The signal is deterministic. The statistics say it’s noise. The message is in the branches.

The Paradox

Take a point c in the complex plane. Iterate f_c(z) = z² + c backward, choosing a preimage branch at each step. The orbit converges to the Julia set J(f_c) and, by the Almost-Sure Invariance Principle, the Birkhoff sums of any Hölder observable are pathwise coupled to Brownian motion:

S_n = W(σ²n) + o(n^ε)

To any linear statistical test — autocorrelation, spectral density, distribution fitting — the orbit looks like noise. But the branch sequence is under your control. Each branch choice is a bit. The orbit carries a message.

Steganography in Julia Orbits

Classical steganography hides messages in redundant data — LSBs of images, phase of audio samples. The cover medium is stochastic, and the message perturbs its statistics. A sufficiently powerful steganalyst can detect the perturbation.

Liminal channels are different. The cover medium is a deterministic dynamical system whose statistics are provably identical to noise. The message isn’t hidden in the noise — it is the noise. The branch sequence that generates the orbit simultaneously encodes information and produces the correct invariant measure. No perturbation. No statistical footprint.

The key is the parameter c. Without it, the interceptor sees a time series indistinguishable from Brownian motion. With it, they can invert the orbit, recover the branches, and read the message.

Interactive Explorer

Explore the liminal channel across four views: the orbit itself, the statistical paradox, the key-reveal mechanism, and the degree-2 vs degree-6 comparison that eliminates ordinal pattern fingerprints.

The Degree Problem

At d = 2, backward iteration z → ±√(z − c) offers only two preimage branches. The binary choice creates forbidden ordinal patterns — certain consecutive orderings that can never appear in the time series. A steganalyst testing for these forbidden patterns can detect the channel.

At d = 6, z⁶ + c has six preimages: three with Re > 0, three with Re < 0. The sign channel (1 bit per sample) encodes the message. The cover channel picks among three branches within each half-plane, and a per-step permutation shuffles the assignment. This eliminates all forbidden patterns. The signal becomes statistically indistinguishable from i.i.d. white noise.

Two-Key Architecture

The degree-6 structure enables a natural separation of concerns:

• Key A (sign channel) — shared with regulators or verifiers for detection. Proves a watermark is present without revealing content.
• Key B (cover channel + permutation) — held by the provider for provenance. Encodes identity, timestamp, or chain-of-custody metadata.

At d = 2, these roles are architecturally inseparable — the single branch choice must serve both detection and provenance. At d = 6, the 3+3 branch structure provides the extra degrees of freedom to split them cleanly.

Capacity

• d = 2: 1 bit per sample. At 16 kHz sampling: 16,000 bits/sec.
• d = 6: 1 bit (sign) + log₂(3) ≈ 1.58 bits (cover) = 2.58 bits per sample. At 16 kHz: 41,280 bits/sec hidden in provably Brownian noise.

The key is a point in ℂ. The message is in the branches. The noise is the channel.