Functions as vectors, integrals as dot products
In the first guide of this rung we found the Fourier coefficients by a magic-looking move: multiply by one wave, integrate over a period, and watch every other term vanish. That move was not magic — it was geometry in disguise, and this guide makes the geometry explicit so you can see why everything stands up. The one idea to hold onto is this: a function on an interval behaves exactly like a vector, only with infinitely many components. To turn that slogan into something you can compute with, you need a way to take the dot product of two functions.
For ordinary vectors u and v in space the dot product is u.v = u_1 v_1 + u_2 v_2 + u_3 v_3 — you pair up matching components and sum. For two functions f and g on an interval, replace 'pair up matching components and sum' with 'multiply the values that line up and add them across the whole interval', which is exactly a definite integral: define the inner product as <f, g> = integral over one period of f(x) g(x) dx. The discrete sum becomes a continuous integral because a function has a value at every point, not just at slots 1, 2, 3. With this one definition, the entire vocabulary of geometry — length, angle, perpendicular, projection — transfers wholesale onto functions.
Orthogonality: the trig waves are perpendicular axes
Now compute that inner product for two waves from our catalogue. The central fact — the orthogonality of the trig system — is that any two DIFFERENT waves have inner product exactly zero. Over the interval from -L to L: the integral of sin(m pi x / L) sin(n pi x / L) dx is 0 whenever m is not n; cosine with a different cosine is 0; and EVERY sine paired with EVERY cosine is 0. The reason is concrete, not mystical: a product of two waves of different frequencies is itself an oscillation that spends exactly as much time above zero as below, so over a full period the positive and negative areas cancel to nothing.
The only time the inner product is NOT zero is when a wave meets its own twin. The integral of cos^2(n pi x / L) over -L to L is L, and likewise for sin^2 — because a squared wave is never negative, so its area cannot cancel. In vector language: every wave has squared length L (it is not a unit vector, just a fixed length), and any two distinct waves are perpendicular. So the family { 1, cos(pi x / L), sin(pi x / L), cos(2 pi x / L), sin(2 pi x / L), ... } is a set of mutually perpendicular axes in the space of functions, exactly the way the standard basis vectors point along x, y, z — only here there are infinitely many of them.
Pythagoras in infinite dimensions: Parseval's theorem
Here is the payoff that orthogonality unlocks. For an ordinary vector written in perpendicular components, its squared length is the sum of the squared components: |v|^2 = v_1^2 + v_2^2 + v_3^2. That is just Pythagoras. Because the trig waves are perpendicular axes, the very same theorem holds for functions — and it has a name, Parseval's theorem. The 'squared length' of f is <f, f> = integral over a period of [f(x)]^2 dx, and Parseval says this equals the sum of the squared contributions of every harmonic, with no cross terms surviving. The perpendicularity is precisely what kills the cross terms, just as the absence of an x.y term in |v|^2 is what makes Pythagoras clean.
PARSEVAL'S THEOREM (trig Fourier series, period 2L)
(1/L) * integral_{-L}^{L} [f(x)]^2 dx
= a_0^2 / 2 + sum_{n>=1} ( a_n^2 + b_n^2 )
WHY it falls out of orthogonality:
square the series f = a_0/2 + sum [ a_n cos_n + b_n sin_n ]
integrate term by term over one period.
- cross terms cos_m cos_n (m != n), sin_m sin_n, cos sin -> integrate to 0
- only the SELF terms survive: integral cos_n^2 = L, integral sin_n^2 = L
divide by L -> each a_n, b_n contributes its own square, nothing else.
ENERGY reading: left side = average power of the signal over a period;
right side = power summed harmonic by harmonic.Why call the left side 'energy'? In physics a great many quantities scale as the square of an amplitude: the energy of a vibrating string, the power dissipated by a current, the intensity of a sound. The time-average of [f(x)]^2 over a period is exactly that average power. So Parseval says: the total power of a signal equals the sum of the powers carried by its individual harmonics. Nothing is lost and nothing is double-counted in passing from the wiggly time picture to the bar-chart frequency picture — the discrete spectrum. That single sentence is why engineers can speak of 'how much energy sits at 60 Hz' as if frequencies were separate, independent bins. Orthogonality is what makes the bins independent.
A worked tally, and a free gift to series
Let us cash this out on the square wave from the first guide — the odd wave equal to +1 on (0, L) and -1 on (-L, 0), whose coefficients were b_n = 4/(n pi) for odd n and zero otherwise. The left side of Parseval is trivial: [f(x)]^2 is just 1 everywhere (since f is plus or minus 1), so its average is 1. The right side is the sum of b_n^2 over odd n, namely (4/pi)^2 times [ 1 + 1/9 + 1/25 + 1/49 + ... ]. Setting the two equal forces 1 + 1/9 + 1/25 + ... = pi^2 / 8. Out of an energy balance for a square wave drops the exact value of a famous infinite sum — a number with no business being easy to get otherwise.
This is a genuine superpower of Parseval, and worth pausing on: it converts a hard infinite series you might never sum by elementary tricks into a routine integral you can. Pick a function, find its Fourier coefficients, write the energy balance, and you have evaluated a sum in closed form. The triangle wave gives you the sum of 1/n^4; other choices reach 1/n^6 and beyond. The whole zoo of these even-power sums (the values of the convergent p-series at even p, historically the Basel problem and its cousins) falls out this way. You did not invent a new summation technique; you reused the energy identity from a different angle.
- Pick a periodic f whose square integrates easily (often a wave that is piecewise constant or a simple polynomial like x or |x|).
- Compute its Fourier coefficients a_n, b_n once, the ordinary way.
- Write Parseval: (1/L) integral of f^2 = a_0^2/2 + sum (a_n^2 + b_n^2). The left side is a single easy integral.
- Solve for the unknown series on the right — it now equals a number you computed by integrating, in closed form.
Completeness: the honest fine print
Now for the caveat that textbooks often slide past. Orthogonality alone guarantees only an INEQUALITY — Bessel's inequality — that the harmonic energies sum to AT MOST the total energy: a_0^2/2 + sum(a_n^2 + b_n^2) is less than or equal to (1/L) integral of f^2. Geometrically that is obvious: projecting a vector onto some perpendicular axes can never capture more length than the vector has. Parseval upgrades the 'less than or equal' to an exact equality, and that upgrade is a separate, stronger claim. It says the trig system is not merely orthogonal but complete: there is no leftover piece of f hiding in some direction your axes fail to point. The energy you collect across all harmonics misses nothing.
There is also a practical lesson buried in the energy sum. Because the harmonic energies add up to a finite total, the terms a_n^2 + b_n^2 must shrink to zero as n grows — they cannot keep contributing forever. The smoother f is, the faster they decay (a continuous function's coefficients fall off faster than a jumpy one's), so most of the energy lives in the first few harmonics. That is the entire principle behind lossy compression: keep the loud low harmonics, throw away the faint high ones, and Parseval tells you in advance exactly how much energy — how much fidelity — you are discarding. The same accounting reappears for non-periodic signals as the Parseval-Plancherel theorem, where the harmonic sum becomes an integral over a continuous spectrum.
Why this is the load-bearing idea of the rung
Step back and see what you have. The whole Fourier method rests on one geometric fact — orthogonality — that lets you read each coefficient independently, and one accounting fact — Parseval — that lets you track energy without ever leaving the frequency domain. Together they say a signal and its spectrum are two faithful descriptions of the same object: change basis, lose nothing. This is exactly the same maneuver as choosing convenient perpendicular axes in ordinary geometry, only the 'axes' are now waves and the 'space' is infinite-dimensional.
And the idea does not stop at sines and cosines. The orthogonality-plus-completeness pattern is the template for everything ahead in this volume: when you solve the heat equation or a vibrating membrane, the natural building blocks are not always trig waves but other orthogonal families — Legendre polynomials, Bessel functions, the eigenfunctions of a boundary problem. Each comes with its own Parseval-style energy identity, and the abstract version, completeness of eigenfunctions, is what guarantees those expansions capture the whole function. Master the trig case as pure geometry here, and every later 'generalized Fourier series' is the same story told in a new alphabet.