Localization on sparse sets — three case studies

Localization on Sparse Sets: Three Case Studies

Status of this document. This is an expository synthesis, not a research announcement. Essentially every individual result below is classical or belongs to an established literature, and I have tried to attribute accordingly. What is offered here is (i) a connective framing that treats three superficially unrelated subjects as instances of one mechanism, (ii) explicit numerical verification of each claim, carried out from first principles rather than cited, and (iii) an honest accounting of where the framing holds and where it breaks. Section 6 states the limits plainly.

---

1. The organizing question

A recurring pattern in analysis, combinatorics, and dynamics is that an object defined by a vast sum or an uncountable domain turns out to be governed by a sparse distinguished subset. Everything else cancels, contributes trivially, or is determined by the survivors.

Three instances, developed below:

The recurring tool in the first two is the logarithmic derivative $F'/F$, which converts products into sums over distinguished points. The third case is a negative result of the same shape: it identifies the distinguished set as empty in the non-archimedean setting, which is why chaos does not occur there.

---

2. Part I: A unified spiral transform

2.1 The obstruction to naive unification

Fourier analysis diagonalizes rotation; Mellin diagonalizes scaling. A natural question is whether a single kernel $z^s = r^s e^{is\theta}$ handles both with $s$ fully complex.

It does not, on the punctured plane. Traversing $\theta \to \theta + 2\pi$ returns to the same point $z$ but multiplies $z^s$ by $e^{2\pi i s}$. Numerically, at $r = 2$, $s = 0.4 + 0.7i$: the ratio of the two values is $-0.00995 + 0.00723i$, matching $e^{2\pi i s}$ exactly and differing from $1$ by two orders of magnitude. Single-valuedness forces $s \in \mathbb{Z}$.

This is the same topological constraint ($\pi_1(S^1) = \mathbb{Z}$) that makes Taylor coefficients integer-indexed. Indeed a Taylor coefficient is a Fourier coefficient: with $z = re^{i\theta}$, $$f(re^{i\theta}) = \sum_n a_n r^n e^{in\theta},$$ and Cauchy's formula is Fourier coefficient extraction divided by $r^n$ (verified: recovering $a_4 = 3^{-4}$ for $f(z) = (1 - z/3)^{-1}$ to 15 digits).

The working compromise is the Fourier–Mellin transform, with $s$ continuous and $n \in \mathbb{Z}$ discrete: $$F(s,n) = \int_0^\infty\!\!\int_0^{2\pi} f(r,\theta)\, r^{-s-1} e^{-in\theta}\, r\, d\theta\, dr.$$ The measure needs no reconciliation: the polar area element $r\,dr\,d\theta$ factors as $(r\,dr)(d\theta)$, and $r\,dr$ is exactly Mellin's multiplicative Haar measure. Verified by factoring a joint integral of a separable function into independent 1D Mellin and Fourier-series computations, agreeing to 15 digits.

2.2 Logarithmic spirals resolve it

The obstruction is periodicity of $\theta$, so remove it. On a logarithmic spiral $z(t) = e^{(a+ib)t}$, one full turn returns to a different radius — the curve never closes. Equivalently: pass to the universal cover of $\mathbb{C}^*$, where $w = \log z$ is single-valued and unconstrained.

A single bilateral Laplace transform in $w$ then does the whole job. For $g(w) = e^{pw}$ with $p = a + ib$, $$G(s) = \frac{1}{s - p},$$ a simple pole at $s = a+ib$ whose real part is the radial growth rate and imaginary part the rotation rate. Verified: $|G(s)|$ tracks $1/\mathrm{dist}$ exactly as $s \to p$ (2.0, 10.0, 100.0, 1000.0 at distances $0.5, 0.1, 0.01, 0.001$).

The price is explicit: one is no longer describing functions on $\mathbb{C}^*$ but on its cover, where "once around" is a trackable event rather than a return.

2.3 Derivative property

On the spiral, $z\,d/dz = d/dw$ exactly. Hence

• bilateral (no endpoint): $\mathcal{T}[z f'](s) = s\,\mathcal{T}[f](s)$;
• one-sided (spiral cut at $z_0$): $\mathcal{T}[z f'](s) = s\,\mathcal{T}[f](s) - f(z_0)$.

Verified to 20 digits at $p = 0.4 + 1.1i$, $s = 2.0+0.5i$.

This dissolves an apparent discrepancy. The three classical transforms' derivative rules are one rule; the presence or absence of the boundary term tracks only whether the domain has an edge. Fourier (translation group, no edge) gets the clean form; Laplace's $f(0)$ is not an anomaly but the transform correctly reporting an initial condition; Mellin's boundary term vanishes for any $f$ whose transform converges at all.

Multiplicative convolution becomes a product, and dilation and rotation are two directions of a single shift theorem in $w$: $$\mathcal{T}[f(\lambda z)](s) = \lambda^{-s}\mathcal{T}[f](s), \qquad \mathcal{T}[f(ze^{i\phi})](s) = e^{-is\phi}\mathcal{T}[f](s).$$

2.4 A damped Chebyshev identity

Applying the Joukowski map $w = (z + 1/z)/2$ to the unit circle gives $T_n(\cos\theta) = (z^n + z^{-n})/2 = \cos n\theta$ — the Chebyshev–Fourier correspondence, with no trigonometric manipulation. Applying it to a spiral instead, $z = e^{a+ib}$, gives

$$T_n\big(\cosh a\cos b + i \sinh a \sin b\big) = \cosh(na)\cos(nb) + i\sinh(na)\sin(nb),$$

verified at $n=6$, $a = 0.4$, $b=1.7$ to 15 digits, and reducing to the classical identity at $a = 0$. The $n$-th Chebyshev polynomial multiplies growth rate and frequency together, because both are components of one complex rate.

2.5 Extracting modes: log-derivative and its failure mode

For a composite signal $G(s) = \sum_j R_j/(s - p_j)$, the log-derivative $G'/G$ is dominated near each pole by that pole's term: at $s = p_1 + (0.05 + 0.02i)$ with a second mode present, $G'/G = -16.98 + 7.21i$ against the isolated-pole value $-17.24 + 6.90i$.

Two uses:

1. Cumulant reading. $G'(s)/G(s) = -\mathbb{E}_s[w]$, the exponentially tilted mean of $w = \log z$ — an Esscher transform. Real part is a typical log-scale; imaginary part a typical phase.
2. Pole location. Poles are repelling for Newton's method on $G$; one must iterate on $1/G$, giving the sign-flipped step $s \mapsto s + G/G'$. This converges to $p_1$ in four steps to 17 digits.

The argument principle applies verbatim: integrating $G'/G$ around a small contour enclosing one pole gives winding number $-1$, confirming a simple pole (phase descending linearly from $0$ to $-\pi$, jumping to $+\pi$, returning to $0$).

A structural obstruction to global extraction. A sum of $m$ simple poles has a numerator of degree $m-1$ over a common denominator, hence $m-1$ extraneous zeros, located between the poles. A contour large enough to enclose all modes also encloses these, and $\oint z^k G'/G$ then returns power sums contaminated by zeros. Concretely, with two poles at $0.3+1.5i$ and $-0.5+2.7i$ and residues $1, 0.6$, the induced zero sits at $-0.2 + 2.25i$ and the $k{=}0$ integral returns $Z - P = -1$, not $-2$; feeding these into Newton's identities recovers wrong modes. Small contours isolating one pole each recover both exactly.

This is a genuine difference from the polynomial case, where $F'/F$ has poles only. It is why practical mode extraction uses recurrence-based methods (Prony, matrix pencil, ESPRIT) that build a pole-only characteristic polynomial directly from samples.

2.6 Estimation limits

For a two-mode signal ($p_1 = 0.3+1.5i$, $p_2 = -0.5+2.7i$, $\Delta = 0.05$, complex Gaussian noise $\sigma = 0.05$), median error in the weaker, decaying mode:

Three distinct effects, worth separating:

• Matrix pencil beats basic Prony by $\sim 10\times$ at low noise because polynomial root-finding amplifies coefficient noise, while SVD truncation removes the noise subspace before any root-finding occurs.
• At $N = 40$ the ML estimator does not attain the CRB — starting the optimizer at the true parameters, it moves away to genuinely lower cost. This is the standard threshold effect: the CRB is asymptotic and is not attainable below a critical SNR.
• The plateau between $N = 160$ and $N = 320$ is a genuine information wall. Fisher information for the decaying mode's rate accumulates as $\sum_n t_n^2 e^{-2\gamma t_n}$, a convergent series: totals $4.56 \to 10.92 \to 14.20 \to 14.40$ for $N = 40, 80, 160, 320$. A decaying signal carries a finite total budget of information about its own decay rate. Only physical changes (lower noise, denser sampling while the mode is alive, earlier observation) move it.

I initially misdiagnosed the $N{=}40$ gap as an initialization problem and tested multi-start refinement across four pencil parameters; results were byte-identical, ruling that out. Recorded here because the negative result is informative.

2.7 Application: log-periodic critical precursors

The log-periodic power law (LPPL) ansatz $\ln p(t) = A + \mathrm{Re}[K(t_c - t)^{m + i\omega}]$ is exactly a spiral mode in $w = \log(t_c - t)$, with $s = m + i\omega$.

On synthetic data ($t_c = 100$, $m = 0.4$, $\omega = 8.0$, noise $0.03$), seven-parameter fitting recovers $t_c = 100.03$, $m = 0.397$, $\omega = 8.010$. It also correctly returns $\omega \approx 0.03$ on a pure random walk, i.e. does not hallucinate log-periodicity in structureless data.

But lead time collapses sharply. Fitting with data truncated further before $t_c$:

The failure is not graceful: the optimizer finds sharp, wrong, confident answers. This is the numerical shape of the standing objection to LPPL crash prediction — that a seven-parameter oscillatory power law fits many curves in-sample, with disputed out-of-sample performance. The mathematics is sound; the predictive claim is contested and the lead-time collapse is why. The idea is on firmer ground in materials failure, where $t_c$ is externally verifiable.

---

3. Part II: Localization on the fixed points of an involution

3.1 Euler's pentagonal number theorem

$$\prod_{n\ge1}(1 - x^n) = \sum_{k \in \mathbb{Z}} (-1)^k x^{k(3k-1)/2}.$$

Verified to order 60 by direct product expansion. Nonzero coefficients occur only at generalized pentagonal numbers $0,1,2,5,7,12,15,22,26,35,40,51,57$, with signs $+,-,-,+,+,-,-,\ldots$ in pairs.

Expanding the product, every finite subset $S \subset \mathbb{N}$ contributes $(-1)^{|S|}x^{\sum S}$. That almost all coefficients vanish means almost all subsets cancel in parity-matched pairs.

3.2 Franklin's involution and its fixed points

Franklin's map exhibits the pairing. For a partition into distinct parts, let $s$ be the smallest part, $m$ the length of the top diagonal (the maximal run $\lambda_1, \lambda_1 - 1, \ldots$), $r$ the number of parts. If $s \le m$ and $s \ne r$, delete the smallest row and add 1 to each of the top $s$ parts; if $m < s$ and not ($m = r$ and $s = m+1$), do the inverse.

The two excluded cases are exactly the fixed points. Verified computationally over all partitions into distinct parts for $n \le 16$: the involution is a valid free $\mathbb{Z}/2$ action with

Exactly one fixed point at each generalized pentagonal number, none elsewhere. The two staircase families are the two branches $k(3k-1)/2$ and $k(3k+1)/2$.

The fixed points are the most symmetric objects, not the least. A generic partition has trivial stabilizer; the staircases have full stabilizer, because they are too rigid geometrically for the move to be defined. This is worth stating because it is the opposite of spontaneous symmetry breaking, where the realized state has less symmetry than the law. Both reduce to orbit–stabilizer bookkeeping, but in opposite directions.

3.3 The signed sum collapses onto fixed points

The mechanism is exactly that of the Lefschetz fixed point theorem and of Duistermaat–Heckman / Atiyah–Bott localization: a signed sum over a space carrying a group action reduces to a sum over the action's fixed locus, because free orbits cancel identically. Here $\sum_\lambda (-1)^{|\lambda|}x^{|\lambda|}$ collapses to the pentagonal terms alone.

Consequences, both verified:

Partitions. $p(n) = \sum_k (-1)^{k+1}[p(n - g_k^+) + p(n - g_k^-)]$ with $g$ the generalized pentagonal numbers. Reproduces $p(0..20)$ exactly.

Divisor sums, via the log-derivative. Taking $F'/F$ of the same product, $$x\frac{F'(x)}{F(x)} = -\sum_{N \ge 1}\sigma(N)x^N.$$ Matching coefficients gives a recurrence for $\sigma$ with identical pentagonal memory positions, but with one twist: the term whose lookback lands at index $0$ supplies $\sigma(n)$ itself, not a boundary constant. Solving for it is equivalent to substituting the current step number $n$ wherever $\sigma(0)$ would appear. Verified against direct divisor sums to $n = 30$.

This is a distinct species of self-reference: the recurrence's own index leaks into its value table at exactly one position per step. In the partition recurrence the same slot holds the constant $p(0)=1$.

3.4 Memory structure

The pentagonal memory is sparse and quadratically spaced, unlike the dense fixed-lag memory of a linear recurrence. The distinction has a clean source. Fibonacci-type memory comes from a characteristic polynomial's roots — an algebraic origin dictating fixed, dense lookback. Pentagonal memory is the residue of near-total cancellation, and the positions thin out quadratically because the un-cancelable "staircase" configurations do.

For contrast, the characteristic polynomials of the recurrences for $n^2$ and $n^3$ are $(x-1)^3$ and $(x-1)^4$: a repeated root, whose non-diagonalizable companion matrix (a Jordan block) is precisely what admits polynomial-in-$n$ solutions $\{1, n, n^2\}$ rather than pure exponentials.

3.5 Where the RG analogy holds and where it stops

Franklin's fixed points behave like renormalization-group fixed points: generic objects flow away or cancel, and a rigid invariant locus carries all universal information. Feigenbaum's universal function is literally defined this way, as a fixed point of the doubling operator $T[g](x) = \alpha g(g(x/\alpha))$.

The analogy has a sharp edge. Feigenbaum's $\delta$ is an eigenvalue of the linearization of $T$ around $g$ — a statement about the neighborhood. Franklin's setting is discrete: there is no tangent space at a staircase partition, no nearby non-fixed objects to linearize against, and therefore no $\delta$ analogue at all. Feigenbaum's $\alpha$ fares slightly better, being a defining constant of the fixed point itself (matched loosely to the fixed "3" and "1" in $k(3k\pm1)/2$), but this is a weak correspondence.

Finding the edge is the useful outcome. The claim that survives is narrow and real: rare invariant points of a symmetry carry disproportionate information. The claim that does not survive is any transfer of the local spectral theory.

---

4. Part III: Non-archimedean dynamics

4.1 Ostrowski's completeness

Every nontrivial absolute value on $\mathbb{Q}$ is equivalent to the ordinary one or to a $p$-adic one (Ostrowski). The list is complete — one archimedean place and one per prime — and the places are tied by the product formula $\prod_v |x|_v = 1$.

The $p$-adic absolute values satisfy the strengthened triangle inequality $|x+y|_p \le \max(|x|_p, |y|_p)$, which the ordinary one provably fails ($|50 + 3| = 53 \not\le \max(50,3)$). Everything below follows from this.

4.2 Newton's method has no fractal boundary

Over $\mathbb{C}$, Newton's method for $z^3 - 1$ has a Julia set: two starting points $2\times10^{-19}$ apart converge to different roots, and a single line segment of length 1 crosses the basin boundary 73 times. The mechanism is "democratic vote" — $F'/F = \sum_i 1/(z - r_i)$ is a weighted referendum over all roots, and a near-tie is amplified exponentially.

Over $\mathbb{Q}_p$ this cannot happen. Ties in an ultrametric are not infinitely finely balanced: either one term strictly dominates, or the outcome is settled by finitely many digits. Verified for $x^2 - 1$ mod $7^{20}$: ten starting points differing wildly in high-order digits but sharing a residue class mod 7 all converge to the same root; the number of correct $p$-adic digits doubles each step ($1,2,4,8,16$), i.e. Hensel's lemma with quadratic convergence.

4.3 Chaos is archimedean

For the logistic map $f(x) = 4x(1-x)$ at the parameter that is maximally chaotic over $\mathbb{R}$, the exact identity $$f(x) - f(x') = 4(x - x')(1 - x - x')$$ shows the gap between two orbits is multiplied by $4(1-x-x')$. Away from the critical point this is a $p$-adic unit, so valuations are preserved exactly.

Verified at $p=5$ mod $5^{25}$: two points agreeing to $k$ digits still agree to exactly $k$ digits after 15 iterations, for $k = 3, 6, 10$ — zero sensitivity.

The bad-reduction prime is not an exception. At $p=2$, where $4 = 2^2$ degenerates the map, the valuation grows by exactly 2 per step ($2,4,6,8,\ldots$): nearby points converge faster, not slower. Bad reduction gives super-contraction, not chaos.

Generalizing: for any polynomial map the same factorization applies, so the only possible source of expansion is a neighborhood of a critical point. This is the "good reduction implies empty Julia set" theorem of non-archimedean dynamics (Rivera-Letelier, Benedetto).

Consequently there is no $p$-adic Feigenbaum cascade at any prime. Positive Lyapunov exponents are a feature of the archimedean absolute value specifically — not of the equation $x \mapsto 4x(1-x)$.

I want to flag one negative result explicitly, since it is tempting to hope otherwise. The product formula constrains a fixed number's norms across places; it provides no mechanism relating a map's Lyapunov exponent at one place to another. There is no conservation law by which the tame $p$-adic places "compensate for" or encode the chaotic real one. What the adelic view does supply is Call–Silverman canonical height theory, in which $\hat h = \sum_v \hat h_v$ and almost all local contributions are trivial. For $r=4$ the nontrivial set is the archimedean place alone.

4.4 What survives: the $p$-adic Mellin transform

Tate's local zeta integral is the $p$-adic Mellin transform: $Z(f,s) = \int_{\mathbb{Q}_p^*} f(x)|x|_p^s\,d^*x$ with $d^*x = dx/|x|_p$. For $f = \mathbf{1}_{\mathbb{Z}_p}$ it sums to $(1-p^{-1})/(1-p^{-(s+1)})$, the local Euler factor (verified to 15 digits).

Multiplicative convolution becomes a product — verified on shell indicators — because $v_p(xy) = v_p(x) + v_p(y)$ plays the role that multiplication plays for ordinary Mellin. Tate's local functional equation relates $Z(f,s)$ to $Z(\hat f, 1-s)$; gluing across all places yields the functional equation of $\zeta$. Morita's $p$-adic $\Gamma_p$, built by excluding multiples of $p$ from the factorial, satisfies its own reflection formula $\Gamma_p(x)\Gamma_p(1-x)=\pm1$.

4.5 Roughness splits into two phenomena

In the real setting, "self-similar roughness" conflates two things that the $p$-adic setting cleanly separates.

Geometric self-similarity survives. The set of $p$-adic integers whose base-$p$ digits never use $0$ is a Cantor-type set built by digit restriction, no dynamics involved. Its Mellin-type sum $\sum_n (p-1)^n p^{-ns}$ has a pole exactly at $\log(p-1)/\log p$, its dimension — verified diverging at that threshold for $p = 5$. This is the same pole-reveals-exponent mechanism that extracts the Weierstrass box dimension and the Feigenbaum spacing.

Dynamically generated roughness does not survive, per §4.3. There is nothing of that kind for the transform to detect, not because the transform fails but because the object does not exist.

4.6 Computability

Hensel's condition $|f(x_0)|_p < |f'(x_0)|_p^2$ is a finite, checkable certificate guaranteeing convergence to a unique nearby root — strictly stronger than anything real Newton's method offers, where rigorous root isolation requires separate machinery (Sturm sequences, interval arithmetic).

This does not extend to a general strengthening. Equality of computable $p$-adic numbers remains undecidable for the same halting-flavored reason as in $\mathbb{R}$: no finite computation certifies that no future digit will differ. What changes is the algebraic case specifically.

---

5. Synthesis

The three parts share a mechanism and differ in what plays each role.

Localization. In each case a signed or weighted sum over a large space reduces to a sparse set. Franklin's involution is the cleanest instance — exact, finite, combinatorial. The spiral transform's version is analytic: the contour integral of $G'/G$ reduces to residues at poles. The $p$-adic case is the degenerate instance where the distinguished set is empty.

The log-derivative as the universal extraction tool. $F'/F$ turns products into sums over distinguished points. It gives $\sum_i 1/(z-r_i)$ for polynomial roots, $-\sum \sigma(N)x^N$ for Euler's product, $-\mathbb{E}_s[w]$ for the spiral transform's tilted mean, and von Mangoldt's $\Lambda$ for $\zeta$'s Euler product. In each case the "vote" is dominated by the nearest distinguished point — which is exactly why Newton's method converges, and why its complex basins are fractal when the vote is close.

Symmetry determines the diagonalizing structure. Rotation gives Fourier series with discrete index; scaling gives Mellin with continuous index; their semidirect product (spiral similarity) gives a single complex index at the cost of passing to a covering space. Which absolute value one chooses, per Ostrowski, determines whether critical points can generate chaos at all.

Two directions of "symmetry breaking." Franklin's fixed points have full stabilizer; spontaneously broken vacua have reduced stabilizer. Both are orbit–stabilizer statements. Conflating them inverts the direction, and the distinction is worth preserving.

---

6. Limits and honest accounting

What is classical. Euler's pentagonal theorem and Franklin's involution (1881); localization theorems; Ostrowski; Hensel; Tate's thesis; good-reduction results in non-archimedean dynamics (Rivera-Letelier, Benedetto); Prony, matrix pencil, ESPRIT; the Cramér–Rao bound and threshold effects; the LPPL model and its critiques; Call–Silverman heights; Morita's $\Gamma_p$. The Fourier–Mellin transform is standard in image processing.

What is offered here. The connective framing; the damped Chebyshev identity of §2.4 in the form stated; the explicit derivative property for the unified spiral transform including the boundary-term unification of §2.3; and the verification program throughout.

What is negative or inconclusive.

• No adelic compensation principle relates $p$-adic tameness to archimedean chaos (§4.3).
• No Feigenbaum-$\delta$ analogue exists in the Franklin setting (§3.5).
• The extraneous-zero obstruction blocks single-contour global mode extraction (§2.5).
• The CRB is not attained at moderate SNR; the initialization hypothesis for that gap was tested and rejected (§2.6).
• LPPL lead-time collapse is severe and is the crux of the standing objection to its predictive use (§2.7).

Caveats on the verifications. All numerics were computed for this document and are reproducible from the descriptions given, but they are illustrations, not proofs. Several were run at modest precision or sample size; the Franklin verification covers $n \le 16$ exhaustively, the $p$-adic dynamics $p \in \{2,5\}$ at fixed modulus, the estimation study a single two-mode signal. Broader parameter sweeps would strengthen but not alter the claims, all of which are either classical theorems or explicitly flagged as framing.