Eigen-Direction Review¶
Analytical classification — instantiates Mixed-Stability Saddle Navigation
Reads the local model's structure to classify which directions decay, which amplify, and how they couple — turning a saddle into a labeled set of stable and unstable modes.
Before anyone touches the system, someone has to say which way is dangerous. Eigen-Direction Review answers that from the model alone: it linearizes the dynamics at the suspected equilibrium and reads the local response off the resulting spectrum — negative-response (decaying) directions are stable, positive-response (amplifying) directions are unstable, and directions whose response sits near zero are neutral and must be treated as unknown. Its defining move is that it never perturbs the live system; the classification comes entirely from the model's structure — the Jacobian's eigenvalues and eigenvectors, or their sensitivity-analysis equivalent. That is what separates it from every sibling that learns by acting: it produces the direction map on paper, cheaply and repeatably, so that all downstream action starts from a labeled geometry instead of a scalar "stuck / unstable" verdict.
Example¶
A team's neural-network training run has flattened onto a long plateau — loss barely moving, but not obviously diverging. Rather than just lowering the learning rate and hoping, they run an Eigen-Direction Review of the loss surface at the current point: using an iterative estimator (power/Lanczos-style curvature probes on the model, not on new data), they recover the extreme curvature directions. The picture that comes back is a saddle, not a minimum: a couple of strongly negative-curvature directions along which loss can still fall sharply (the escape modes), a broad cluster of near-zero-curvature directions that are flat and undecided, and the rest well-curved and stable. Reading the leading escape eigenvector, they notice it is nearly aligned across a normalization-scale parameter and a downstream weight — the two move together, so nudging one will drag the other. Illustratively: "two escape directions, ~40 flat modes, and one coupled pair to watch." None of that required perturbing the run; it told them the plateau is a pass, where the escape lies, and that a naive single-parameter step would spill sideways.
How it works¶
- Linearize at the point. Take the local model at the suspected equilibrium and form its response operator (Jacobian, or curvature/Hessian for a gradient system).
- Sort the spectrum. Classify each mode by its eigenvalue: decaying → stable, amplifying → unstable, near-zero → neutral/unknown. Keep the growth/decay rates, not just the signs.
- Read the eigenvectors. Translate each mode back into the real variables it moves, so "unstable direction" becomes a named combination of controllable quantities.
- Inspect the coupling. Where eigenvectors are non-orthogonal (a non-normal operator) or off-diagonal terms are large, flag the pairs of modes that move together — a control aimed at one will bleed into the other.
Tuning parameters¶
- Linearization radius — how far from the point you trust the linear picture. Wider coverage buys reach but nonlinearity eventually invalidates the classification.
- Neutral-band width — the eigenvalue magnitude below which a mode is called "unknown" rather than stable. A wider band is more cautious and flags more directions as undecided.
- Spectrum resolution — full spectrum versus just the extreme eigenpairs. Estimating only the ends is cheap but can miss a mid-band mode that matters.
- Model source — analytic derivatives versus finite-difference or sampled sensitivities; higher fidelity for higher cost.
- Refresh cadence — one-shot versus recomputed as the operating point moves, since the spectrum drifts along a trajectory.
When it helps, and when it misleads¶
Its strength is converting a scalar impression into a labeled, rate-annotated geometry without disturbing anything live — the cheapest possible way to learn a saddle's shape, and the natural first step before any probing or crossing. Its central failure mode is that the linear picture is only locally valid: the Hartman–Grobman guarantee[n1] holds near a hyperbolic point but lapses exactly at the non-hyperbolic (near-zero-eigenvalue) directions — which are frequently the very directions that decide the outcome, since the slow center-manifold behavior is invisible to the eigenvalues. The classic misuse is trusting the classification globally, far outside the radius where the linearization holds, or on a model that is a poor proxy for the real system. The guarding discipline is to bound every claim to the validity radius, label near-zero modes as unknown rather than benign, and hand the risky directions off for a live probe rather than pretending the spectrum settled them.
How it implements the components¶
Eigen-Direction Review fills only the analysis-side geometry of the archetype:
stability_direction_map— its primary output: the sorted list of stable, unstable, and neutral modes with their growth/decay rates, expressed in real variables.cross_mode_coupling_watch— the eigenvector geometry (non-orthogonality / off-diagonal structure) exposes which modes move together, marking where single-axis controls will spill.
It does not implement reversible_perturbation_budget or directional_control_policy — injecting a small, recoverable perturbation into the live system to read its actual response is the Reversible Nudge Test; this review only classifies the model. Nor does it lay the result out as a shared artifact — that is the Saddle Neighborhood Map.
Related¶
- Instantiates: Mixed-Stability Saddle Navigation — supplies the direction classification the rest of the pattern acts on.
- Sibling mechanisms: Saddle Neighborhood Map · Reversible Nudge Test · Separatrix Crossing Checklist · Unstable Mode Dashboard · Basin Arrival Review
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Eigen-Direction Review operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it reads the local model's structure to classify which directions decay, which amplify, and how they couple — turning a saddle into a labeled set of stable and unstable modes.
Independent corroboration: The frozen evidence defines Eigen-Direction Review as 'Reads the local model's structure to classify which directions decay, which amplify, and how they couple — turning a saddle into a labeled set of stable and unstable modes', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Mathematics
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Linear algebra and dynamical-systems analysis cohered eigenvectors as invariant directions whose eigenvalues classify local decay, amplification, and saddle structure.
Related originating lineages:
- Systems Thinking & Cybernetics — Stability and modal analysis made eigen-direction interpretation an operational diagnostic for coupled systems.
Review resolution: The current reviewers agree that mathematics is primary. For the reported differences (origin_mode_disagreement, encyclopedia_synthesis_disagreement), the evidence supports single_lineage, multi_domain, and systems_cybernetics; these choices preserve materially formative origins without conflating later domain reach.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The review is offline and effectively one-shot: its map is accurate at the point where it was computed and goes stale as the operating point moves. It is the which directions half of direction-awareness; the Unstable Mode Dashboard is the how fast, right now half. Neither substitutes for the other.
[n1] Hartman–Grobman theorem — near a hyperbolic fixed point (no eigenvalues on the imaginary axis), a nonlinear system's local behavior is topologically equivalent to that of its linearization. This is what licenses reading stability from eigenvalues, and precisely why the guarantee fails at the non-hyperbolic, near-zero-eigenvalue directions. ↩