A trace of the intellectual path that produced this repo.
The rotation matrix for 90°:
M = [[0, -1],
[ 1, 0]]
has no real eigenvectors. Every vector in the plane gets knocked off its own line, so there's nothing for the transformation to "stretch." The characteristic polynomial yields λ = ±i — imaginary roots.
The standard reaction is to treat this as a failure. But it isn't. Those eigenvalues are exactly the complex number that performs the same rotation. The matrix and the scalar are doing the same job in two different languages, and the eigenvalue computation is the translator.
For the general rotation-by-θ matrix:
R(θ) = [[cos θ, -sin θ],
[sin θ, cos θ]]
The eigenvalues are cos θ ± i sin θ = e^(±iθ) — exactly the complex number you'd multiply by to rotate a point through angle θ.
The observation: rotation looks complicated in ℝ², but becomes a single scalar multiplication in ℂ. The eigenvalues aren't just numbers attached to the matrix — they're the matrix's true identity, expressed in the number system rich enough to hold it.
This suggests a general principle:
- Systems look complex in one representation
- They become simple in a richer one
- Eigenvalues are the "compression layer" — the invariant form of a transformation
The same pattern appears in Fourier transforms (diagonalizing convolution), quantum mechanics (observables as operators), and PCA (rotating data into independent axes).
Reframe: systems don't look complex because they are complex. They look complex because we're expressing them in the wrong basis.
The unlearnable interference research showed something unexpected:
- Increasing model capacity → marginal improvement
- Increasing discontinuity rate → strong degradation
- Even large MLPs with universal approximation properties fail
The naive explanation is "the signal is too complex." The eigenvalue framing suggests something sharper:
The signal has no eigenbasis that remains stable long enough to be discovered under the learner's representation constraints.
Just as a rotation matrix has no eigenvectors in ℝ² (you must move to ℂ), some signals have no stable structure in the space the learner operates in. But the interference system goes further — it actively destroys the possibility of finding a stable basis.
| Linear Algebra | Unlearnable Interference |
|---|---|
| Matrix | Signal dynamics |
| Eigenvectors | Stable invariant modes |
| Eigenvalues | True identity of the system |
| Diagonalization | Finding a learnable representation |
| Complex plane | Extended representation space |
The key insight: drift and noise are not the same thing. Noise is random. Drift is structure with direction. Some "noise" directions are actually natural modes of the system — just like rotation has no real eigenvectors but has perfectly stable complex ones.
The initial intuition: maybe the system is unlearnable because we didn't model rotations.
This is partially correct. The signal does contain rotational structure at any short time window — it's locally sinusoidal with phase. But the rotations are:
- Hidden (private-key basis transitions)
- Switching (orthogonal jumps between frames)
- Non-stationary (logistic-map chaos driving phase)
The correct framing: the system is not missing rotations. It is preventing rotations from stabilizing long enough to be modeled.
Formally, if the signal is:
x(t) = sin(2π f_k t + θ_k(t)), k changes unpredictably
then there is no single frequency basis, no stable phase trajectory, no consistent coordinate system. Even Volterra filters, KLMS, and large MLPs fail — because they all assume there exists a consistent mapping f(x). Your system enforces that f itself keeps changing.
Combining the above:
A quantum (or adaptive) system's "true computational identity" is not its state, but its eigen-structure under drift.
More concretely for the interference system:
Unlearnability is not absence of structure — it is instability of the structure's coordinate system.
This splits into two regimes:
- Transient invariants: Structure exists locally but dies before learning can lock onto it
- Structural annihilation: Structure ceases to exist altogether (spectral gap → 1)
Empirical results show the interference system operates primarily in the transient-invariant regime (spectral gap > 3), not annihilation. Structure is present — it just doesn't persist.
The core competition:
τ_structure = lifetime of dominant invariant structure
τ_alignment = minimum time for learner to stabilize gradient direction (∆W)
Empirically, a system is learnable only when τ_structure ≥ τ_alignment.
The formal biconditional — learnability if and only if the inequality holds — is the natural hypothesis, but what has been demonstrated is strong empirical sufficiency: every tested regime where τ_structure falls below τ_alignment produces convergence failure. The necessity direction (does τ_structure ≥ τ_alignment always suffice?) remains to be tested across other signal families and architectures.
τ_alignment approximates the minimum temporal horizon required for the learner's gradient updates to establish a consistent descent direction. It is not simply a function of learning rate — experiments varying LR across 0.0001–0.001 showed minimal change in τ_alignment, suggesting it is primarily an architectural property (memory length, depth) rather than a hyperparameter.
The ChaoticAnchor layers interact as follows:
- L1 (Chaos): Slow phase drift — trackable in isolation
- L2 (Adaptive chaos): Phase drift rate changes — harder but still learnable in isolation
- L3 (Private key): Basis switch at hidden intervals — destroys accumulated alignment
- L4 (Orthogonal jumps): Instantaneous frame change — resets the coordinate system entirely
The anchor controls τ_structure through:
tau_samples→ L1/L2 coherencekey_transition_interval→ L3 reset frequencyorthogonal_switch_prob→ L4 stochastic breaks
τ is a programmable parameter. The system doesn't fail randomly — it is designed to enforce dB/dt > τ_alignment.
From BoundaryProbe.py with the corrected EigenDriftTracker:
| Anchor speed | τ_structure @0.5 | Spectral gap | Late SINR |
|---|---|---|---|
| slow | 411 samples | ~4.0 | -0.32 dB |
| normal | 53 samples | ~3.3 | -0.71 dB |
| fast | 18 samples | ~3.4 | -1.30 dB |
Key observations:
- Structure exists in all cases (gap > 3 in all runs — not annihilation)
- τ_structure drops monotonically and controllably
- SINR tracks τ_structure, not model capacity
- τ_learning remains near zero at 0.7 threshold in all cases — the learner never achieves stable directional alignment
The LR sweep (varying learning rate, fixed anchor = normal) showed no meaningful recovery, confirming that τ_alignment is an architectural property, not just a hyperparameter.
Early formulation (too simple):
τ_structure < τ_learning → failure
Correct version (grounded in results):
Learnability emerges only when the persistence of invariant structure exceeds the time required for directional alignment of the learning updates.
Or shorter:
Learning is bounded by the lifetime of exploitable structure.
This framework reframes several standard problems:
- Catastrophic forgetting: not an overwriting problem, but a basis-instability problem where τ_structure of old tasks falls below τ_alignment
- Non-stationary learning: the failure mode is not distribution shift per se, but the rate at which invariant structure changes relative to the learner's adaptation horizon
- Adversarial robustness: the most effective adversaries don't increase complexity — they decrease τ_structure
These implications are hypothesized extensions. The principle is currently validated on one signal family (chaotic anchor with logistic-map phase modulation) and one learner class (online MLP with backprop). Whether τ_structure ≥ τ_alignment governs learnability in other non-stationary systems — recurrent learners, transformers, different signal families — is an open question, not a confirmed result.
The key upgrade to standard practice:
Instead of asking "is this signal learnable?", ask "what is τ_structure, and does it exceed τ_alignment for our learner?" This reframes learnability from a binary property of the signal to a relative property between signal coherence and learner dynamics.