The number line. The triangle inequality. Why errors accumulate.
Begin with counting: one, two, three — the natural numbers $\mathbb{N}$. Add zero and negatives for the integers $\mathbb{Z}$. Insert fractions — the rational numbers $\mathbb{Q}$, dense but with gaps. Fill all gaps to obtain the real numbers $\mathbb{R}$ — the continuous line.
The real numbers satisfy the Archimedean property: for any positive $x, y$, no matter how small $x$ and how large $y$, you can add $x$ enough times to exceed $y$. There are no infinitesimals. This property seems obvious — but it is precisely what causes errors to accumulate.
On $\mathbb{R}$, the distance between two points is simply:
$$d_{\mathbb{R}}(x, y) = |x - y|$$This satisfies all three metric axioms — non-negativity, symmetry, and the triangle inequality.
$d(3, 7) = 4 \qquad d(-2.5, 1.5) = 4 \qquad d(0.001, 0.002) = 0.001$
Suppose a system maintains a target value $T$. Noise causes perturbations $\varepsilon_1, \varepsilon_2, \ldots$ over time. After $N$ perturbations:
$$x_N = T + \sum_{i=1}^{N} \varepsilon_i$$The total error, by repeated application of the triangle inequality:
$$E_N = \left|\sum_{i=1}^{N} \varepsilon_i\right| \leq \sum_{i=1}^{N} |\varepsilon_i| \leq N \cdot \varepsilon$$The total error can be as large as the sum of the individual errors. With enough perturbations, the error can exceed any finite bound.
Even with zero-mean random perturbations, the RMS error grows as $\sigma\sqrt{N}$ — without bound. Continuous memory inevitably decays unless actively corrected.
A sensor at 5.000 V, with ±0.001 V noise per second:
Left: Continuous memory — error walks away from zero. Right: Hierarchical memory — threshold suppresses errors. Click Start and watch the difference.
The only defense against accumulation is active error correction: constantly measure, compare, and correct. This costs energy — at minimum $k_B T \ln 2$ per erased bit (Landauer's principle).
For an $n$-bit state with noise bandwidth $B$, the power dissipation scales as:
$$P_{\text{correction}} \gtrsim n \cdot B \cdot k_B T \ln 2$$For quantum computers with thousands of logical qubits and millions of physical qubits measured at gigahertz rates, the energy cost becomes prohibitive. This is the thermodynamic wall — the point where active correction consumes more resources than computation itself.