenmaps.
Key Characteristics
- Approximates each sample neighborhood with a local tangent space
- Uses local PCA or SVD coordinates rather than reconstruction weights
- Removes arbitrary local translation before aligning overlapping charts
- Solves a global eigenproblem for mutually consistent coordinates
- Depends on output dimension, neighborhood size, local rank, and overlap
- Is sensitive to curvature, noise, boundaries, branches, and sparse sampling
Common Use Cases
- Unfolding smooth manifolds with reliable local tangent estimates
- Comparing tangent-chart alignment with reconstruction-based LLE
- Analyzing controlled shape, pose, sensor, or process trajectories
- Testing whether one global coordinate system explains overlapping local patches
- Studying the effect of intrinsic-dimension choices on nonlinear embeddings
Example
Loading code...Frequently Asked Questions
How does LTSA differ from Locally Linear Embedding?
LLE preserves neighbor reconstruction weights. LTSA estimates a tangent-coordinate subspace in each neighborhood and seeks global coordinates whose local restrictions align with those charts. They use related sparse eigenproblems but encode different local invariants.
How is the local tangent space estimated in LTSA?
The neighborhood is centered and decomposed with local PCA or SVD. The leading `d` singular directions approximate the tangent basis, and the projected samples become local coordinates. Reliable estimation requires a clear local rank and well-distributed neighbors.
How should the LTSA neighborhood size be selected?
Use enough points to estimate a stable `d`-dimensional tangent space and overlap adjacent charts, but not enough to cross curvature, folds, branches, or density regimes. Compare singular gaps, alignment residuals, connectivity, and stability over a declared range.
Does LTSA estimate the intrinsic dimension automatically?
No. Standard implementations require an output dimension, which also controls every local tangent estimate. Estimate plausible dimensions separately, inspect local singular spectra, and validate several candidates rather than inferring truth from one attractive embedding.
Can LTSA embed unseen samples?
Classical LTSA is transductive. An implementation may interpolate from a fitted neighborhood or train a separate mapping, but chart selection, preprocessing, tangent basis, and global coordinates must remain versioned and tested on shifted samples.