The square root problem

The law that governs every deadline

Render a frame with a hundred samples per pixel and it is noisy. Double the samples to two hundred and the noise drops — but only by a factor of roughly 1.4, the square root of two. To halve the noise you need four times the original count. To halve it again, four times that. The relationship is merciless: error falls as one divided by the square root of the sample count, and nothing about modern hardware changes the exponent.

This is not a software limitation or a temporary one. It follows directly from what a renderer is doing: estimating an integral by averaging a finite number of random measurements. Each sample is an independent draw from a distribution, and the standard deviation of the mean of n independent draws is the standard deviation of one draw divided by √n. That is the Central Limit Theorem applied to light transport, and it holds whether you are path-tracing on a laptop or on a thousand-core farm. As noise is a budget, the only question is what you are spending it on.

What it costs in practice

Say a shot takes one hour at two hundred samples and the result is too noisy. The supervisor asks for half the noise. That means eight hundred samples — four times the count — and, assuming linear scaling with sample count, four hours. Halve it again and the render is sixteen hours. The clean, pristine frame you are imagining lives at some sample count you cannot afford before the delivery date, and the square root curve means each marginal improvement costs more than the last.

The practical response is to stop thinking of sampling as a quality dial and start thinking of it as a budget with diminishing returns. Early samples buy a lot: going from twenty-five to one hundred cuts noise in half. Going from ten thousand to forty thousand does the same thing but costs forty thousand samples to get there. The curve is steep at the low end and nearly flat at the high end, which is why a noisy render can look dramatically better with a modest increase in samples, and why a near-clean render is so expensive to finish.

This asymmetry shapes every production decision about where to spend render time. Uniform sampling — the same count everywhere in the frame — is almost always wasteful, because a clear blue sky and a shadow caustic do not have the same variance and do not need the same budget. Adaptive sampling is the systematic answer to that: measure the variance per pixel and spend more samples where the estimate is still uncertain. It does not repeal the square root law; it applies it more efficiently by concentrating the budget where the return is still on the steep part of the curve.

Denoising is a different kind of answer. Rather than buying the last factor of two through sampling, you stop short and use a spatial filter to suppress the remaining noise. The tradeoff is detail: the filter cannot always distinguish a genuine high-frequency texture from sample variance, and some of what it removes is real. But on a sixteen-hour render the alternative is sixty-four hours, so the tradeoff is usually taken. What matters is understanding that denoising is not free noise reduction — it is a trade of one kind of error for another, made because the square root curve has made the honest path too expensive.

The number that changes a deadline is not the sample count itself but where you currently sit on the √n curve. At low sample counts you are on the steep slope, and more samples help fast. At high counts you are on the flat shoulder, and nothing short of importance sampling, better light paths, or a looser noise tolerance will get you home.