One Rule, No Randomness, and Nothing You Can Forecast

The map x to 4x(1-x) contains no randomness and is still useless for prediction, because substituting x = sin squared of pi t turns it into angle doubling: one binary digit of your measurement is spent per step, so fifty steps eat fifteen decimal digits. The resulting series has autocorrelation exactly zero at every lag, proved by orthogonality of distinct cosine frequencies rather than measured. What that does not establish is anything about real return series, and the article says so.

Take the interval from zero to one and the rule $x \mapsto 4x(1-x)$ . There is no randomness in it anywhere. Hand me a starting value and I can compute every future value exactly, forever, with nothing but a pencil. Now hand me a starting value measured to fifteen decimal places and ask where the orbit sits after fifty steps. I cannot tell you.

That gap between a rule you fully know and a future you cannot state is the whole content of the word chaos, and it can be quantified precisely for this map. The answer is one binary digit per step.

Determinism is a property of the rule, not of you

A system is deterministic when the next state is a function of the current state. That is a statement about the rule. Predictability is a statement about the map from what you know about the current state to what you can say about a later one, and those two things come apart the instant the rule amplifies the difference between two nearby states.

For this map, differences do not merely grow. They double, every step, exactly.

The map is angle doubling in costume

Substitute $x = \sin^{2}(\pi t)$ and watch what happens:

4\sin^{2}(\pi t)\bigl(1 - \sin^{2}(\pi t)\bigr) \;=\; 4\sin^{2}(\pi t)\cos^{2}(\pi t) \;=\; \sin^{2}(2\pi t)

(1)

The messy quadratic on $x$ is the doubling map on $t$ , seen through a change of coordinates. Iterating is now free, because doubling an angle $n$ times multiplies it by $2^{n}$ :

x_{n} \;=\; \sin^{2}\!\bigl(2^{n}\pi t_{0}\bigr), \qquad t_{n+1} = 2t_{n} \bmod 1

(2)

Conjugacy

Two maps $f$ and $g$ are conjugate when some invertible $h$ satisfies $h \circ g = f \circ h$ , so that $h$ carries orbits of $g$ onto orbits of $f$ . Here $h(t) = \sin^{2}(\pi t)$ carries the doubling map onto $4x(1-x)$ . Everything dynamical about one map is therefore a fact about the other, and the doubling map is the one you can read off by eye.

One binary digit per step

Write $t_{0}$ in binary. Doubling shifts the binary point one place to the right, and reducing modulo one throws away the integer part, so a step of the dynamics deletes the leading bit and promotes the next one into the position that matters. Fifty steps in, the value you are trying to predict is governed by the fifty-first bit of your measurement.

Fig. 1 — A step of the dynamics is a shift. The digits you measured are a fuel tank, and the map burns one per step from the front.

The same fact in the language of calculus: the doubling map has derivative two everywhere, so a perturbation $\varepsilon$ becomes $2^{n}\varepsilon$ . The growth rate of the logarithm of the separation is the Lyapunov exponent, and here it is exactly $\log 2$ . Measured along a 300,000-step orbit built in exact arithmetic, it comes out at 0.69315 against $\log 2 = 0.69315$ .

Now the budget. Fifteen decimal digits is worth

\left\lfloor 15 \log_{2} 10 \right\rfloor = 49 \quad \text{doublings}, \qquad 2^{49} < 10^{15} < 2^{50}

(3)

so the fiftieth step is the one that spends the last of them. Two starting values a trillionth apart stay within a billionth of each other for ten steps, which looks reassuring, and are 0.66 apart by step forty, which is the width of the whole interval. Nothing went wrong in between. The gap simply doubled thirty more times.

Fig. 2 — A straight line on a logarithmic scale is a constant doubling rate. The flattening at the top is not the dynamics calming down, it is the interval running out of room.

The horizon this implies is the depressing part. An initial uncertainty $\varepsilon$ reaches order one after about

n^{\star} \;\approx\; \frac{1}{\lambda}\log\frac{1}{\varepsilon} \;=\; \log_{2}\frac{1}{\varepsilon} \quad \text{steps}

(4)

which at $\varepsilon = 10^{-12}$ is a shade under forty, matching the measurement above. Note how the horizon depends on the precision: logarithmically. Halving your measurement error buys exactly one extra step. Ten more decimal digits buy thirty-three.

The autocorrelation is exactly zero

Line the values up as a time series and they pass every linear test for randomness, which is the property that makes this map an uncomfortable thing to think about. The reason is visible in equation (2). Rewrite it with a double angle:

x_{n} \;=\; \frac{1 - \cos\bigl(2\pi \cdot 2^{n} t_{0}\bigr)}{2}

(5)

so the covariance between $x_{i}$ and $x_{j}$ is a quarter of the average of $\cos(2\pi 2^{i} t)\cos(2\pi 2^{j} t)$ over the circle. Distinct integer frequencies are orthogonal there:

\int_{0}^{1} \cos(2\pi m t)\,\cos(2\pi n t)\,\mathrm{d}t \;=\; 0 \qquad (m \ne n)

(6)

and $2^{i} \ne 2^{j}$ whenever $i \ne j$ . So the autocorrelation is not small, and it is not approximately zero. It is zero, at every non-zero lag, as an identity. The orbit also has a well-defined long-run mean, $\int_{0}^{1}\sin^{2}(\pi t)\,dt = 1/2$ , so there is a distribution for the correlation to be measured against. Empirically, over 200,000 exact samples, the worst absolute autocorrelation across lags one to ten is 0.00307, and the integral behind it evaluates to $7 \times 10^{-17}$ .

What that does not say is that the values are independent. They are the opposite of independent: each one determines all the rest. Square them and the correlation reappears at once. Zero linear correlation is a far weaker property than independence, and treating the two as the same thing is the same error as treating determinism as predictability, run in the opposite direction.

The computer runs out of digits at step fifty-two

There is a trap in verifying any of this. The obvious way to produce a reference orbit is to iterate $4x(1-x)$ in double precision. That orbit is worthless past about fifty steps, for exactly the reason under discussion: double precision carries roughly fifteen or sixteen decimal digits, and the map spends one bit per step. Using it as the reference would be circular, a simulation that assumes the conclusion in order to display it.

The honest construction is equation (2) with $t_{0} = p/q$ for an odd prime $q$ , because then the angle orbit is integer arithmetic modulo $q$ and never drifts at all. Checked against that reference, the conjugacy of equation (1) holds to $6 \times 10^{-16}$ across 1,999 grid points, and the floating-point orbit tracks the exact one and then parts company from it at step 52. That divergence is not a bug in the arithmetic. It is the claim of the article, applied to the machine making the claim.

What this does and does not say about prices

Here is the whole of what has been established. A rule with no randomness in it can produce a series whose linear statistics are indistinguishable from noise, and the number of steps you can forecast is set by your measurement precision divided by the Lyapunov exponent. If a price series were generated by a map like this one, forecasting fifty steps ahead would require fifteen decimal digits of the current state that you do not have, and there is no meaningful sense in which a market has a state measurable to fifteen digits.

Here is what has not been established, and this article will not pretend otherwise. Nothing above says whether real return series are chaotic. Nothing above evaluates any particular statistical test for chaos in returns, and no such claim belongs in a piece whose evidence is a quadratic map on the unit interval. The mechanism is the result. Whether it applies to any given market is a question about data, and it deserves its own evidence rather than a borrowed metaphor.

The one transferable lesson is equation (4). When a system amplifies errors at a constant rate, forecast horizon grows like the logarithm of precision, which is the worst possible return on effort. Doubling the resolution of your instruments buys a single extra step, and it will keep buying a single extra step no matter how many times you do it.

Sources and further reading

The general setting Chaos theory
The map itself, at the parameter used here Logistic map
What equation (2) turns it into Dyadic transformation
The rate in equation (4) Lyapunov exponent
The statistic that equation (6) kills Autocorrelation

Both identities here were verified symbolically rather than sampled, the orthogonality at two separate lags. The orbit statistics come from exact integer arithmetic on the angle, which is the only reference this particular claim can honestly be checked against.