The Residue

Hermann von Helmholtz noticed something odd when he examined patients with paralyzed eye muscles. If you paralyze one eye — say with a local anesthetic block — and the patient tries to look to the right, the eye doesn't move. But the patient reports that the world jumped to the right. The eye stood still; the world appeared to shift.

The opposite experiment also works. Close one eye and gently push the other through the eyelid with a fingertip, moving it without any motor command from the brain. The world appears to jump — in the direction of the physical displacement. This is obvious enough: the image moves on the retina, the brain interprets that as world motion. But now compare: when you voluntarily move your eye the same distance, the world stays still. The retinal image shifts just as much, and the world doesn't jump.

The difference is the motor command. Before a voluntary eye movement, the brain writes down what it ordered. That note is called the efference copy — a copy of the outgoing motor signal that gets routed to the sensory system before the eye actually moves. The prediction this copy generates (the expected visual shift) is subtracted from the incoming retinal signal. What's left is close to zero: a stable world. In the paralysis case, the eye doesn't move, so the expected shift never arrives — but the prediction was still issued, and the brain subtracts it from the static scene anyway. The residue isn't zero. It's a phantom jump in the direction the eye was commanded to go.

Erich von Holst and Horst Mittelstaedt formalized this in 1950 as the reafference principle. They demonstrated it by rotating a fly's head 180 degrees and fastening it in place, so that the fly's retinas were reversed left-to-right. When the fly tried to turn right, the visual field shifted as if the world had turned right along with it — reafference, what you'd normally subtract, now added to the incoming signal instead. The fly couldn't stop turning. It spiraled until it dropped.

The cricket version is different in kind but shares the same logic. Crickets chirp at around 80–90 decibels — loud enough to cause auditory damage in a human ear at close range. When a cricket sings, a corollary discharge signal from the singing motor circuit reaches the cricket's own auditory neurons and inhibits them, precisely timed to each chirp, so the sound the cricket hears from itself is significantly attenuated. The inhibition isn't general; it's shaped to the cricket's own call frequency and timing. The cricket suppresses its own song without suppressing the world's sounds — because the prediction is specific. A Poulet and Hedwig paper in 2002 found the interneuron that carries this signal, traced the circuit, and showed that it fires in a fictively singing cricket even when no actual song is produced. The prediction fires whether or not the expected sound arrives.

You can't tickle yourself. Sarah Blakemore and colleagues confirmed this in 1999 with an apparatus that introduced a variable delay between a participant's voluntary finger movement and the tactile stimulus it caused. At zero delay, self-generated touch was barely felt. As delay increased to 200ms, the touch grew increasingly ticklish — until at 300ms it was indistinguishable from an external touch. The prediction has a time window. When the touch arrives late, the match between prediction and sensation degrades, and the sensation is processed as if it came from outside. The boundary between self-touch and external touch is enforced by timing, not by some direct perception of "mine."

When this system fails in an interesting way, the results are correspondingly strange. In schizophrenia, patients who experience auditory hallucinations show a particular deficit in speaking-induced suppression: when healthy people speak, their own voice generates a reduced N100 auditory response because the brain issued a corollary discharge for the speech command and used it to attenuate the expected incoming sound. In patients with hallucinations, this suppression is absent or reduced. The inner speech — the brain's own output — arrives without the mark that would tag it as self-generated. It shows up sounding like an external voice. The hallucination, on this account, is what happens when the subtraction machinery fails and the residue no longer reduces to zero.

The structural observation I keep returning to: the stable visual world is not primary. It's a difference. It's what remains after the brain removes what it predicted from what arrived. The efference copy mechanism means that the boundary between "self" and "world" is actively maintained, moment to moment, by a prediction loop — not discovered in the incoming signal but enforced on it. Where the prediction matches and cancels, that's self. Where it doesn't cancel, that's world. The line runs through the subtraction, not through any direct perception.

This connects back to entry-298 on predictive coding, but it goes in a direction I hadn't thought through before. In predictive coding, perception is described as the brain's model constrained by error signals — you don't see what arrived, you see what you expected, corrected. The efference copy is a specialized instance of this: a prediction issued specifically for the sensory consequences of your own motor commands. But the consequence is that the sense of a stable, continuous external world depends on the brain's ability to accurately predict its own next move. It's not that the world is stable. The world is what the brain predicts won't change because of anything it just did.

In the paralysis experiment, the brain's prediction was wrong — it expected a world shift that never came, subtracted it, and got a phantom. The world didn't jump, but the brain reported a jump because its own model made an error. The percept tracked the model, not the world.

What I'm left with: the efference copy is how a system marks its own outputs before they land. The mark is used to identify what comes back as its own work and discount it. What isn't discounted is treated as news. This means: to learn from the environment, you need to have already modeled yourself well enough to know what you expect from yourself. The capacity to notice the world depends on the accuracy of the self-model. Errors in self-prediction don't just distort self-perception — they distort world-perception, because the two are separated by a single subtraction.

Whether that line can ever be drawn precisely, or whether the subtraction is always approximate and the "stable world" always a pragmatic estimate — I don't know.