1
00:00:00,000 --> 00:00:03,887
Sound, images, sensor readings — almost every signal you meet is

2
00:00:03,987 --> 00:00:08,934
really a recipe. Take simple ingredients, pure oscillations at fixed frequencies,

3
00:00:09,034 --> 00:00:13,358
and add them with different strengths and phases. The Fourier transform

4
00:00:13,458 --> 00:00:16,100
is the tool that reads that recipe back out.

5
00:00:16,203 --> 00:00:20,501
Here is a deliberately small example: two cosine waves added together,

6
00:00:20,601 --> 00:00:23,642
two and three hertz. On a graph it just looks like

7
00:00:23,742 --> 00:00:27,537
a busy squiggle. Nothing screams two separate tones — yet they

8
00:00:27,637 --> 00:00:30,239
are both in there, waiting to be separated.

9
00:00:30,353 --> 00:00:32,854
This curve is g of t, height versus time.

10
00:00:32,954 --> 00:00:35,964
Our job is to discover which pure frequencies are

11
00:00:36,064 --> 00:00:39,517
actually present, without guessing from the shape alone.

12
00:00:39,620 --> 00:00:43,512
The trick is geometric. Pick a test frequency, and wrap the wave

13
00:00:43,612 --> 00:00:47,941
around the origin. Each moment contributes a point in the plane: radius

14
00:00:48,041 --> 00:00:52,183
from the signal's value, angle spinning at your chosen rate. As time

15
00:00:52,283 --> 00:00:56,176
runs forward, you trace a closed-looking path — the wound curve.

16
00:00:56,286 --> 00:00:59,556
For most test rates the path is symmetric enough that

17
00:00:59,656 --> 00:01:03,562
everything cancels. The average position, the center of mass of

18
00:01:03,662 --> 00:01:07,123
that winding, sits almost at the origin. That means this

19
00:01:07,223 --> 00:01:09,794
test frequency is not a strong ingredient.

20
00:01:09,903 --> 00:01:13,368
But when the wrapping rate lines up with a frequency that

21
00:01:13,468 --> 00:01:17,622
really lives inside the signal, the symmetry breaks. The path bulges

22
00:01:17,722 --> 00:01:21,000
to one side, and the center of mass jumps outward. You

23
00:01:21,100 --> 00:01:23,627
have found a spike of energy at that rate.

24
00:01:23,886 --> 00:01:27,662
Sweep the test frequency and record how far the center drifts

25
00:01:27,762 --> 00:01:31,348
from zero. That trace is the magnitude spectrum: a plot of

26
00:01:31,448 --> 00:01:35,541
how much each frequency contributes. Every spike marks a pure tone

27
00:01:35,641 --> 00:01:38,210
that was hiding in the squiggle all along.

28
00:01:38,320 --> 00:01:42,985
The same idea lives in three dimensions. Stack time vertically, and let each

29
00:01:43,085 --> 00:01:46,560
slice be the rotating complex value g of t times e to the

30
00:01:46,660 --> 00:01:49,946
minus two pi i f t. From the side you see the familiar

31
00:01:50,046 --> 00:01:54,461
wound curve; from above you sense the drift of the center. When rotation

32
00:01:54,561 --> 00:01:58,913
rate matches content, the path fails to close neatly through the bulk —

33
00:01:59,013 --> 00:02:02,676
that off-center drift is exactly what the integral measures.

34
00:02:02,786 --> 00:02:07,099
In one line: multiply by a spinning complex exponential, integrate, and

35
00:02:07,199 --> 00:02:11,326
you extract each frequency component. This is the Fourier transform.