The grid and the invariant rays

A matrix is an action with coordinates attached

A two-by-two matrix can be read as four entries, but the entries are not the most durable fact. They describe what a linear rule does to two chosen basis vectors. Change the basis and the entries change, even when the physical transformation does not. The durable object is the rule that maps an input vector x to an output vector Ax.

Linearity means two constraints hold. Scaling the input scales the output by the same amount, and transforming a sum gives the sum of the transformed parts. If x = u + v, then Ax = Au + Av. This permits a difficult motion to be understood by decomposing the input into simpler directions, transforming those directions separately and recombining them.

Read the columns as controlled interventions

There is a practical way to recover the action hidden inside the table. Feed the transformation the first basis vector e₁ = (1, 0). The output is the first column. Feed it e₂ = (0, 1). The output is the second column. For the shear matrix A = [[1, 1], [0, 1]], the first probe returns (1, 0) and the second returns (1, 1). The machine leaves horizontal motion unchanged but adds one horizontal unit for every vertical unit.

Now take x = (3, -2). Linearity lets us predict its image without applying four memorised arithmetic rules: x = 3e₁ - 2e₂, so Ax = 3Ae₁ - 2Ae₂ = (1, -2). The columns are therefore not four unrelated constants. They are the outcomes of two basis interventions, sufficient to predict every other input. A different basis would ask different probe questions and produce different columns, while preserving the same underlying action.

The smallest useful definition

Take a non-zero vector v. Apply the transformation. If the result remains on the same line through the origin, then v is an eigenvector. The transformation may stretch it, shrink it or reverse it, but it does not send it into a new direction. The scalar multiplier is the eigenvalue λ.[1]

If λ = 2, one pass doubles distance along the ray. If λ = 0.5, it halves it. If λ = -1, the vector reverses direction while keeping length. If λ = 0, that direction collapses to the origin. These are not four unrelated cases. They are four forms of one question: what does the transformation do to a direction that it cannot bend away from itself?

How the invariant directions are found

The equation Av = λv can be rearranged as (A - λI)v = 0. A non-zero solution exists only when A - λI collapses at least one direction, which means its determinant is zero. This is why the characteristic polynomial appears. It is not a ritual for manufacturing special numbers. It asks which attempted scale factors make the transformed and scaled copies line up.

For A = [[3, 1], [0, 2]], the determinant of A - λI is (3 - λ)(2 - λ). The eigenvalues are 3 and 2. Their eigenvectors may be chosen as v₁ = (1, 0) and v₂ = (-1, 1). The ordinary vertical vector is not itself invariant, but it decomposes as (0, 1) = v₁ + v₂. After k steps it becomes 3^kv₁ + 2^kv₂ = (3^k - 2^k, 2^k). The eigenbasis reveals two independent rates inside a trajectory whose visible coordinates remain coupled.

flowchart LR
  A["Opening case"] --> B["The grid and the invariant rays"]
  B --> C["Repetition becomes selection"]
  C --> D["Modes in working systems"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#007D68,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Repetition becomes selection

Why powers of a matrix simplify in the right basis

Suppose a square matrix has enough independent eigenvectors to form a basis. Put those eigenvectors into the columns of a matrix S, and put their eigenvalues on the diagonal of Λ. Then A = SΛS^-1. The inverse S^-1 expresses the current state in modal coordinates, Λ scales each mode independently, and S returns to the original coordinates.[2]

Now apply A repeatedly. Adjacent factors S^-1S cancel, leaving a power only on the diagonal matrix. This is the algebraic form of the stretching-sheet story.

This equation explains persistence without anthropomorphic language. A mode with |λ| > 1 grows. One with |λ| < 1 decays. One with |λ| = 1 preserves magnitude, though it may reverse or rotate. If one magnitude is strictly larger than the rest and its initial coefficient is non-zero, its direction eventually dominates after normalisation. Numerical eigensolvers exploit the same filtering principle: repeated matrix polynomials amplify selected invariant components and attenuate others.[3]

A recurrence that discovers the golden ratio

The Fibonacci rule says the next number is the sum of the previous two. That verbal recurrence can be written as one matrix transformation:

The matrix has eigenvalues φ = (1 + √5)/2 and -1/φ. The first has magnitude about 1.618. The second has magnitude about 0.618. Unless the initial state is specially chosen to cancel the first mode, repeated updates suppress the second relative to the first. Consequently, the ratio Fₙ₊₁/Fₙ approaches φ.

When no real ray is invariant

A real matrix can have no real eigenvector in a two-dimensional plane. A pure rotation by an angle other than 0 or 180 degrees bends every real ray. Its eigenvalues and eigenvectors live in the complex numbers. This is not a mathematical trick added after the fact. A complex-conjugate pair is a compact way to encode two coupled real behaviours: radial scaling and angular rotation.

For a rotation-contraction matrix, the eigenvalue has the form re^iθ. The modulus r multiplies radius each step. The phase θ adds an angle each step. If r < 1, trajectories spiral inward. If r = 1, they orbit without radial decay. If r > 1, they spiral outward.

Modes in working systems

Probability mass that forgets its starting point

Consider a system that moves among three states. Its state vector does not record one definite location. It records a probability distribution. A column-stochastic transition matrix P moves probability mass from the current distribution to the next one. A stationary distribution π satisfies Pπ = π, so it is an eigenvector with eigenvalue 1.[4]

Take the synthetic transition matrix below. Its dominant eigenvalue is 1, and the remaining eigenvalues are approximately 0.521 and 0.379. The corresponding stationary distribution is approximately (0.336, 0.361, 0.303). Any initial difference along the smaller modes decays. The spectral gap therefore controls how quickly the system forgets where it began, subject to the chain’s structural conditions.

The same mathematics appears in historical link ranking. PageRank treated the web as a directed link graph and computed a recursively defined importance distribution over pages.[5] The point here is not that modern search ranking is reducible to the original algorithm. It is that a global pattern can be the eigenvector fixed by repeated local transfers.

Estimating modes from observations

Sometimes the matrix is known because an engineer designed it. Sometimes only a sequence of snapshots is available. Dynamic mode decomposition fits a linear evolution map between successive observations and decomposes that map into spatial modes and temporal rates. It has become a widely used way to extract dynamic patterns from time-series data, especially in fluid mechanics.[6]

The method’s usefulness can tempt an overreach. A fitted eigenmode may summarise regularity in observed data without being a unique physical cause. Hidden variables, forcing, noise and nonlinear state dependence can generate the same fitted mode. A mode becomes a mechanism claim only after interventions show that changing the relevant state component changes the trajectory as predicted.

Why eigenstructure enters machine intelligence

A recurrent system repeatedly updates a hidden state. Near a fixed point, a nonlinear update h_t+1 = f(h_t, u_t) can be approximated by its Jacobian matrix with respect to the state. The Jacobian’s eigenstructure then describes local directions of contraction, persistence, oscillation or expansion. This is one reason spectral design appears in recurrent neural networks and implicit models.

Unitary recurrent architectures were proposed partly to keep eigenvalue magnitudes at one and reduce vanishing or exploding long-range dynamics.[7] Deep equilibrium models instead solve for a fixed point of a repeated nonlinear layer, and later work regularised the fixed-point Jacobian to improve convergence stability.[8][9] These results show that spectral structure can affect optimisation and state dynamics. They do not show that eigenvalues alone determine learning, memory or intelligence.

The example exposes three distinct questions. The spectrum asks whether local perturbations eventually vanish. The eigenvectors ask which coupled patterns carry the two rates. Direct powers of J ask whether a transient crosses a real decision boundary. Convergence is a property of an infinite horizon; acceptability belongs to the horizon at which an action is released.

stateDiagram-v2
  [*] --> Claim
  state "A matrix is an action with<br/>coordinates attached" as Claim
  state "Read the columns as controlled<br/>interventions" as Test
  state "The smallest useful definition" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

When the spectrum lies by omission

Two stable spectra, two incompatible short-term stories

The clean modal picture assumes more than the existence of eigenvalues. It assumes the eigenvectors form a usable coordinate system. Symmetric and, more generally, normal matrices are especially friendly because they admit orthogonal eigenvectors. Modal components do not require large cancelling coefficients, and the matrix norm of each power is tightly governed by the eigenvalue magnitudes.

Non-normal matrices can behave differently. Their eigenvectors may be far from orthogonal. Representing an ordinary state in that basis can require large positive and negative modal coefficients that nearly cancel at the start. As the coefficients decay at different rates, the cancellation can temporarily disappear. The state grows before it eventually shrinks.

This is not a rare technicality confined to contrived two-by-two matrices. Nonnormality is central in fluid stability, control, numerical iteration and network dynamics. Trefethen showed that for matrices far from normal, eigenvalues may have little to do with finite-horizon quantities such as ‖Aⁿ‖, and that pseudospectra can reveal sensitivity and transient growth missed by the spectrum alone.[10] Earlier work on hydrodynamic stability demonstrated why spectrally stable linearised flows can still exhibit large transient amplification.[11]

A defective matrix makes the issue sharper. If there are fewer independent eigenvectors than dimensions, no eigenbasis exists. A Jordan block contributes polynomial factors such as kλ^k-1 in addition to the exponential λ^k. Even when |λ| < 1, the polynomial factor can delay decay. In numerical work, a Schur decomposition is often safer because it always exists over the complex numbers and uses an orthonormal basis, though its triangular form does not pretend the modes are independent.

The modal trust audit

Practitioners often jump directly from a plotted spectrum to a story about modes. The missing step is an assurance test for whether the modal coordinates deserve explanatory weight. The following audit is a practitioner synthesis grounded in matrix perturbation and numerical linear algebra.[12][13][14]

The audit does not produce a consciousness score, intelligence score or universal stability certificate. It answers a narrower question: at the declared horizon and perturbation scale, may this eigenstructure be used as the primary explanation of repeated behaviour?

Seven gates before a modal claim

The first gate is operator scope. Specify what state is transformed, whether the same operator is repeated and where any inputs or forcing enter. The second is invariance. For every reported pair, compute a residual such as ‖Av - λv‖/(‖A‖‖v‖). A visually persuasive arrow is not an eigenvector if this residual is large.

The third gate is completeness. Check whether the proposed eigenvectors span the analysed state space. If they do not, use invariant subspaces, Jordan structure or an orthonormal Schur basis rather than pretending each coordinate evolves independently. The fourth is conditioning. The condition number of the eigenvector matrix S estimates how strongly small state or operator errors can distort modal coefficients. There is no universal safe threshold, so report the value alongside the numerical precision and perturbation scale that matter in the application.

The fifth gate is finite-horizon gain. Calculate or bound max₀≤k≤H ‖A^k‖ for the actual decision horizon H. If inputs persist, inspect forced response as well, because a stable unforced system can amplify repeated disturbances. The sixth is perturbation stability. Recompute the spectrum, subspaces and predicted rates after small structured changes to the operator. A mode that changes identity under harmless perturbations is a fragile explanatory object.

The final gate is intervention. Add, remove or vary the state component associated with the proposed mode and test whether the subsequent trajectory changes with the predicted rate and pattern. This step separates a descriptive decomposition from a causal account. Passing all seven gates permits a bounded claim: eigenstructure is the primary explanation for this operator, state region, perturbation scale and horizon. Failing a gate does not invalidate the matrix. It changes the representation and evidence needed next.

Executable lab: make the hidden assumptions visible

The browser lab below accepts any real two-by-two matrix, an initial vector and an iteration horizon. It draws the trajectory and reports eigenvalues, spectral radius, non-normality, eigenvector conditioning when a real basis exists, and the largest finite-horizon gain max_k≤H‖A^k‖₂.

The lab encodes three assumptions. The same linear operator is applied at every step. The state has exactly two real coordinates. Floating-point arithmetic is adequate for the selected values. A positive result means the trajectory, modal rates and finite-horizon gain agree well enough to justify an eigenmode account at this scale. A negative result does not prove chaos or nonlinearity. It means the spectrum alone is not a sufficient explanation and should be supplemented by direct simulation, Schur structure, singular-value growth or perturbation analysis.

from __future__ import annotations

from dataclasses import dataclass
import numpy as np


@dataclass(frozen=True)
class ModalAudit:
    eigenvalues: np.ndarray
    spectral_radius: float
    eigenbasis_condition: float | None
    nonnormality: float
    peak_gain: float
    peak_step: int
    final_state: np.ndarray


def spectral_norm(matrix: np.ndarray) -> float:
    return float(np.linalg.svd(matrix, compute_uv=False)[0])


def audit_matrix(
    matrix: np.ndarray,
    initial_state: np.ndarray,
    horizon: int = 20,
) -> ModalAudit:
    matrix = np.asarray(matrix, dtype=float)
    state = np.asarray(initial_state, dtype=float)

    if matrix.shape != (2, 2):
        raise ValueError("This minimal lab expects a 2x2 matrix.")
    if state.shape != (2,):
        raise ValueError("The initial state must have two coordinates.")
    if horizon < 1:
        raise ValueError("The horizon must be positive.")

    eigenvalues, eigenvectors = np.linalg.eig(matrix)
    spectral_radius = float(np.max(np.abs(eigenvalues)))

    rank = np.linalg.matrix_rank(eigenvectors)
    condition = None if rank < 2 else float(np.linalg.cond(eigenvectors))

    commutator = matrix.T @ matrix - matrix @ matrix.T
    denominator = max(np.linalg.norm(matrix, "fro") ** 2, 1e-15)
    nonnormality = float(np.linalg.norm(commutator, "fro") / denominator)

    power = np.eye(2)
    gains: list[float] = []
    for _ in range(horizon + 1):
        gains.append(spectral_norm(power))
        power = matrix @ power

    peak_step = int(np.argmax(gains))
    final_state = np.linalg.matrix_power(matrix, horizon) @ state

    return ModalAudit(
        eigenvalues=eigenvalues,
        spectral_radius=spectral_radius,
        eigenbasis_condition=condition,
        nonnormality=nonnormality,
        peak_gain=float(gains[peak_step]),
        peak_step=peak_step,
        final_state=final_state,
    )


if __name__ == "__main__":
    normal = np.diag([0.8, 0.7])
    nonnormal = np.array([[0.8, 4.0], [0.0, 0.7]])

    for name, matrix in [("normal", normal), ("nonnormal", nonnormal)]:
        result = audit_matrix(matrix, np.array([0.0, 1.0]), horizon=15)
        print(name, result)

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: Opening case
  M->>E: A matrix is an action with coordinates attached
  E-->>M: compare rival explanations
  E->>D: Trust a stable mode only when it stays stable
  D-->>Q: revise scope or act

Trust a stable mode only when it stays stable

When confronted with a matrix, many people ask for its eigenvalues first. That is often the wrong first decision. Ask instead what is being repeated, over which states, for how long and under what disturbances. Only then does the spectrum acquire an operational meaning.

If the operator is fixed, diagonalizable, well-conditioned and tested at the relevant horizon, eigenvectors provide a powerful explanatory basis. They reveal invariant patterns, eigenvalues reveal rates, and matrix powers reveal selection. In that regime, replacing direct step-by-step calculation with modal reasoning is more than elegant. It exposes which parts of an initial condition persist and why.

If the matrix is non-normal, defective, highly sensitive or only a local Jacobian, the decision changes. Do not discard eigenvalues, but narrow their claim. Use them for asymptotic rates or local hypotheses. Pair them with direct finite-horizon simulation, eigenvector conditioning, singular-value gain, Schur form and perturbation tests. In a learned system, check whether the same mode survives across seeds, states and small weight changes before naming it as a mechanism.

A useful triage has three questions. Can the current state be decomposed into stable invariant components without extreme cancellation? Do those components predict the largest excursion before the decision deadline? Does the interpretation survive a perturbation comparable to measurement, estimation or training variation? A “yes” to only the first question supports algebraic description, not operational assurance. A “yes” to all three supports a modal decision rule.

The deepest shift is from treating eigenvectors as special arrows to treating them as a test of decomposability. They ask whether repeated behaviour can be separated into patterns that carry their own rates. When the answer is yes, a complicated transformation becomes a set of independent scalar stories. When the answer is no, the failure itself is informative: interaction among directions is part of the mechanism and must not be erased by a tidy spectrum.

That is the practical consequence of eigenstructure. It tells you not only what may remain stable under repetition, but also when the very idea of a stable mode deserves trust.