systems theory//engineering patterns//error relocation

Error relocation is the engineering principle that error in a real system is never removed, only moved from one place to another, so design consists of choosing where it should live and checking with a budget that it fits there; it is used to see the hidden cost of every fix before paying it in the field. A gyro is noisy, so the readings are filtered harder. The noise drops. But every filter that smooths also delays, the controller now corrects with old information and the drone starts to oscillate: the error moved from noise to delay, which for a control loop is a worse place to keep it.

Error relocation is the engineering principle that error in a real system is never removed, only moved from one place to another, so design consists of choosing where it should live and checking with a budget that it fits there; it is used to see the hidden cost of every fix before paying it in the field. A gyro is noisy, so the readings are filtered harder. The noise drops. But every filter that smooths also delays, the controller now corrects with old information and the drone starts to oscillate: the error moved from noise to delay, which for a control loop is a worse place to keep it.

The same exchange appears in every chapter of a loop, each time with a knob that slides error from one form into another. Smoothing trades noise for lag (digital filter). Regularization trades variance for bias (bias-variance trade-off): a physics model swapped for a network that fits the test data better lowers model error and brings it back as generalization error the day the drone carries a heavier payload than any it trained on. Discretization and linearization trade accuracy for tractability. A lower detection threshold trades detection delay for false alarms. More links in a consensus protocol trade slowness for fragility under delay. In a feedback loop the exchange is an identity, S+T=1S+T=1S+T=1 (sensitivity function): rejecting disturbances at a frequency means passing sensor noise at that same frequency. And a simulator trades field cost for the sim-to-real gap.

Since the error will go somewhere, the useful question is where it does least harm, and the tool that answers it is the error budget: every source carried to the same units at the same point, combined, and the largest fixed first. Each fix is followed by a new reading of the table, because the dominant source changes: on the book's facade-inspection drone it changed twice in three steps.

A fix that fattens another source is the common trap. A narrower innovation gate rejects more outliers and loses the target in a fast turn; a faster sampling rate cuts aliasing and sampling delay and adds CPU load and jitter to every other task. Re-running the budget after each change is the only honest check.

Some places are worse than others for the same amount. Delay inside a feedback loop costs phase margin and can destabilize; the same error as a small static offset may be harmless. Choosing the place is therefore judged by what the error does where it lands, with the magnitude only half of the answer.

Not every improvement is relocation. A better sensor, a timestamp that removes uncompensated latency or more informative data can lower the total; the principle says there is no free knob inside a fixed design, so a real reduction costs a new sensor, new information or new computation.

The principle is the book's central maxim, and it connects with flyswatter rule (the simple method fails where you can see it) and with the siblings in engineering patterns.