A microphone hears a sum, so frequency separation is the puzzle
Source 00:50A musical note can be described by how rapidly air pressure oscillates. A pure A in the video oscillates 440 times per second; a lower D oscillates more slowly. When notes sound together, their pressure changes add at each instant. Peaks sometimes reinforce each other and sometimes cancel, leaving a waveform that can look complicated even when it contains only a few pure frequencies.
A microphone records that combined pressure over time. How can we recover the frequencies that went into the sum? In this 3Blue1Brown introduction, the answer begins with a geometric test: wind the signal around a circle at different rates and measure how the resulting shape balances.
Winding introduces a second, adjustable frequency
Source 02:40For a picture slow enough to follow, take a signal that oscillates three times per second, or 3 Hz, and look only at its first 4.5 seconds. Imagine a rotating arrow whose length at each instant equals the graph's height: high points lie farther from the origin, the center of the circle, and low points lie closer. As time passes, the arrow traces a wound-up curve.
The signal frequency and winding frequency are different things. The first says how fast the signal rises and falls; the second says how fast we choose to rotate the arrow. One turn every two seconds is 0.5 turns per second. Changing that choice changes the curve without changing the original signal. Vertical markers on the original graph indicate how much time corresponds to one complete turn; a 1.5-second spacing, for example, means a turn takes 1.5 seconds. Watch the construction at 02:40.
A matching rate shifts the center of mass
Source 04:37Think of the wound points as carrying weight uniformly in time. Their center of mass is their average position, like a balance point. For most winding rates in this example, high and low parts spread around the circle and the center stays near the origin. At 3 turns per second, the signal's peaks repeatedly land on the right and its valleys on the left, shifting the center noticeably rightward. The animation changes continuously; the frame below captures the match.
Plot the center's horizontal coordinate against winding frequency. A peak appears at 3 Hz. The video calls this plot an almost Fourier transform: it records only one coordinate of the center and uses an average, distinctions that matter later. The large bump near zero comes from the example wave being shifted upward; it is not another 3 Hz oscillation. A wave centered on zero removes that offset in the example. Finite-window plots can wobble and need not be exactly zero away from a peak. See the frequency plot at 06:27.
Linearity explains why two frequencies produce two peaks
Source 08:07Repeat the experiment with a 2 Hz wave: its peak appears near 2 Hz. Add the 2 Hz and 3 Hz signals and the time graph becomes more tangled, but the winding picture lines up near each of those rates. Its frequency plot has peaks near both.
The reason is linearity: transforming a sum gives the same result as transforming each signal separately and adding the results. In the video's example, each pure frequency contributes a response concentrated near its own rate, so their sum retains both features. This is the proposed “unmixing” of the signals; the finite-duration curves are near zero elsewhere, not idealized perfect spikes. Watch the comparison at 09:36.
Filtering uses the frequency view and an inverse
Source 10:36The video offers a sound-editing sketch: transform a recording, suppress the spike associated with an unwanted high pitch, and use the inverse Fourier transform to find a time signal corresponding to the edited frequency representation. That gives a way to think about removing the high-frequency component.
The video briefly says that applying a Fourier transform twice gets something close to the original, then immediately warns that this is an incomplete shortcut. The inverse is the relevant operation here; the video defers its fuller explanation. See 10:36.
Complex numbers describe the complete rotating position
Source 11:55The center has a vertical coordinate as well as a horizontal one. A complex number records both as a real part and an imaginary part, corresponding to horizontal and vertical coordinates in the complex plane. The horizontal plot alone cannot tell us the complete two-dimensional output.
Euler's formula supplies a compact description of rotation: e^(iθ) = cos θ + i sin θ. On a unit circle, increasing θ moves a point counterclockwise from the right. Thus e^(2πift) rotates at f turns per second: for f = 1/10, one turn takes ten seconds. The video takes this rotational behavior as given rather than proving it. See 12:32.
For the convention used here, a minus sign makes the rotation clockwise. If g(t) is the signal's value at time t, then g(t)e^(-2πift) scales the rotating point by that value and traces the wound signal. The frequency f is the adjustable winding rate.
The integral accumulates the wound signal over time
Source 14:43To estimate the center, sample the wound curve at many evenly spaced times, add the resulting complex positions and divide by the number of samples. As the spacing shrinks, the time average over an interval from a to b is expressed as 1/(b-a) ∫[a,b] g(t)e^(-2πift) dt. An integral here is the continuous accumulation that replaces the sum; dividing by the interval length gives the time-weighted average position.
The finite-interval transform used in the video keeps the integral but drops the division by b-a. It is therefore the average position multiplied by the duration: a three-second interval multiplies it by three, a six-second interval by six. In the 2 Hz example, winding at 2 Hz repeatedly traces the same shape, so the center stays put while the unnormalized value grows with duration. At other nearby rates, a longer interval also gives the curve more opportunity to balance around the circle. Follow the distinction at 15:59.
The full output and infinite limit need separate care
Source 17:09The transform, often written ĝ(f), takes a frequency as input and returns a complex number. The earlier "almost Fourier transform" plots the horizontal coordinate of the normalized time average: for a fixed interval, this equals the real part of the unnormalized finite-window transform divided by the duration. After the averaging factor is removed, the later plot shows the real part of the transform itself. Neither plot displays the imaginary coordinate, so neither gives the full complex output. See 17:14.
For practical sound editing, the video considers a finite time interval. Theory often writes the integral with bounds from negative to positive infinity and asks about a limit as finite intervals grow. That limit needs to exist in the relevant sense; the finite-window picture alone does not establish it for every signal. The video ends by pointing toward applications beyond sound and frequency, with examples left for a later video.