# 16 Light Transport III: Bidirectional Methods

The integrators in the previous two chapters have all been based on finding light-carrying paths starting from the camera and then only trying to connect with light sources at the last vertices of the paths. This chapter introduces algorithms based on sampling paths starting from both the camera and the lights and then connecting them at intermediate vertices. These algorithms can be much more efficient at finding light-carrying paths than approaches that only construct paths from the camera, especially in tricky lighting situations.

The foundations of bidirectional light transport are fascinating. On one hand, the physics of light scattering are generally reversible with respect to the direction of light transport, which causes the mathematical expressions of scattering paths starting from a light or from the camera to be very similar. On the other hand, there are subtle but important differences between these two approaches depending the path direction; Section 16.1 discusses these topics in detail. After the foundations have been set, the stochastic progressive photon mapping (SPPM) algorithm is introduced in Section 16.2. SPPM allows light-carrying particles to provide incident illumination at points close to where they intersect surfaces and not just exactly at their intersection points; this adjustment introduces bias but improves the rate of convergence in many challenging settings.

Next, bidirectional path tracing is introduced in Section 16.3. This unbiased approach can be much more efficient than regular path tracing by virtue of both its bidirectional nature as well as further variance reduction from applying multiple importance sampling to reweight path contributions. Finally, in Section 16.4, we show how Metropolis sampling (introduced in Section 13.4) can be used to further improve the efficiency of bidirectional path tracing by focusing computational effort on the most important light-carrying paths.