Same-State Pairs Make One JEPA Proof of the Platonic Representation Hypothesis Vacuous
Organizations: Shenzhen University of Advanced Technology
Abstract
In a Joint-Embedding Predictive Architecture (JEPA), the JEPA proof of the Platonic Representation Hypothesis (PRH) assumes that every observation determines the world state and that a spectral gap makes minimizers unique up to rotation. Under the hard-whitening population objective, same-state positive pairs make the induced world kernel the identity: every centered whitened state function is a global minimizer, and, when , two minimizers can have best-rotation squared residual . The uniqueness claim is conditional; for , where is the number of world states, this limit makes its premise unsatisfiable, so the claim has no content there. The experiments that motivate it leave the assumption: a caption sees only part of an image, and the paper states no experiment meets its hypotheses exactly. The theorem does not reach those experiments, and for its idealized same-state limit is vacuous. One of the paper's four quantitative results survives: a gap-free lower bound. Its bound vanishes only at perfect scene-explained fraction; it bounds alignment error but cannot certify good-view alignment. A canonical frame is determined only up to a rotation and a latent permutation. Paired samples fix the relative frame at effective rank ; a global rotation remains. Component semantics fix the relative permutation only up to signs; value anchors fix it once the mixing frame is pinned. A resource--gauge ledger records which resource fixes which gauge.
Figures & tables
| Original result | What it asserts | Premise relevant here | Status in the same-state limit and in the experiments |
|---|---|---|---|
| Theorem 1 (with Proposition 1) | shared representation up to one rotation | injective renderings; shared ; minimizers unique up to rotation (spectral gap) | premise empty in the same-state limit for ; experiments leave the injective premise |
| Theorem 2 (kernel mismatch) | misalignment bounded by kernel mismatch | spectral solutions; eigengap | no gap for in the same-state limit; at the theorem applies trivially (bound zero) |
| Theorem 3 (nuisance costs alignment) | gap-free lower bound from the scene-explained fraction | centered, whitened encoders; no gap | survives; positive bound for imperfect views; cannot certify good-view alignment |
| Theorem 4 (views and data set the kernel) | a spectral solution is close to the kernel’s spectral embedding, and paired modalities align within the theorem’s two error terms; a gap separates state from nuisance | gap ( largest nuisance eigenvalue) | no gap for in the same-state limit; at the theorem applies trivially ( ); experiments do not verify its hypotheses |
| Proposition 3(i) | a good-view JEPA learns a function of the state ( ) | gap | no gap for ; trivial at |
| Proposition 2 (k pairs) | pairs in general position identify the rotation | paired rank | refined below by the partial-rank case |
| Resource | Continuous ? | Discrete ? |
|---|---|---|
| Paired samples | relative frames only (at effective rank ); global remains | no |
| Homoscedastic known noise | no (whitening-neutral) | no |
| Heteroscedastic calibrated noise | exact under identifiability: yes if (4) has no continuous part | no (absolute remains a gauge under free factorization) |
| Correspondence anchors | no (labels only) | yes, up to a signed-permutation subgroup |
| Value anchors (known coordinate values) | no (the frame must already be pinned) | yes, for nonzero anchored coordinates |
| Sparse or nonnegative constraints | only if rotation symmetry is broken | sparse: yes; nonnegative: no |
| Test | Type | Result | Statement |
|---|---|---|---|
| Partial pairing, | L2 estimated | : (no-information gauge), : one gauge dimension, : discrete, : , : | effective-rank threshold |
| ICA and component fingerprints, | L2 estimated | relative-permutation accuracy ( – ); same-shape control mean ( – ), component-level chance | discrete gauge needs semantics |
Representation Identifiability
Representation identifiability studies when learned internal codes can be uniquely recovered from observations or behavior, and which ambiguities remain.
See all 43 papers