DEMYSTIFIED

TERRY A. LORING

1. You’ve convolved before Convolution is made mysterious, but it comes up in a familiar place: polynomial multiplication. 2 3 2 Consider f(x) = a0 + a1x + a2x + a3x and g(x) = b0 + b1x + b2x . The annoying way to compute and fg, which is a polynomial, and to find all its coefficients, is like this:

[fg](x) = f(x)g(x) ¡ 2 3¢ ¡ 2¢ = a0 + a1x + a2x + a3x b0 + b1x + b2x ¡ 2¢ = a0 b0 + b1x + b2x ¡ 2¢ + a1 b0 + b1x + b2x ¡ 2¢ + a2 b0 + b1x + b2x ¡ 2¢ + a3 b0 + b1x + b2x 2 5 = a0b0 + a0b1x + a0b2x + ... + a3b2x which is past my pain threshold. You’ve all noticed long ago that the way a degree-n monomial is created is when a degree-r and a degree-s monomial come together and r + s = n. The coefficient of xn is thus

Date: September 5, 2003. X arbs. Thus the quadratic term in fg is (a0b2 + a1b1 + a2b0) . Indeed r+s=n

[fg](x) = (a0b0)

= (a0b1 + a1b0) x 2 = (a0b2 + a1b1 + a2b0) x 3 = (a1b2 + a2b1 + a3b0) x 4 = (a2b2 + a3b1) x 5 = (a3b2) x

X The meaning of a sum like arbs is to create all possible pairs r+s=n (r, s) add these all together. There are only the order in a finite sum of reals in irrelevant, and yet most people feel better if an order is forced. Thus such a sum is usually rewritten to imply an order: X Xn arbs = arbn−r r+s=n r=0 We’re almost to convolution. Next we pad the polynomials f and g with zeros. For j ≤ −1 and for j ≥ 4, aj = 0. For k ≤ −1 and k ≥ 3 set bk = 0. Then (for all real values of x, except perhaps zero) we have X∞ j f(x) = ajx j=−∞ and X∞ k g(x) = bkx k=−∞ (You can ignore ² and δ for these sums. Just toss the terms that equal zero and sum the finitely many terms that remain.) What’s more, we see that X∞ n [fg](x) = cnx n=−∞ where, for all n, X∞ cn = arbn−r r=−∞ 2 2. Convolving sequences We’ll consider all types of bi-infinite sequences later. Here’s a good place to start:

Definition 1. A doubly-infinite sequence x = hxni (index set the inte- gers), has finite support is there are integers C1 and C2 such that

(n < C1 or n > C2) ⇒ xn = 0. (Equivalently, all but finitely many of the terms in the sequence are zero.) A good convention with a doubly infinite sequence is to mark the term of index zero with an underline. Thus z = h... 0, 0, −1, 2, 1, 0, 0,...i is the sequence that has all terms zero except

z−1 = −1, z0 = 2, z1 = 1. To give x a playmate, we let w = h... 0, 0, 1, 0, 1, 0, 0,...i. Definition 2. Suppose x and y are double-infinite sequences of finite support. Then the convolution x∗y is again a double-infinite sequence, where X∞ [x ∗ y]n = xryn−r r=−∞ So, for example, z ∗ w = h... 0, −1, 2, 0, 2, 1, 0,...i. It is no coincidence that the following if true for all real x 6= 0 : ¡ ¢ ¡ ¢ −x−1 + 2 + x x−1 + 0 + x = −x−2 + 2x−1 + 0 + 2x + x2

3. A homorphism Mathematicians are always excited when one operations is turned into another by a . For example, the log function is beloved precisely because of what it does to products: ln(ab) = ln(a) + ln(b) for positive a and b. 3 A pure mathematician is excited because the structure of the product operation can be explained via the structure of . An applied mathematician gets excited an invents the sliderule. All the variations on the perform a similar trick, converting regular multiplication (pointwise multiplication) of functions into convolution, and vice-versa. We found something very close to the Fourier transform above. Consider the set P of all functions from (−∞, 0] ∪ [0, ∞) to(−∞, ∞) that are polynomials in x and x−1. Let S denote the set of all doubly infinite real sequences of finite support. We make the following simple definition: If −2 −1 1 2 f(x) = ... + a−2x + a−1x + a0 + a1x + a2x ... then Φ(f) = ha−2, a−1, a0, a1, a2, i. (It is simple once you look past the notation. It just strips off the coefficients.) Here’s the type of formula that creates all the fuss: (†) Φ(fg) = Φ(f) ∗ Φ(g). We would rather have the reverse formula, that explains the convolu- tion of sequences in terms of the pointwise product of functions. That is not really possible if we stick with real variables. But for a pure mathematician, (†) is inspiration enough to pursue this subject.

4. Differentiation made “easy” We will see that the Fourier Transform for R takes as input certain functions F : R → R and returns a new function Φ(f): R → R. Of the many formulas that hold, here’s one: Φ(f 0)(x) = xΦ(f)(x) In words, the Fourier transform converts differentiation into multipli- cation by x. (Multiplying by x is easier than taking derivatives, hence “differen- tiation made easy.” Of course, the Fourier Transform is harder than differentiation, so this is a bit of a mirage.) This formula opens lots of doors. Using the inverse of the Fourier Transform, Φ−1, we can get this: Dnf(x) = Φ−1(xnΦ(f)(x)) 4 for n = 1, 2,... Now, just because we can, we put in a non-integer for n and call this a definition. In particular, for n = 1/2, we define D1/2(f) = Φ−1(x1/2Φ(f)(x)). We now have the order “1/2” derivative of (some) functions! Study- ing this is outside the scope of this class. I just present this to you to give you more of a sense of why the Fourier Transform is so essential in pure math.

5