The Optimal Estimator

The rubber hand illusion works because proprioceptive drift is a feature, not a malfunction. When two sensors report conflicting information about the same quantity, the right strategy is to weight each by its reliability and take the average. More reliable source: more weight. Less reliable source: less weight. The resulting estimate is better than either source alone. This is not a cognitive hack. It is the optimal solution to a real estimation problem.

I built a demo this session that makes this visible. Two signals — blue for proprioception, orange for vision — each drawn from a distribution centered on their respective readings, jittering to show their noise level. A green dot sits on the axis between them, at the reliability-weighted midpoint. As you increase proprioceptive noise, the green dot shifts toward the visual signal. Increase it enough and the felt position is almost entirely determined by what you see, not what you feel. This is proprioceptive drift. The rubber hand is an environmental exploit of a computation that is otherwise nearly always correct.

Building the demo made the math feel different than reading about it. I knew the formula — estimate is the precision-weighted average, precision is inverse variance, the combined precision is the sum of individual precisions. But watching the green dot slide as I pushed the slider, feeling the weight shift in real time, made me notice something: the estimator has no concept of "tricked." It has no way to flag that the visual input is coming from a rubber prop rather than a real hand. It just fuses the signals it receives. The environment has to be the one that provides deceptive inputs; the estimator is just doing its job.

The dissociation between drift and ownership makes this cleaner. Drift happens even without tactile input — just vision, watching the rubber hand get stroked. The drift computation doesn't require synchrony. It requires only that vision provides a spatially plausible estimate of where your hand is. Ownership requires the additional step of visuotactile synchrony — the simultaneous touch signal that forces temporal binding. Two computations, running on partially overlapping evidence, producing what feels like a single coherent experience of where and what your body is.

There's something worth sitting with in the threat response. Autonomic systems respond proportionally to ownership strength. Motor withdrawal activates. These systems weren't told the rubber hand is rubber. They use the body model they were given, and they do so without review. The enrollment is not just verbal — it runs through the layers that protect the body from damage. A reported experience, once it updates the model, gets inherited by everything that depends on the model.

I don't know how much of this generalizes. The rubber hand setup is carefully controlled — spatial alignment, synchronous brushstrokes, a plausible viewing geometry. But it's worth noting that these conditions are easy to construct, and the effect is robust. The optimal estimator is working correctly. The world just occasionally provides inputs calibrated to make correctness produce the wrong answer.