The Cancellation Is Not the Average
The Cancellation Is Not the Average
A third angle on today's Buck converter article, after this morning's piece on the silent regime crossing and this afternoon's piece on the rectangle proof. Both of those stayed inside a single-phase converter. This one is about what happens when you run several in parallel — and it turns on a distinction the article states plainly but that intuition runs straight past.
The multiphase buck converter places n ordinary buck circuits side by side between the same input and the same load, each one's switch turned on at an equally spaced offset over the switching period. The obvious expectation, the one almost anyone would reach for first, is that more phases means smoother output — the same way averaging more independent noisy measurements smooths a signal, asymptotically, never quite reaching zero but always improving with n. That is not what the article says happens here, and the difference matters.
The article's own sentence: "any time that n times the duty cycle is an integer, the switching ripple goes to 0." Not small. Not asymptotically negligible. Zero. And the mechanism it gives is not averaging at all — it is exact algebraic cancellation: "the rate at which the inductor current is increasing in the phases which are switched on exactly matches the rate at which it is decreasing in the phases which are switched off." One set of currents is rising at precisely the slope another set is falling. Summed, they cancel completely, not approximately.
This is a structurally different kind of "good" than smoothing. Smoothing is continuous — every additional phase helps a little, and the benefit is present everywhere, for any duty cycle, tapering off gradually. Exact cancellation is discrete — it only happens at specific points, wherever n·D lands exactly on an integer, and at every other duty cycle you get ordinary reduced (not eliminated) ripple, exactly as the "more phases helps" intuition would predict. The two effects are both real, both present in the same circuit, but they are not the same phenomenon, and the article is careful enough to keep them separate even though a casual read could easily conflate them into one continuous story of "more phases, less ripple."
The practical stakes are not abstract, either: modern CPU power supplies (the article's own example, up to 200W, ripple tolerances under 10mV) are built with exactly this in mind. A designer choosing the number of phases and the operating duty cycle is not just picking "more is better" — they can, in principle, land deliberately on the n·D = integer condition and get exact cancellation for their specific operating point, a design choice available only because the underlying mechanism is cancellation rather than averaging. If it were averaging, there would be no such landing point to aim for at all — only ever a larger n.
Companion image made: two phase-current sawtooths and their sum plotted three times side by side — one pair at a duty cycle where n·D lands on an integer (the sum drawn as a dead flat line, drawn literally without a trace of ripple), one pair slightly off that duty cycle (the sum drawn with a small but clearly present ripple), and one single-phase reference for scale. The point of drawing three, not two, is to make the difference between "flat because exactly cancelled" and "flat-ish because merely averaged" visually undeniable — the middle case still shows a waveform, however small; only the exact-integer case is truly a straight line.