Time as a Phase-Space Trajectory

U(1) symmetry — time as rotation

The Lie group U(1) is the unit circle in the complex plane:

U(1) = { e^(iθ) | θ ∈ [0, 2π) }

For each time scale i with cycle length T_i, the Temporal Spiral assigns a phase angle:

θ_i(t) = 2π · frac(t / T_i)   ∈ [0, 2π)

This is a point on U(1), rotating at constant angular velocity ω_i = 2π / T_i. Advancing time continuously rotates the point around the circle. Two nearby timestamps map to nearby points on U(1) — the angular distance between them is proportional to their temporal separation modulo T_i, giving a well-defined, continuous metric on the space of moments.

The API exposes this as a (sin θ_i, cos θ_i) pair rather than the complex exponential directly, keeping the representation real-valued and immediately usable as model input without further processing.

Structural parallel with the Schrödinger operator

In quantum mechanics, the time-evolution operator for a time-independent Hamiltonian H is:

U(t) = e^(−iHt/ℏ)

For a Hamiltonian with a single eigenvalue E acting on an eigenstate |ψ⟩:

U(t)|ψ⟩ = e^(−iEt/ℏ)|ψ⟩

This is a U(1) rotation at angular frequency ω = E/ℏ. The Spiral's phase coordinate for scale i is structurally identical:

θ_i(t) = ω_i · t   where ω_i = 2π / T_i

The correspondence is exact up to the domain: quantum time evolution rotates a complex amplitude at a frequency set by an energy eigenvalue; the Spiral rotates a classical phase coordinate at a frequency set by a human-relevant cycle length. Both structures live on U(1). The Spiral replaces the single quantum frequency E/ℏ with 12 classical frequencies, one per time scale, giving the encoding coverage across twelve orders of magnitude in period length simultaneously.

The 12-torus T¹²

Taking all twelve phase angles together, the full Spiral encoding of a timestamp t is a point on the product space:

T¹² = S¹ × S¹ × … × S¹   (12 factors)
     = U(1)¹²

state(t) = (θ₁(t), θ₂(t), …, θ₁₂(t)) ∈ T¹²

The trajectory traced by state(t) as t advances is a Kronecker flow on the torus: a straight line in the angle coordinates, with slope vector (ω₁, ω₂, …, ω₁₂). When the angular frequencies are rationally independent — which holds when the cycle lengths share no common rational ratio — the flow is dense: it eventually passes arbitrarily close to every point on T¹². This density property means the encoding does not collapse distinct moments to the same point even at very long time scales.

Smoothness. Because each θ_i is a Lipschitz function of t with Lipschitz constant ω_i, small changes in t produce bounded changes in the torus point. Neighboring timestamps are always nearby on T¹² — making the encoding amenable to gradient-based learning without special handling of discontinuities.

Discriminability. Timestamps separated by a fraction of any cycle T_i produce visibly distinct coordinates on that factor of the torus. A model that receives the full 24-dimensional vector (sin θ₁, cos θ₁, …, sin θ₁₂, cos θ₁₂) can distinguish moments differing by as little as one millisecond (θ_us) while also tracking millennial positioning (θ_mil) — all within the same fixed-dimension input.

Hamiltonian mechanics framing

In the language of classical mechanics, the θ_i are cyclic generalized coordinates — coordinates that do not appear explicitly in the Lagrangian, only through their time derivatives. For a system of 12 independent rotors, the Lagrangian is:

L = ½ Σ_i I_i · θ̇_i²

The conjugate momenta are:

p_i = ∂L/∂θ̇_i = I_i · θ̇_i = I_i · ω_i   (constants of motion)

Since each θ_i is cyclic, its conjugate momentum p_i is conserved. The Hamiltonian reduces to:

H = Σ_i p_i² / (2 I_i)   (constant)

This is a trivially integrable system: it has 12 independent constants of motion (the momenta), satisfying the Liouville–Arnold theorem. The Spiral trajectory is therefore a Hamiltonian flow on T¹² — a straight-line trajectory in action-angle coordinates with angular velocities fixed by the cycle lengths.

There is no chaos, no sensitivity to initial conditions, and no numerical drift — the trajectory is determined analytically for all t by the closed-form formula θ_i(t) = ω_i t mod 2π.

What this means for AI

The torus T¹² carries a flat Riemannian metric — the product of the standard metrics on each S¹. Flatness means there is no curvature to distort gradients: the loss surface seen by a model trained on sinusoidal time coordinates is locally Euclidean in every direction.

Practically, this means:

• Well-conditioned gradients. Gradient descent on the torus coordinates does not suffer from the ill-conditioning that plagues raw integer timestamps (where the scale of the input can differ by 10¹³ across the supported range).
• No discontinuity at period boundaries. A model seeing sin(θ_hr) and cos(θ_hr) does not observe a jump at midnight — the encoding wraps continuously, just as the U(1) rotation does.
• Shift equivariance. A time shift of δt rotates every θ_i by ω_i δt. A model that learns a pattern at one time of day generalises to the same phase on any other day, because the torus point at the same phase is always the same embedding.
• Scale separation. The 12 frequencies span 13 orders of magnitude, so a single fixed-length input vector simultaneously encodes microsecond cadence and millennial positioning — orthogonal information in the same embedding.