Monte Carlo integration estimates the rendering-equation integral by averaging random samples

The rendering (light transport) equation expresses light leaving a point as an integral over all incoming directions — a continuous quantity with no closed-form solution for general scenes. Monte Carlo integration approximates it numerically: draw a finite number of random samples (directions, points, or whole light paths), evaluate the integrand at each, and average. The average is an unbiased estimator that converges to the true integral as sample count N grows, but slowly — error falls only as 1/sqrt(N), so quartering the error (halving noise) needs roughly four times the samples, and residual error appears as image noise. Unlike deterministic quadrature, Monte Carlo handles arbitrary BRDFs and complex light paths without special cases. How samples are drawn matters enormously: variance-reduction techniques — importance sampling (concentrating samples where the integrand is large, e.g. by the BRDF), stratified sampling, Russian roulette, and multiple importance sampling — reach a clean image with fewer samples.

Examples

Area-light soft shadow: sample 16 random points on the light, trace a shadow ray to each, average visibility → a smooth penumbra; one sample gives all-or-nothing noise. Path tracing: cast rays per pixel, choosing bounce directions randomly by the BRDF, and average many paths. Estimating pi by testing random points against an inscribed circle uses the same sampling logic.

Assessment

Explain how ray tracing approximates the rendering-equation integral with a finite number of samples, and why importance sampling from the BRDF reduces variance versus uniform sampling. If 100 samples give noise level X, roughly how many give X/10, and why? Why does Monte Carlo handle complex BRDFs where grid quadrature fails?