Latent on the Spectrum, in JAX

Almost everything here is one function call, jnp.linalg.eigh, applied to a matrix the size of your label set. Diagonalise a label-similarity kernel, keep its strongest modes, and the codebook’s geometry is already decided: that is the explainer’s claim, and the code is short enough to check it in an afternoon. What follows is the classical-MDS embedding with its square-root scaling done right, the flat→simplex and graded→horseshoe morph, kernel-target alignment, and the split of a trained representation into a prototype frame and an information tail.

Prefer a notebook? Run the whole thing on Kaggle: the same code, block by block, with the math written out and every figure reproduced from a real run.

Everything here is plain JAX. The objects are tiny (a C × C label kernel, a handful of features) so the linear algebra is legible.

A codebook is the spectral embedding of a kernel#

How does a target similarity S between classes become actual coordinates? You read them off the spectrum. Ignoring the unit-norm step for a moment, the best rank-d Gram is $U_d\Lambda_d U_d^\top$, and a coordinate matrix that realises it is $\Lambda_d^{1/2}U_d^\top$: the top-d eigenvectors scaled by the square roots of their eigenvalues. If we want cosine codes, we normalise afterward.

import jax
import jax.numpy as jnp


def spectral_codebook(S, d):
    """S: (C, C) symmetric label kernel -> C unit-norm codes in R^d."""
    w, V = jnp.linalg.eigh(S)                 # ascending eigenpairs
    w, V = w[::-1], V[:, ::-1]                # to descending
    coords = V[:, :d] * jnp.sqrt(jnp.clip(w[:d], 0.0, None))   # Lambda_d^{1/2} U_d^T : rows are codes
    return coords / (jnp.linalg.norm(coords, axis=1, keepdims=True) + 1e-9)

That single eigendecomposition is classical multidimensional scaling, the spectral fact behind kernel PCA and Laplacian eigenmaps. The reconstruction error of the rank-d codebook is exactly the tail of the spectrum:

def gram_error(S, d):
    w, V = jnp.linalg.eigh(S)
    w, V = jnp.clip(w[::-1], 0.0, None), V[:, ::-1]
    Shat = (V[:, :d] * w[:d]) @ V[:, :d].T
    return jnp.linalg.norm(Shat - S) / jnp.linalg.norm(S)     # falls as the kept modes capture more of S

The spectrum decides the geometry#

Three label kernels, three spectra, three shapes. The structureless kernel is flat and gives the even simplex; a block kernel is two-peaked and gives clusters; a graded kernel (similarity falling off with class distance) is the horseshoe of Diaconis, Goel & Holmes (2008).

C = 9
i = jnp.arange(C)

flat   = jnp.eye(C) - jnp.ones((C, C)) / C                       # structureless: the simplex
blocks = (i[:, None] // 3 == i[None, :] // 3).astype(jnp.float32)  # 3 superclasses of 3: clusters
graded = jnp.exp(-((i[:, None] - i[None, :]) / 2.2) ** 2)        # falls with class distance: the horseshoe

for name, S in [("flat", flat), ("blocks", blocks - blocks.mean()), ("graded", graded - graded.mean())]:
    codes = spectral_codebook(S, d=2)
    w = jnp.clip(jnp.linalg.eigh(S)[0][::-1], 0.0, None)
    print(name, "top-3 spectrum:", jnp.round(w[:3] / (w[0] + 1e-9), 2))
# flat   -> [1.  1.  1. ]  (no preferred direction -> even ring)
# blocks -> [1.  1.  0. ]  (two dominant modes, then nothing -> clusters)
# graded -> [1.  0.7 0.3]  (a smooth tail -> a curved 1-D manifold, the horseshoe)

The hero figure above is exactly this spectral_codebook(S, 2) evaluated on the flat and graded kernels, with the two embeddings orthogonally aligned so the contrast is legible.

Kernel-target alignment#

The reason this works is an old idea that has aged well: the ideal embedding kernel is the label kernel (Cristianini et al., 2002). The match between them is one cosine between two Gram matrices:

def alignment(A, B):
    return jnp.sum(A * B) / (jnp.linalg.norm(A) * jnp.linalg.norm(B) + 1e-9)


codes = spectral_codebook(graded - graded.mean(), d=2)
gram = codes @ codes.T                                   # the codebook's own similarity
print(alignment(gram, graded - graded.mean()))           # high: the 2-D code already captures the kernel

Raise d and the alignment climbs toward 1 as the codebook is allowed to reproduce more of the kernel’s spectrum. The figure below shows that budget directly: for a few values of d it rebuilds the rank-d Gram from the top eigenmodes and lays the reconstruction next to the target kernel, with alignment and gram_error charted across every budget.

A handful of modes already recover most of a graded kernel, which is the whole reason a low-dimensional codebook works: the kernel’s mass lives in its top eigenvalues, and the rest is a tail you can drop.

The prototype frame and the information spectrum#

What does a trained representation keep of all this? Switch objects, from the target label kernel to a feature matrix, and split it into between-class and within-class covariance. The between-class part $\Sigma_B$ has rank at most C-1 and spans the prototypes: it is the separation channel, the codebook. The within-class part $\Sigma_W$ is everything else: the gradations the codebook does not carry, the information.

def class_covariances(Z, y, C):
    mu = jnp.stack([Z[y == c].mean(0) for c in range(C)])
    g, N = Z.mean(0), Z.shape[0]
    SB = sum((y == c).sum() * jnp.outer(mu[c] - g, mu[c] - g) for c in range(C)) / N
    SW = jnp.mean(jax.vmap(lambda z, c: jnp.outer(z - mu[c], z - mu[c]))(Z, y), 0)
    return SB, SW


SB, SW = class_covariances(Z, y, C)
eig_B = jnp.clip(jnp.linalg.eigvalsh(SB)[::-1], 0.0, None)   # <= C-1 nonzero: the prototype frame
eig_W = jnp.clip(jnp.linalg.eigvalsh(SW)[::-1], 0.0, None)   # the information tail
print("nonzero prototype modes:", int((eig_B > 1e-6 * eig_B[0]).sum()), "  (C-1 =", C - 1, ")")

The split is clean only when the between-class variance dominates, as it does near neural collapse; with large within-class variance the two regimes overlap. The figure below trains a small encoder with cross-entropy and watches it happen: $\Sigma_B$ sharpens into a C-1-mode simplex frame while the $\Sigma_W$ spectrum is ground toward zero (Papyan, Han & Donoho, 2020).

Dark knowledge is the coefficients#

The explainer’s last move: the information that survives lives between the prototypes, as the soft mixture a feature makes over them, the dark knowledge a teacher distils (Hinton et al., 2015). In this frame view it is a coefficient vector, a soft assignment over the prototype frame:

def soft_assignment(z, prototypes, tau=0.1):
    """z: (..., d) feature; prototypes: (C, d). Returns coefficients over the codebook."""
    return jax.nn.softmax((z @ prototypes.T) / tau, axis=-1)


a = soft_assignment(z, mu)        # e.g. [0.70, 0.27, 0.02, 0.01]: "mostly cat, a bit dog"

Sharpen tau → 0 and the vector collapses to one-hot, erasing the relation between classes, which is the effect label smoothing and a low temperature have. The information is the off-one-hot mass. On a graded codebook the point is visible in the geometry: a feature near class k lends its mass to k’s neighbours on the horseshoe, so the assignment encodes which classes are similar, and cooling the temperature burns that structure away.

Rendering the figures#

All four figures are generated with Python, JAX, and matplotlib: the linear algebra in JAX, the drawing in matplotlib. Three are static contrasts (a kernel dial, a budget, a temperature are knobs, not processes); only neural collapse is a real training run, so only it animates.

The codebook figure computes spectral_codebook(S, 2) and the eigenspectrum for the flat and graded kernels, Procrustes-aligned so the contrast reads: a flat ring against a graded horseshoe. The reconstruction figure rebuilds the rank-d Gram and recomputes alignment and gram_error across every budget d. The collapse animation actually trains a small encoder, recomputing class_covariances every few optimizer steps so you watch the $\Sigma_W$ spectrum collapse happen. The dark-knowledge figure fixes a graded codebook and evaluates soft_assignment at a warm and a cold temperature. None is a benchmark; they are visual audits of the spectrum.

Boundary of the result#

A production embedding would use far more classes, real data, and a deep encoder, and the label kernel would be estimated rather than handed to you. The point this companion keeps is the one the explainer named: a codebook is a spectrum, the prototypes are its top, and the information is the tail.

References: classical MDS / kernel-target alignment from Cristianini et al. (2002); the horseshoe from Diaconis, Goel & Holmes (2008); neural collapse from Papyan et al. (2020); dark knowledge from Hinton et al. (2015); the Welch bound from Welch (1974).

@misc{bouhsine2026latentonthespectrumjax,
  author       = {Bouhsine, Taha},
  title        = {Latent on the Spectrum, in JAX},
  year         = {2026},
  month        = {jun},
  howpublished = {\url{https://tahabouhsine.com/blog/latent-on-the-spectrum-jax/}},
  note         = {Blog post, Records of the !mmortal Data Scientist}
}