Fire up a path tracer, let it throw a single ray-bundle at each pixel, and the first frame looks like a photograph buried under television static — a grainy, speckled mess. Let it keep going and the grain slowly melts away until the image is clean. What you are watching is Monte Carlo integration at work: the renderer is estimating the colour of every pixel by averaging random light paths, and the graininess is the estimator's error, called variance, quietly shrinking.

This page explains why physically-based rendering is noisy in the first place, exactly how fast that noise fades (spoiler: painfully — as 1/\sqrt{N}), and the clever variance-reduction tricks — importance sampling, stratification, quasi-Monte Carlo, and multiple importance sampling — that are what actually make a path tracer usable, far more than just firing more rays. It builds directly on the rendering equation and path tracing, and leads into denoising, which cleans up the noise that no amount of sampling can economically remove.

The one idea: an integral is an average

The colour arriving at a pixel is an integral — a sum over every direction light could have come from, weighted by how the surface reflects it. There is no formula for that integral in a real scene; it depends on the whole world of geometry and lights. Monte Carlo integration turns the integral into something we can compute: a random average.

Applied to rendering, each random light path traced from the camera into the scene is one sample X_i; the light it carries back, divided by the probability of having chosen that path, is one term f(X_i)/p(X_i); and the pixel colour is the average of those terms. One sample is a wild guess. A thousand samples average out into a picture.

Why it is noisy: variance and the 1/\sqrt{N} wall

Each single-sample estimate f(X_i)/p(X_i) scatters around the true answer. The spread of that scatter is the variance \sigma^2. Averaging N independent samples shrinks the variance of the average by a factor of N:

The visible grain in the image is that standard error, and it falls only as 1/\sqrt{N}. This is the central, brutal fact of path tracing: to halve the noise you must take four times as many samples; to make the image ten times cleaner costs a hundred times the rays. Brute force runs into a wall of diminishing returns almost immediately — the last bit of grain is fantastically expensive to remove by counting alone.

Seeing the wall: error versus sample count

Here is the whole tragedy in one plot. The curve is the Monte Carlo error \sigma/\sqrt{N} (with \sigma=1) as a function of the number of samples N. Drag N and watch the marker: early on, each extra sample buys a big drop in noise; but the curve flattens fast, and past a few hundred samples you are pouring rays in for almost no visible gain. The faint straight line is a hypothetical 1/N decay — what a good quasi-Monte Carlo sequence can approach, and why it is worth the trouble.

The gap between the two curves is exactly the prize that variance reduction chases. You cannot change the 1/\sqrt{N} law for plain random sampling, but you can shrink the constant \sigma in front of it — often by orders of magnitude — and you can switch to samples that decorrelate the error toward 1/N. That is what the rest of the page is about.

Worked example: shrinking \sigma with importance sampling

Forget rendering for a moment and estimate a concrete integral, I = \int_0^1 2x\,dx = 1. We will estimate it two ways and compare the variance.

Uniform sampling. Draw X uniformly on [0,1], so p(x)=1. Each sample contributes f(X)/p(X) = 2X. The mean is \int_0^1 2x\,dx = 1 (unbiased, good), and the variance of a single sample is \sigma^2_{\text{unif}} = \int_0^1 (2x)^2 dx - 1^2 = \tfrac{4}{3} - 1 = \tfrac{1}{3} \approx 0.333. The samples scatter a lot, because 2x is 0 at the left and 2 at the right — most of the "value" hides near x=1, where uniform sampling rarely looks.

Importance sampling. Now put the density where the integrand is large: choose p(x) = 2x (a valid density — it integrates to 1 on [0,1]). Then every sample contributes \frac{f(X)}{p(X)} = \frac{2X}{2X} = 1 — a constant! Every single sample returns exactly the right answer, so the variance is \sigma^2_{\text{IS}} = 0. The estimator is perfect from one sample.

In a renderer we do not know f — it is the full light transport — but we know its factors: the BRDF has a bright specular lobe, and the light sources are where the energy comes from. So we sample directions towards the light and along the BRDF lobe instead of uniformly over the hemisphere. Each path then carries more weight and less scatter, and the same picture emerges in a fraction of the samples.

The variance-reduction toolbox

Practical path tracers stack several tricks, each attacking the noise from a different angle.

Why some scenes stay stubbornly noisy

Even with the whole toolbox, a few light-transport effects fight back. Caustics — the bright ripples where light focuses through glass or off water — arrive along rare, tightly-focused paths that unidirectional sampling almost never finds, so they stay speckled for a very long time. Small, intense lights (a bare bulb, the sun through a keyhole) concentrate enormous energy into a tiny solid angle: miss it and you get black, hit it and you get a blinding firefly — a single pixel spiking far above its neighbours because one sample had a huge f/p. High variance is firefly pixels.

These are exactly the cases where throwing more samples is uneconomical, and where the next lesson picks up: rather than integrate the noise away, we denoise — use a smart filter (often a neural one) to reconstruct a clean image from a noisy, low-sample render, trading a little bias for a huge saving in rays.