Fourier analysis — flashcards

The sum of a function over a lattice equals the sum of its Fourier transform over the dual lattice. It is the bridge back from harmonic analysis to the Bernoulli and zeta material, and the reason the functional equation of ζ looks like a symmetry rather than a coincidence.

Fourier (on R, kernel e^{iξx}), Laplace (on [0,∞), kernel e^{st}), and Mellin (on (0,∞), kernel x^s) are not three unrelated formulas but the same construction — diagonalize translation by the relevant group — applied to translation, causal translation, and dilation respectively. The Mellin kernel's exponent is not a free convention: writing the multiplicative Haar measure on (0,∞) as dx/x forces the kernel x^{s-1}dx = x^s·(dx/x), a character (x^s) times the invariant measure, exactly parallel to e^{iξx}dx on the additive side. Evaluated at integer s, the Mellin transform hands back power sums of roots directly (moment order: Re(s) indexes power sums, Im(s) a log-frequency), feeding Newton's identities (st/newton-identities) as a bridge between the transform and elementary symmetric functions. Further members of the same family exist for other groups/geometries: the Hankel transform (radial Fourier, rotation group, Bessel functions), the Helgason–Fourier transform (hyperbolic space), and wavelets (simultaneous translation and dilation, when no single group suffices).

A kernel z^s=r^se^{is\theta} that diagonalizes both rotation (like Fourier) and scaling (like Mellin) at once fails for a structural reason, not a technical one: evaluating z^s at the same point reached by two different angular paths (θ vs θ+2π) gives values whose ratio is exactly e^{2πis}, confirmed numerically to differ sharply from 1 for non-integer s — single-valuedness on a loop enclosing the origin forces s∈Z. The practically-used compromise (the genuine Fourier–Mellin transform, used in image processing for scale/rotation-invariant matching) keeps the angular part discrete and only the radial part continuous. The full resolution removes the obstruction outright: a logarithmic spiral z(t)=e^{(a+ib)t} never closes on itself, so working on the universal cover (w=log z, unrestricted) makes a single spiral mode g(w)=e^{pw} transform (via a bilateral Laplace transform in w) to G(s)=1/(s−p) — one simple pole whose real part is growth rate and imaginary part is rotation rate, previously two separate quantities now encoded in one complex number. On the spiral, z·d/dz = d/dw exactly, and this single fact unifies the three classical transforms' apparently different derivative rules: the bilateral case gives clean multiplication by s, the one-sided case (spiral cut at a starting point z₀) gives a boundary term −f(z₀) — verified to 20 digits in both cases — so Fourier's clean rule, Laplace's f(0)-carrying rule, and Mellin's vanishing-boundary-term rule are one rule, distinguished only by whether the domain has an edge.

z^s satisfies Laplace's equation ∇²u=0 for any complex s, not just integers — verified numerically to finite-difference precision — as an automatic consequence of being holomorphic; the general principle (any PDE invariant under both dilation and rotation has z^s as a separable eigenfunction) has named physical instances: spiral vortex sheets in 2D fluid dynamics (self-similar Kelvin–Helmholtz roll-up), and unstable photon orbits near a black hole's photon sphere, where the winding rate is tied to the orbit's own Lyapunov exponent. Applying the Joukowski map to a logarithmic spiral instead of the unit circle produces a 'damped Chebyshev' identity, T_n(cosh(a)cos(b)+i·sinh(a)sin(b)) = cosh(na)cos(nb)+i·sinh(na)sin(nb) — verified numerically to 15 digits — generalizing T_n(cosθ)=cos(nθ) by giving both frequency and growth rate simultaneously, and identified as the actual mechanism underneath the Chebyshev/Fourier correspondence, requiring no trigonometric identity manipulation.

For a composite signal built from several spiral modes, G'/G is dominated by the nearest pole (verified numerically) — the same logarithmic-derivative-locates-structure principle used elsewhere on this map, applied here to pole location rather than zero location. Locating that pole via ordinary Newton's method on G itself fails outright (the iteration diverges), for an identifiable reason: poles are repelling fixed points of Newton's method applied to G, since Newton's method finds zeros, not poles. Applying Newton's method to 1/G instead converges cleanly, to 17 digits in four steps. A further structural obstruction: a sum of m poles produces m−1 extraneous zeros of G itself, contaminating any single large contour meant to extract all modes at once — small, pole-isolating contours recover exact values instead, though no general fix for the multi-pole case (beyond this practical workaround) was found.

The log-periodic power law (LPPL) model used for critical-transition precursors is identified as exactly a spiral mode of st/spiral-transform-unifies-fourier-mellin in w=log(t_c−t): parameters are recovered accurately from noisy synthetic data, but usable lead-time collapses sharply and non-gracefully as the available data before t_c shrinks, rather than degrading gradually. A claim about the model's robustness to false positives, made from a single test on structureless (random-walk) data returning an unremarkable frequency parameter, was corrected after a proper multi-realization test: across many independent random walks, the fit typically returns large frequency values that would read as genuine log-periodicity — the model hallucinates apparent structure on pure noise most of the time, the opposite of the original claim.

Ramanujan's Master Theorem: if f(x)=∑(−1)^k/k!·φ(k)·x^k, then the Mellin transform M[f](s)=Γ(s)φ(−s) — verified on f(x)=e^{−x} (φ≡1, recovering Γ(s) itself) and f(x)=1/(1+x) (φ(k)=k!, recovering Γ(s)Γ(1−s)), both matching to high precision. Carlson's theorem (a function of sufficiently controlled exponential growth that vanishes at every nonnegative integer must vanish identically) is the rigorous justification behind Ramanujan's own formal derivation. Why Γ enters by multiplication rather than division — unlike ordinary Taylor's theorem, which divides by k! to compensate for repeated differentiation — resolves once you notice the 1/k! is already part of the given data, not something the theorem introduces: a Mellin–Barnes contour shift shows Γ(−s)'s poles at every nonnegative integer k have residue exactly (−1)^(k+1)/k! (verified for k=0,…,4), so the entire infinite family of reciprocal factorials is already encoded in Γ's own residue structure, and residue-sum reconstruction of the original series is what recovers the 1/k! normalization — not a division anywhere.

The Hirschman–Beckner entropic uncertainty principle, H(|f|²)+H(|f̂|²) ≥ log(πe) with equality only for Gaussians (verified directly: both entropies computed for a Gaussian of arbitrary width, summing to exactly log(πe) independent of width), is the precise quantitative version of a tension recurring across this map between precise local knowledge and diffuse global knowledge. Three previously-established facts are each a separate instance of the same trade-off, not merely reminiscent of it: (1) a polynomial's roots-as-multiset and its coefficient list (Vieta's formulas, st/newton-identities) are a lossless bijection — zero information lost converting either way; (2) the argument principle's root count Z−P is a drastic, lossy compression from O(N) numbers to one integer, with each additional weighted contour moment restoring exactly one further degree of freedom until, after N moments, the information is recovered in full — a literal lossy-then-lossless compression channel; (3) a Taylor series' radius of convergence being controlled by the nearest singularity (st/borel-summation) is a statement that analytic functions are maximally compressible — an infinite-dimensional piece of purely local data (every derivative at one point) is exactly equivalent to the function's entire global behavior, with no redundancy either way.

Analytic continuation and the functional equation of Hecke L-functions, reproved as Fourier analysis on the adeles. The functional equation becomes Poisson summation.

st/tate-thesis reproves Hecke L-function analytic continuation via Fourier analysis on the adeles. At a single p-adic place, the local piece of that machinery — Tate's local zeta integral Z(f,s) = ∫_{Q_p*} f(x)|x|_p^s d*x — is exactly the p-adic analogue of the (archimedean) Mellin transform, using the multiplicative Haar measure d*x in place of dx/x. For f the indicator of Z_p, it sums to (1−p⁻¹)/(1−p^{-(s+1)}), the local Euler factor itself — so the Euler factor at p is literally a Mellin transform value, not just an analogue of one by loose resemblance.

st/tate-thesis reproves the Hecke L-function functional equation as Poisson summation on Q embedded diagonally in its adele ring A_Q — a consequence of the product formula ∏_v |x|_v = 1 over all places v (finite and infinite), which underlies the whole construction and was checked directly for a specific rational number. Carrying this local/global structure one step further: the log-derivative ξ'(s)/ξ(s) of the completed zeta function decomposes exactly into two pieces — an archimedean piece built from the digamma function (from the Gamma factor at the infinite place) plus a finite-place piece built from the von Mangoldt function (from the Euler product over primes) — verified numerically to 17 digits. The local zeta integral at a single prime p, Z(f,s) for f the indicator of Z_p, sums directly to the Euler factor 1/(1-p^{-s}); the global decomposition is what happens when every place's local computation is assembled at once under the product-formula constraint.

Siméon Denis Poisson, 1823.

John Tate, 1950.

A locally compact abelian group ↔ its group of characters, with Fourier transform the dictionary. Ordinary Fourier series is the case of the circle ↔ the integers.

Held, and closed a loop rather than opening a new one: Poisson summation is the bridge back from Fourier analysis to the Bernoulli/zeta material, and it's the reason the functional equation of ζ looks like a symmetry instead of a coincidence.

Decomposing functions into waves. Turns differentiation into multiplication and convolution into product.