Is the 'low-dimensional subspace that constrains adaptation' one shared mathematical object across neural manifolds and LLM intrinsic dimension — or three analogies that share a shape?

neural-manifoldsdimensionalitycross-domain-convergencefine-tuninglearninglarge-language-models

This is the saved hook behind observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets. The same shape — useful adaptation confined to a low-dimensional subspace of a much larger space — recurs across three literatures:

• Neural manifolds in motor cortex (claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists): the low-D structure is the covariance/principal subspace of population firing.
• Plastic recurrent networks (claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry): the low-D structure is the activity subspace an RNN's recurrent weights already support.
• LLM fine-tuning (claim-aghajanyan-2020-fine-tuning-low-intrinsic-dimension): the low-D structure is the intrinsic dimension of a fine-tuning objective in weight space — a different space (parameters, not activations) from the first two.

The observation note deliberately refuses to assert these are the same math, exactly as question-gradient-geometry-one-object-or-three-analogies refuses it for gradient geometry. The open questions:

1. Same object? Is there a formal reduction linking an activation-space neural manifold to a parameter-space intrinsic dimension — e.g. via the tangent/Jacobian map from weights to activity, or via the Fisher-information geometry that also underlies question-gradient-geometry-one-object-or-three-analogies? Or are activation-subspace and weight-subspace low-dimensionality genuinely distinct phenomena?
2. Why does it recur? The capture noted the Platonic Representation Hypothesis (networks converging on shared low-D representations) as an adjacent candidate explanation for why low-D structure is ubiquitous. Worth chasing as a possible mechanism.

Candidate empirical next move (a direct test of the analogy): are LLM "within-manifold" edits (low-rank LoRA updates) safe while "outside-manifold" edits cause catastrophic forgetting — the AI mirror of Sadtler's within-/outside-manifold asymmetry? If so, the bridge is more than metaphor. If a genuine reduction or a clean empirical parallel is found, this cluster may warrant a "low-dimensional-adaptation" MOC.

Progress — 2026-07-25 (partial; question stays open). Promotion of 10-inbox/raw/2026-07-19-is-the-low-dimensional-subspace-that-constrains-adaptation.md substantially advances sub-question 1 ("same object?") and leaves sub-question 2 ("why does it recur?" / Platonic Representation Hypothesis) and the empirical within-/outside-manifold test untouched — so this is a progress note, not a closure.

What settled on sub-question 1, from four new claim-notes reading the primary definitions on each side:

• The neuroscience notion is defined on activity — claim-jazayeri-ostojic-2021-neural-manifold-intrinsic-dimension-parametrizes-activity (minimal continuous variables parametrizing the population activity manifold).
• The LLM notion is defined on weights — claim-li-2018-intrinsic-dimension-objective-landscape-codimension-parameter-space (codimension of a solution set in parameter space; the ancestor of Aghajanyan/LoRA). These are different mathematical objects on different spaces.
• ML already overloads "intrinsic dimension" onto two constructs — claim-ansuini-2019-two-intrinsic-dimensions-representation-vs-weight-space — and only the representation-space (activation) one is the same type of object as the neuroscience manifold; the famous fine-tuning/LoRA one is not.
• The only located formal bridge is ML-internal — claim-gelora-2024-representation-intrinsic-dimension-lower-bounds-lora-rank (representation ID lower-bounds LoRA rank) — and a targeted search found no paper bridging neuroscience manifold ID to LLM weight-space ID, recorded as [unverified — could not confirm/deny after search]: evidence of absence, not proof of non-equivalence.

Net: the definitional and mechanism evidence weighs toward "analogies that share a shape" over "one shared object," but as evidence of absence it cannot close the door on a reduction. Sub-question 1 is answered provisionally against a shared object; the question stays open for sub-question 2 and the empirical test, and because a bounded negative search is not a proof.

Progress — 2026-09-19 (partial; question stays open). Promotion of 10-inbox/raw/2026-09-15-is-qi-et-als-2023-finding-that-ten.md advances the candidate empirical within-/outside-manifold test named above (are LLM "within-manifold" edits easy while "outside-manifold" edits resist, mirroring Sadtler?). A targeted search found no primary source drawing that comparison for Qi et al.'s fine-tuning jailbreak — recorded as a bounded negative at observation-no-primary-source-links-qi-2023-jailbreak-to-sadtler-2014-within-manifold-asymmetry. The empirical arm is therefore unresolved by the literature, not answered.

What the search did surface — three primary papers on the geometry of safety-relevant fine-tuning — sharpens rather than settles the premise, and pulls in two directions:

• claim-ponkshe-2025-safety-subspaces-not-linearly-distinct-entangled-with-general-learning — safety is entangled with general learning, not a distinct direction (favors "same subspace").
• claim-springer-2026-finetuning-orthogonality-unstable-curvature-steers-into-alignment-subspace — the alignment-sensitive subspace is distinct; early updates avoid it and curvature steers them in (favors "outside-manifold, but reachable by a curved path" — a reframing of the analogy, not a clean within/outside split).
• claim-zhang-liu-shao-2023-intrinsic-finetuning-subspaces-task-specific-no-global-subspace — there is no demonstrated single global fine-tuning subspace at all; the premise's definite article ("the subspace") is not earned.

Net: sub-question 2 and a clean empirical parallel remain open; the literature offers partial, conflicting bearing evidence and no source that makes the Sadtler comparison. Stays open.