Why does inverse-polynomial noise allow quantum separation while inverse-polylogarithmic noise destroys it—what changes at that threshold?

This is a sharp question about a genuine phase transition in quantum complexity theory. The answer hinges on how noise *scales* with system size—and there's a critical exponent where quantum advantage flips on and off.

The Core Threshold

The dividing line is whether noise decays faster than logarithmically. [AG-2024.05-1599] proves:

• Inverse polynomial noise (e.g., $1/n$, $1/n^2$): quantum-classical separation *persists*. Shallow quantum circuits can solve problems that require exponentially deep classical networks, even in the presence of noise.

• Inverse polylogarithmic noise (e.g., $1/(\log n)^2$): the separation *vanishes*. No super-polynomial quantum advantage survives.

Why does this threshold exist? The intuition is about signal-to-noise ratio. An inverse-polynomial noise rate decreases *slowly enough* that the quantum circuit's nonlocal correlations—the feature giving it computational power—can still dominate the computation before errors overwhelm the signal. Inverse-polylogarithmic noise, by contrast, decreases *so slowly* that errors accumulate faster than the quantum advantage can compound, even in shallow circuits.

What Changes at the Boundary

[AG-2024.05-1599] shows this is not arbitrary: the separation persists under depolarization noise (a standard realistic noise model) whenever the noise strength is $O(1/\text{poly}(n))$, and fails when noise is $\Omega(1/\text{polylog}(n))$. The quantum nonlocality property that distinguishes quantum from classical circuits simply cannot survive noise that decays *that* slowly.

This matters because it establishes the *minimum* noise suppression real quantum devices need to demonstrate genuine quantum advantage in learning tasks, even with constant-depth circuits that sidestep the need for error correction [AG-2024.05-1935].

Related work explores noise-induced *phase transitions* in other quantum settings—for instance, [AG-2024.01-2041] studies how size-dependent noise probabilities $q = p/L^\alpha$ trigger transitions in entanglement structure, finding a sharp shift at $\alpha = 1$. The broader picture is that noise doesn't degrade quantum systems monotonically; rather, its effect depends critically on *how it scales*.