6fbd3ac Δ f6bf760 Δ 56bc728 Δ b1f9bc4 Δ b34550f Δ 3ff9228 Δ 49c2b3a
editors: nobody
# The universal optimality of the triangular lattice

*An expository walkthrough, written to be readable with high-school mathematics, while exactly stating every theorem and every step of the proof.*

---


- For a one-page summary: Section 12.6.
- For a map of every fact invoked: Appendix C.

## Contents

- **0. Prologue** -- the paper in one sentence.
- **1. Cast of characters** -- lattice, configurations, kernels, energy.
- **2. Main theorem** -- universal energy minimality; history; quantifiers; Riesz and log corollary.
- **3. Strategy** -- turn energy into Fourier signs; sharp pairs; roadmap.
- **4. Stage A: candidate functions** -- sine product, cardinal functions, Fourier transform, jet operator.
- **5. Stage B: finite data** -- truncation, Fejer quadrature, mass matrix, power sums, tables.
- **6. Stage C: exact interpolation** -- operator equation, Schur complement, correction bound.
- **7. Signs** -- deleted-slack quotient, Bernstein rows, quadratic barrier.
- **8. Gaussian to energy** -- density LP, sharp pairs, Poisson summation.
- **9. Gaussian to universal** -- positive mixtures, singular kernels, shifted potentials.
- **10. Renormalized Riesz and jellium** -- field energy, heat comparison, screening, thermodynamic limit.
- **11. Arithmetic verification** -- rational enclosures, scalar budgets, floor-rounding, Arb scripts.
- **12. Epilogue** -- summary, quantifier subtlety, open questions.
- **Appendix A** -- where the constants come from.
- **Appendix B** -- floor-rounded algorithm reference.
- **Appendix C** -- statement of every hypothesis.

---

## 0. Prologue: the paper in one sentence

Among all locally finite point sets in the plane with average density one, the triangular (hexagonal) lattice is the minimizer of average pairwise energy per point for every nonnegative completely monotone function of squared distance -- including when the energy is infinite. The same lattice also minimizes the logarithmic and Riesz renormalized energies with 0 le s < 2, and the corresponding jellium problems.

If that was too dense, here is a friendlier version. Pick a favorite pairwise interaction. Some examples:

- **Inverse power**: g(t) = t^{-p}, p > 0. Points close together repel strongly.
- **Gaussian**: g(t) = e^{-pi alpha t}, alpha > 0. Repulsion dies off quickly.
- **Logarithm**: g(t) = -(1)/(2) log t. The planar Coulomb potential, singular at t = 0 and slowly growing at infinity.
- **Riesz**: g(t) = t^{-s/2}, 0 < s < 2. A power law interpolating between Coulomb-like and slower decay.

Now consider all ways to scatter identical points in the plane at one point per unit area. The paper proves that the triangular lattice minimizes the average pairwise energy per point for every smooth completely monotone kernel (all inverse powers, all Gaussians, and any positive mixture of these). It also proves the analogous statement for the log and Riesz kernels, which are too singular to compare directly and require a standard neutralization ("renormalization") step.

The word "universal" means: one configuration minimizes the whole family at once, not one kernel at a time. The Cohn-Kumar conjecture (2007) predicted this for the triangular lattice, the E_8 lattice in dimension eight, and the Leech lattice in dimension twenty-four. This paper settles the planar case.

---

## 1. The cast of characters

### 1.1 The triangular lattice

Let b = sqrt 3 / 2. The triangular lattice A of unit covolume is

A = b^{-1/2}bigl{ j,(1,0) + k,(1/2, b) : j, k in Z bigr}.

The prefactor b^{-1/2} normalizes area: the parallelogram spanned by b^{-1/2}(1,0) and b^{-1/2}(1/2, b) has area exactly one. So A is the triangular lattice at density one.

Two constants appear everywhere:

b = (sqrt 3)/(2) = 0.8660254...,  B = frac{1}{b^{2}} = (4)/(3).

B is the "area scale" that will reappear in every Gaussians Fourier-transform computation.

The **dual lattice** A^{*} is the set of all w in R^{2} such that w · a in Z for every a in A. Since the basis of A has determinant one, a quarter-turn J gives A^{*} = JA. Because A is rotation-invariant, A and A^{*} have the same set of lvert a rvert^{2} values -- this will let us make contacts on the direct and dual side simultaneously.

### 1.2 Configurations, density, and disks

A **configuration** is a locally finite subset C subset R^{2}: infinite, but with only finitely many points in any bounded region.

For R > 0, let B_{R} be the closed disk of radius R centered at the origin, C_{R} = C cap B_{R}, and N_{R} = # C_{R}.

The configuration C has **centered disk density one** if

frac{N_{R}}{pi R^{2}} → 1  (R to ∞).

This is *weaker* than periodicity. Points can cluster locally, so long as the global average count per unit area is one, and there is no separation assumption at all.

### 1.3 Completely monotone kernels

Let g : (0, ∞) to [0, ∞) be smooth. Say that g is **completely monotone** if

(-1)^{r} g^{(r)}(t) ge 0  for every integer  r ge 0  and every  t > 0.

The derivatives alternate in sign, and every (-1)^{r} g^{(r)} is nonnegative.

Examples:

- **Inverse powers**: g(t) = t^{-p}. Its rth derivative is (-1)^{r} p(p+1) ... (p+r-1), t^{-p-r}, and (-1)^{r} g^{(r)}(t) = p(p+1) ... (p+r-1), t^{-p-r} > 0.
- **Gaussians**: g(t) = e^{-alpha t}. Its rth derivative is (-alpha)^{r} e^{-alpha t}, so (-1)^{r} g^{(r)}(t) = alpha^{r} e^{-alpha t} ge 0.
- **Products**: t^{-p} e^{-t} has the same sign pattern, by the Leibniz rule applied to nonnegative-magnitude derivatives.
- **Positive mixtures**: if each g_{alpha} is completely monotone and mu is a positive measure, then int g_{alpha}(t) , dmu(alpha) is completely monotone. This closure under mixtures is the engine of Section 9.

The classical **Hausdorff-Bernstein-Widder theorem** (proved in shifted form in Section 9.1) says the above is *equivalent* to the representation

g(t) = int_{[0,1]} v^{t} , drho(v)

for some positive Borel measure rho on [0, 1]. The change of variables v = e^{-alpha} turns this into a Laplace transform:

g(t) = int_{0}^{∞} e^{-alpha t} , dtilderho(alpha).

For g = t^{-p}, the representing measure has infinite mass at alpha = 0: the decomposition is valid, but the total mass is not finite. This is exactly why the paper will *shift* the potential before mixing (Section 9.2).

### 1.4 The energy of a configuration

Fix a nonnegative kernel g : (0, ∞) to [0, ∞). The **average ordered-pair energy** at radius R is

E_{g}(C_{R}) ;=; frac{1}{N_{R}} sum_{substack{x, y in C_{R}  x ne y}} g(lvert x - y rvert^{2}).

The sum counts *ordered* pairs: (x, y) and (y, x) are separate terms; x = y is excluded.

The full energy is

E_{g}(C) ;=; liminf_{R to ∞} E_{g}(C_{R}) in [0, ∞].

We use liminf rather than lim because E_{g}(C_{R}) may oscillate with R; the lower limit captures the asymptotic "worst-case-lower" energy. Nonnegativity of g ensures E_{g}(C) is well-defined in the extended reals.

The lattice has a clean value:

E_{g}(A) = sum_{a in A ∖ {0}} g(lvert a rvert^{2}).

For g = t^{-p} with p ge 1 this sum is infinite, which is fine: the theorem's conclusion is that every competitor is also infinite.

---

## 2. The main theorem: universal energy minimality

**Theorem (Universal energy minimality).** *Let g : (0, ∞) to [0, ∞) be smooth and completely monotone. For every locally finite configuration C subset R^{2} of centered disk density one,*

E_{g}(C) ;ge; E_{g}(A) ;=; sum_{a in A ∖ {0}} g(lvert a rvert^{2}),

*where both sides are interpreted in the extended nonnegative reals [0, ∞].*

Three things stand out.

- **Universal.** One configuration A witnesses the inequality for the whole family of completely monotone g at once.
- **Against everything.** The inequality is not restricted to lattices, or to periodic configurations, or to bounded density fluctuations; it applies to any locally finite C with centered disk density one.
- **Infinite energies allowed.** When the right-hand side is infinite (e.g. g = t^{-1}), the inequality says E_{g}(C) = ∞ for every competitor. The formulation is meaningful throughout.

The theorem immediately covers every inverse power g(t) = t^{-p}, p>0, and every Gaussian g(t) = e^{-alpha t}, alpha > 0. It covers each positive mixture of these by linearity of the representing measure.

### 2.2 The subtlety about quantifiers

The Cohn-Kumar conjecture asked for a single auxiliary function f, allowed to depend on the kernel g, with certain properties (see Section 3.3). This is strictly stronger than what the present paper establishes.

The present paper constructs a sharp auxiliary function f_{alpha} for *each Gaussian scale* alpha > 0. For a general mixed potential (e.g. t^{-1/2}), it does not construct a single auxiliary function; it only proves the *energy* bound by integrating the Gaussian-scale bounds over the positive-mixture representation. The paper states this honestly: the stronger per-potential quantifier of Cohn-Kumar's original conjectured form is not claimed.

The energy conclusion (the theorem) is fully established regardless.

### 2.3 Corollary: Riesz and log

For 0 < s < 2, the raw sum sum_{a ne 0} lvert a rvert^{-s} diverges. So the "energy comparison" above is trivially ∞ ge ∞ and carries no information. Physically this is because an infinite configuration of all-same-sign charges has infinite energy: standard Coulomb physics neutralizes the charges against a uniform oppositely charged background.

Two renormalizations are standard.

- **Periodic.** Fix a torus T = R^{2} / L. Neutralize the points against a uniform background of opposite total charge on T. Remove each point's self-energy. The result is a finite periodic energy W_{s, L}.
- **Field.** Build the infinite "jellium" configuration: point charges plus uniform background. Introduce a self-energy cutoff eta > 0, remove it, and take the field-energy per unit area. Minimize over compatible gradient fields. The result is the infinite-system energy W_{s}.

The paper proves the following corollary (Section 10.4).

**Corollary (planar renormalized and jellium energies).** *Let 0 le s < 2, with interaction g(x) = -log lvert x rvert at s = 0 and g(x) = lvert x rvert^{-s} for 0 < s < 2.*

*(i) For each n ge 1, the configuration A/nA (i.e. the n^{2} distinct classes of the triangular lattice in the torus R^{2}/(nA)) minimizes W_{s, nA} among all configurations of n^{2} distinct points in that torus.*

*(ii) The infinite-system energy W_{s}(A) minimizes W_{s} over all density-one planar configurations. In addition, the ordered-pair thermodynamic jellium minimum equals W_{s}(A)/kappa_{s}, where kappa_{0} = 2pi and kappa_{s} = 4pi for 0 < s < 2.*

The proof routes through the Gaussian comparison proved in the main theorem, by way of a "heat-kernel representation" (Petrache-Serfaty, 2020) that turns the renormalized energy into an integral of Gaussian energies over a heat-time parameter (Section 10.2).

---

## 3. The road: from energies to Fourier signs

The theorem has a simple shape after Fourier transformation. We describe the shape now; every piece gets its own section afterward.

### 3.1 Fourier transform and nonnegativity

For a Schwartz function f : R^{2} to R (smooth, rapidly decaying, and so is every derivative), define

widehatwidehat f(f(xi) ;=; int_{R^{2}} f(x), e^{-2pi i, x · xi}, dx.

For finite signed measures mu, nu on R^{2} set

Q_{f}(mu, nu) = iint f(x - y), dmu(x), dnu(y).

Fourier inversion and Fubini give

Q_{f}(mu, nu) = int_{R^{2}} widehatwidehat f(f(xi), widehatmu(xi)bar, widehatnu(xi), dxi.

If widehatwidehat ff ge 0 pointwise, then Q_{f} is positive semi-definite: Q_{f}(mu, mu) ge 0 for every finite measure mu. So Cauchy-Schwarz applies to Q_{f}.

For a configuration C let mu_{R} = sum_{x in C_{R}} delta_{x}. Then Q_{f}(mu_{R}, mu_{R}) = sum_{x, y in C_{R}} f(x - y). The diagonal terms contribute N_{R} · f(0), and the off-diagonal terms are the pairwise energy. Hence

sum_{substack{x, y in C_{R}  x ne y}} f(x - y) = Q_{f}(mu_{R}, mu_{R}) - N_{R}, f(0).

When widehatwidehat ff ge 0 we get Q_{f}(mu_{R}, mu_{R}) ge 0 and therefore

sum_{substack{x, y in C_{R}  x ne y}} f(x - y) ge -N_{R}, f(0).

Dividing by N_{R} and taking the lower limit gives the trivial lower bound

liminf_{R to ∞} frac{1}{N_{R}} sum_{substack{x, y in C_{R}  x ne y}} f(x - y) ge -f(0).

This is non-vacuous when widehatwidehat f(0)f(0) - f(0) > 0, but it neglects density. Density one sharpens the bound below.

### 3.2 The Cohn-Zhao linear program

Density one enters by comparing mu_{R} to the area measure nu_{R} = mathbf 1_bf1_{B_{(1+ε)R}} (uniform area measure with total V_{R} = pi (1+ε)^{2} R^{2}). This trick is due to Cohn-Zhao (2014) and was made explicit for packings by Cohn-de Courcy-Ireland (2018).

Fix ε > 0. For every x in B_{R} the disk B_{(1+ε)R} - x contains B_{ε R}. So

Q_{f}(mu_{R}, nu_{R}) = N_{R}, widehatwidehat f(0)f(0) + O!(N_{R} · int_{lvert z rvert ge ε R} lvert f(z) rvert, dz),

and the error term is *uniform* in the positions of the points of C_{R}: clustering does not help or hurt. Simultaneously Q_{f}(nu_{R}, nu_{R}) / V_{R} to widehatwidehat f(0).f(0).

Cauchy-Schwarz gives Q_{f}(mu_{R}, mu_{R}), Q_{f}(nu_{R}, nu_{R}) ge lvert Q_{f}(mu_{R}, nu_{R}) rvert^{2}. Divide by N_{R}^{2}:

frac{Q_{f}(mu_{R}, mu_{R})}{N_{R}} · frac{V_{R}}{N_{R}} · frac{Q_{f}(nu_{R}, nu_{R})}{V_{R}} ge ( frac{Q_{f}(mu_{R}, nu_{R})}{N_{R}} )^{2}.

As R to ∞: V_{R} / N_{R} to (1+ε)^{2}, and the two ends converge to widehatwidehat f(0)^f(0)^{2} each. Taking lower limits in R then ε ↓ 0 yields

liminf_{R to ∞} frac{Q_{f}(mu_{R}, mu_{R})}{N_{R}} ge widehatwidehat f(0).f(0).

Removing the N_{R} diagonal terms gives the **density LP inequality**:

liminf_{R to ∞} frac{1}{N_{R}} sum_{substack{x, y in C_{R}  x ne y}} f(x - y) ;ge; widehatwidehat f(0)f(0) - f(0).

If f le g pointwise off the origin and widehatwidehat ff ge 0, then the same argument runs with f in place of g inside the sum (termwise domination), giving

E_{g}(C) ;ge; widehatwidehat f(0)f(0) - f(0).

**Key message.** A *minorant* f le g with nonnegative Fourier transform produces a lower bound on the average g-energy.

### 3.3 The Gaussian "sharp pair"

The main theorem will be proved in the Gaussian case, g(t) = e^{-pi alpha t}, by constructing a suitable minorant at each alpha. Define G_{alpha}(x) = e^{-pi alpha lvert x rvert^{2}}. We want a real radial Schwartz function f_{alpha} with

f_{alpha} le G_{alpha},  widehatwidehat f_f_{alpha} ge 0,  (two sign conditions)

and two families of *contacts*:

f_{alpha}(a) = G_{alpha}(a)  for every  a in A ∖ {0},  (direct contacts)

widehatwidehat f_f_{alpha}(w) = 0  for every  w in A^{*} ∖ {0}.  (dual contacts)

**Why the contacts produce the sharp estimate.**

Poisson summation for the lattice A says, for any decaying f,

sum_{a in A} f(a) = sum_{w in A^{*}} widehatwidehat f(w).f(w).

Split off the origin of each side:

f(0) + sum_{a ne 0} f(a) = widehatwidehat f(0)f(0) + sum_{w ne 0} widehatwidehat f(w).f(w).

With f = f_{alpha}, the dual contacts give sum_{w ne 0} widehatwidehat f_f_{alpha}(w) = 0, so

sum_{a ne 0} f_{alpha}(a) = widehatwidehat f_f_{alpha}(0) - f_{alpha}(0).

The direct contacts set f_{alpha}(a) = G_{alpha}(a) at every nonzero lattice point, so

sum_{a ne 0} G_{alpha}(a) = widehatwidehat f_f_{alpha}(0) - f_{alpha}(0).

Combining this identity with the density LP inequality for f_{alpha} against G_{alpha} yields

E_{e^{-pialpha t}}(C) ge widehatwidehat{f_f_{alpha}}(0) - f_{alpha}(0) = sum_{a ne 0} G_{alpha}(a) = E_{e^{-pialpha t}}(A).

That is the Gaussian case of the main theorem. The remaining sections construct the pair (f_{alpha}, widehatwidehat f_f_{alpha}) for every alpha > 0.

### 3.4 The interpolation obstruction

The two sign conditions and the two contact conditions must hold *simultaneously*, *at every scale* alpha, and at every radius.

- Contact conditions are *interpolation* conditions. There are infinitely many of them, one at every nonzero lattice shell.
- Sign conditions are *continuous* everywhere.

Neither condition alone is hard. Simultaneously, they multiply the difficulty.

A further subtlety: values and first derivatives of f and widehatwidehat ff on the actual shells of A do *not* suffice to determine a radial Schwartz function. Sardari (2021) proved that infinitely many linearly independent radial Schwartz functions have all those data zero. So the interpolation problem must be enlarged: we work with a superset of shells (the periodic node set N of Section 4.1), and impose contacts at every node of the superset. In a summable coefficient space this *does* determine the function uniquely (Section 6).

### 3.5 The three-stage strategy

The construction splits into three stages, each handled in its own section.

**Stage A (Section 4).** Set up the *machinery*: express every admissible interpolation datum (value and first derivative at every node) as an entire function of a complex variable, Fourier-transform it, and prove the resulting Schwartz function pair has explicit, uniform bounds. In this stage there is no "interpolation" yet; we just write down the *form* of every candidate.

**Stage B (Sections 5-6).** Solve the *interpolation problem* exactly. Turn the coupled set of "value-and-slope at every node" equations into a linear operator equation in a summable space. Approximate it by an explicit finite matrix D (Stage B1, Section 5), invert it there, and solve the remaining infinite tail by a Schur-complement argument (Stage B2, Section 6). The crucial small quantity is that the operator's outputs at far nodes are *tiny*: interpolation with distant inputs never significantly reinforces a given node.

**Stage C (Section 7).** Prove the *two signs*. This turns the contact data from Stage B into a genuine pair of upper-lower bounds. The strategy is:

1. Subtract value and first derivative at every node, so the residues are quotient functions with bounded behavior near nodes.
2. Bound the quotient from below on each *half-gap* between adjacent nodes using *Bernstein polynomial* coefficient estimates (Section 5.3 provides the input, Section 7.2 applies it).
3. Bound the quotient from below on the *unbounded tail* using a *quadratic barrier* for the sine product (Section 7.4).

The final sections (8-10) convert the Gaussian pair into the universal mixture inequality and then into the renormalized/jellium corollary. Two appendices supply the numerical inequalities and the alternative rational verification procedure used in Stages B and C.

The next sections build Stage A concretely.

---

## 4. Stage A: building entire candidate functions

We now construct the entire infinite family of "candidate" functions that the interpolation problem of Sections 5-6 will pin down to a specific pair. These candidates are of the shape "sine product times a rational sum." Their point is: each candidate has control over value and first derivative at every node of the periodic set N, and the map from "dataset" to "function" is *linear* and *summable*.

### 4.1 Shells, nodes, and the periodic set N

The construction uses four numerical parameters:

b = (sqrt 3)/(2),  h = (2)/(5),  B = (4)/(3) = frac{1}{b^{2}},  s = b lvert x rvert^{2}.

s is the "rescaled squared radius" used throughout Sections 4-7. h = 2/5 is a Gaussian damping exponent chosen to make numerical bounds tight; other values would work but produce worse constants.

Take the basis

u = sqrt{2/√(3)};(1, 0),  v = sqrt{2/√(3)};(1/2, ,b).

The determinant of (u, v) is one, so this is a unit-covolume lattice -- exactly the triangular lattice A from Section 1.1.

For any j, ell in Z:bbZ:

b lvert j u + ell v rvert^{2} = j^{2} + j ell + ell^{2}.

This quadratic form takes the values 0, 1, 3, 4, 7, 9, 12, 13, 16, 21, ... -- the *shells* of A.

Two local computations pin down these shells modulo 12.

- **Modulo 3.** j^{2} + j ell + ell^{2} equiv (j - ell)^{2} pmod 3, so its residue is 0 or 1 in {0, 1, 2}.
- **Modulo 2.** j^{2} + j ell + ell^{2} is even iff both j and ell are even, in which case it is divisible by 4.

The two restrictions, combined, force

j^{2} + j ell + ell^{2} in {,0, 1, 3, 4, 7, 9,} pmod{12}.

Define

L = {0, 1, 3, 4, 7, 9},  Z = 12Z + L,  N = Z cap (0, ∞),  N_{0} = {0} cup N.

**Claim.** Every nonzero shell of A lies in N.

**Claim (strictness).** The inclusion is proper: for example 24 in N, but 24 is not a shell of A. Indeed, if j^{2} + jell + ell^{2} = 24, parity forces j, ell even, so (j/2)^{2} + (j/2)(ell/2) + (ell/2)^{2} = 6; parities force j/2, ell/2 even too, so the original value would be divisible by 16, contradicting 16 nmid 24.

The extra nodes (like 24) are *artificial*: they are not shell radii of A. But they will also carry interpolation data. This is exactly the enlargement that the Sardari non-uniqueness theorem requires (Section 3.4).

Finally, recall the dual lattice A^{*} = JA from Section 1.1 --- the same shells. So contacts at the nodes of N serve *both* the direct (A) side and the dual (A^{*}) side simultaneously.

### 4.2 The normalized Gaussian pair

The construction will produce *two* entire functions H_{1}, H_{2} and a single radial Schwartz f. Everything is normalized to the parameter k > 0.

For each k ge 2.36 define

T_{k}(s) = k^{-1} e^{-k(s - 1)}.

T_{k} is a stretched exponential "reference profile," scaled so T_{k}(1) = 1/k and decaying away from s = 1.

**Theorem (Normalized Gaussian Fourier pair).** *For every finite real k ge 2.36, there exist entire functions H_{1}, H_{2} real on the real axis, and a real radial Schwartz function f on R^{2}, with*

f(x) = e^{-pi h b lvert x rvert^{2}}, H_{1}(b lvert x rvert^{2}),  widehatwidehat f(f(xi) = e^{-pi h b lvert xi rvert^{2}}, H_{2}(b lvert xi rvert^{2}),

*such that for every real s ge 0:*

H_{1}(s) le T_{k}(s),  H_{2}(s) ge 0,  (two signs)

*and at every node n in N:*calN:*

H_{1}(n) = T_{k}(n),  H_{1}'(n) = T_{k}'(n) = -e^{-k(n-1)},  H_{2}(n) = H_{2}'(n) = 0.  (four contacts per node)

This is the technical heart of the paper. Everything else unfolds from it.

**Why a pair, and why this normalization.** The two functions H_{1}, H_{2} will eventually form the two halves of f and widehatwidehat f.f. The Gaussian damping e^{-pi h b lvert x rvert^{2}} ensures f, widehatwidehat ff are Schwartz; the H_{i} are *entire* and control the jets at nodes. The two constants b and B = 1/b^{2} are dictated by the Gaussian-transform algebra: the transform of a Gaussian with parameter pi h b is another Gaussian with the same parameter, and the coupling constants B, h inside H_{i} are chosen to make this work out.

**Converting to the Gaussian of the main theorem.** If we take k = pi(alpha / b - h), a straightforward algebra computation gives

k e^{-k} e^{-pi h s} T_{k}(s) = e^{-(pi h + k) s} = e^{-pi alpha lvert x rvert^{2}} = G_{alpha}(x).

So k e^{-k} · f is a minorant of G_{alpha} (with f le e^{-pi h s} T_{k} from the theorem), its Fourier is nonnegative (multiplied by k e^{-k} > 0, retains nonnegativity), and its direct and dual contacts at N ⊇ A ∖ {0} match G_{alpha} and zero respectively. This is the Gaussian pair of Section 3.3, for every alpha ge 1.

For 0 < alpha < 1 (i.e. k too small), a Fourier-complement trick handles it (Section 8.2). We focus on alpha ge 1 for the construction.

**What "existence" means here.** The theorem asserts *there exists* a pair (H_{1}, H_{2}, f). It does not claim uniqueness --- in fact the space of such pairs, in a fixed coefficient system, is one-dimensional (Section 6 will pin it down by the *specific* structure we build). The construction is explicit; the numerical inequalities are certified separately.

### 4.3 The sine product P(s)

The first ingredient is a *periodic* function with a double zero at every node and nowhere else. Define

P(s) = prod_{a in L} ( 2 sin(pi(s - a))/(12) )^{2}.

Six squared sine factors, one for each residue class of N. Each factor is 2-periodic in the appropriate variable (period 12 in s); the product is nonnegative, periodic with period 12, has a double zero at every point of the periodic set Z, and no other real zeros.

P has a short complex-exponential expansion:

P(s) = sum_{ell = -6}^{6} P_{ell}, e^{i pi t_{ell} s},  t_{ell} = (ell)/(6),  P_{-ell} = overlineoverline{P_P_{ell}}.

**Coefficient values.** Setting w = e^{i pi s / 6} converts the product to w^{-6} prod_{a in L} (w - e^{i pi a / 6})^{2}. The scalar factor is 1 because sum_{a in L} a = 24 and the leading coefficient picks up exactly 1. Multiplying the degree-12 polynomial gives

(P_{0}, P_{1}, P_{2}, P_{3}, P_{4}, P_{5}, P_{6}) = ( 5, ; -1 + 2 i b, ; 1 + 2 i b, ; -2, ; -(1)/(2) + i b, ; -1 - 2 i b, ; 1 ).

The absolute values that matter later:

( lvert P_{1} rvert, ..., lvert P_{6} rvert ) = (2, 2, 2, 1, 2, 1),  sum_{ell = 1}^{6} lvert P_{ell} rvert = 10,  sum_{ell = 1}^{6} t_{ell} lvert P_{ell} rvert = (16)/(3).

**Node constants Q_{n}, D_{n}.** At any node n in N_{0}, write

P(n + u) = Q_{n} u^{2} ( 1 + D_{n} u + O(u^{2}) ).

The periodicity of P makes Q_{n}, D_{n} depend only on the residue of n mod 12:

| n bmod 12 | Q_{n} | D_{n} |
|---|---|---|
| 0 | pi^{2}/3 | -7 pi sqrt 3 / 18 |
| 1 | (2pi^{2}/3)(2 - sqrt 3) | pi(3 + sqrt 3)/18 |
| 3 | (2pi^{2}/3)(2 - sqrt 3) | -pi(3 + sqrt 3)/18 |
| 4 | pi^{2}/3 | 7 pi sqrt 3 / 18 |
| 7 | (2pi^{2}/3)(2 + sqrt 3) | -pi(3 - sqrt 3)/18 |
| 9 | (2pi^{2}/3)(2 + sqrt 3) | pi(3 - sqrt 3)/18 |

To get the table: pull out one vanishing sin^{2} factor, take the value and log-derivative of the remaining product. Concretely,

Q_{a} = ( (pi)/(6) )^{2} prod_{d in L ∖ {a}} ( 2 sin(pi(a - d))/(12) )^{2},  D_{a} = (pi)/(6) sum_{d in L ∖ {a}} cot(pi(a - d))/(12).

Two bounds (proved by rational comparisons of pi, sqrt 3; see Appendix A):

Q_{n} > (7)/(4) = 1.75,  lvert D_{n} rvert < (53)/(25) = 2.12.

These constants show up again in every jet bound and every sign estimate. The specific rational witnesses are (157)/(50) < pi < (22)/(7) and sqrt 3 < (1733)/(1000).

### 4.4 The entire cardinal functions

Now we turn "specify a value and a slope at every node" into a single entire function. The trick: multiply the sine product by a sum of *double poles* (to control values) and *simple poles* (to control slopes).

**Definition (coefficient list).** A coefficient list is data of the form

a = ( C ; ; {(c_{n}, d_{n})}_{n in N_{0}} ),

where C in R is a constant, and c_{n}, d_{n} in R for each node n. Its *unweighted* norm is

lVert a rVert_{1} = lvert C rvert + sum_{n in N_{0}} ( lvert c_{n} rvert + lvert d_{n} rvert ).

Summability (lVert a rVert_{1} < ∞) is the fixed admissibility criterion.

Define

p_{a}(s) = P(s) ( C + sum_{n in N_{0}} ( frac{c_{n}}{(s - n)^{2}} + frac{d_{n}}{s - n} ) ),  s in C ∖ N_{0}.

The double poles c_{n}/(s-n)^{2} interact with the double zero of P at n to give a *finite* value Q_{n} c_{n} at n. The simple poles give a finite derivative. Away from nodes, p_{a} is analytic.

**Lemma (Cardinal measure).** *For every summable a:*

*(a) p_{a} extends to an entire function, real on the real axis, with jets*

p_{a}(n) = Q_{n} c_{n},  p_{a}'(n) = Q_{n}( d_{n} + D_{n} c_{n} )  (n in N_{0}).

*(b) There is a finite complex measure mu_{a} on [-1, 1] such that*

p_{a}(s) = int_{-1}^{1} e^{i pi t s} , dmu_{a}(t)  (s in C),

*and mu_{a}(- E) = overline{mu_{a}(E)} for every Borel set E, plus the TV bound*

lVert mu_{a} rVert_{TV} le 25 lvert C rvert + 36 sum_{n in N_{0}} ( lvert c_{n} rvert + lvert d_{n} rvert ).

*(c) The same p_{a} has a "folded" evaluation on real s:*

p_{a}(s) = Re int_{0}^{1} e^{i pi t s} , dnu_{a}(t)  (s in R),

*where nu_{a} is a finite complex measure supported on [0, 1], with atoms C e_{j} P_{j} at t_{j} = j/6 (with e_{0} = 1, e_{j} = 2 for 1 le j le 6) and a density on each open interval ((j-1)/6, j/6) given by*

d_{j}(t) = 2pi sum_{n in N_{0}} e^{-i pi t n} sum_{ell = j}^{6} [ -pi c_{n}(t_{ell} - t) + i d_{n} ] P_{ell}, e^{i pi t_{ell} n}.

**Proof sketch.** For one node n, use u = s - n and the elementary integral identities

frac{e^{i pi t_{ell} u} - 1}{u} = i pi int_{0}^{t_{ell}} e^{i pi t u}, dt,  frac{e^{i pi t_{ell} u} - 1 - i pi t_{ell} u}{u^{2}} = -pi^{2} int_{0}^{t_{ell}} (t_{ell} - t), e^{i pi t u}, dt.

Multiply the first by P_{ell} e^{i pi t_{ell} n} and sum over ell; the subtracted constant and linear terms cancel because P(n) = P'(n) = 0. The result is an integral representation of c_{n}/(s-n)^{2} and d_{n}/(s-n) as spectral measures. The negative-frequency parts are conjugate reflections, so the "folding" gives item (c). Summability over n gives (b). QED

**Why these particular constants 25 and 36.** The constant column C P has variation lvert P_{0} rvert + 2 sum_{ell = 1}^{6} lvert P_{ell} rvert = 5 + 2 · 10 = 25. The c_{n} column has variation pi^{2} sum_{ell=1}^{6} t_{ell}^{2} lvert P_{ell} rvert = 65 pi^{2}/18 < 36. The d_{n} column has variation 2pi sum_{ell = 1}^{6} t_{ell} lvert P_{ell} rvert = 32pi/3 < 36. So 25 and 36 are just convenient upper bounds.

### 4.5 The Gaussian transform of a cardinal function

The whole point of the sine product coefficients P_{ell} is that they pair beautifully with complex Gaussian transforms. Define, for t in [-1, 1],
z(t) = -(B)/(t + i h) - i h,  lambda(t) = (i)/(b(t + i h)).
Both are holomorphic in t except at t = -ih (which is not on [-1, 1]; h = 2/5, so the pole is below the contour).

**Lemma (Transformed cardinal).** For every summable a, the *transformed cardinal function*
K_{p_{a}}(s) = int_{-1}^{1} lambda(t), e^{i pi z(t) s}, dmu_{a}(t)
is entire, and satisfies the reality condition K_{p_{a}}(sbar) = overlineoverline{K_K_{p_{a}}(s)} for all complex s. In particular it is real on the real axis.

**Why the formula works.** The complex Gaussian transform in dimension two, with our convention fhat(xi) = int f(x) e^{-2pi i x · xi} dx, is
widehatwidehat{e^e^{-pi gamma lvert x rvert^{2}}}(xi) = gamma^{-1}, e^{-pi lvert xi rvert^{2}/gamma}  (Re gamma > 0).
Apply it to the Gaussian e^{-pi h b lvert x rvert^{2}} · e^{i pi t s} with s = b lvert x rvert^{2}. The result is another Gaussian with parameter gamma = b(h - it), whose transform is another Gaussian with parameter 1/gamma. Writing gamma = b(h - it) = -ib(t+ih) and simplifying, the transform of e^{-pi h b lvert x rvert^{2}} e^{i pi t s} equals
e^{-pi h s}, lambda(t), e^{i pi z(t) s}.
Integrating against dmu_{a}(t) and swapping integrals gives the claimed K_{p_{a}}.

**Pointwise frequency consistency.** z is an involution on the upper half-plane: z(z(t)) = t. And lambda(t) lambda(z(t)) = 1. These identities hold pointwise; they do not define a composition of the folded integral because z does not preserve the integration contour [-1, 1]. They are also not needed for anything in the argument except as sanity checks; the Fourier-pair conclusion below is proved from the actual complex Gaussian formula, not from these.

**Lemma (Fourier pairs).** For every summable a, the radial functions e^{-pi h s} p_{a}(s) and e^{-pi h s} K_{p_{a}}(s), viewed as functions on R^{2} via s = b lvert x rvert^{2}, are real Schwartz functions, and are Fourier transforms of one another:
widehatwidehat{(e^(e^{-pi h s}, p_{a})} = e^{-pi h s}, K_{p_{a}},  widehatwidehat{(e^(e^{-pi h s}, K_{p_{a}})} = e^{-pi h s}, p_{a}.

**Lemma (Coupled pair).** Suppose a_{1}, a_{2} are summable. Define
H_{1}(s) = p_{a_{1}}(s) + K_{p_{a_{2}}}(s),  H_{2}(s) = p_{a_{2}}(s) + K_{p_{a_{1}}}(s).
Then the radial function f(x) = e^{-pi h b lvert x rvert^{2}} H_{1}(b lvert x rvert^{2}) has Fourier transform fhat(xi) = e^{-pi h b lvert xi rvert^{2}} H_{2}(b lvert xi rvert^{2}).

Both lemmas are proved in the paper by direct use of the complex Gaussian transform and the exchange of Fourier integration with the finite measure integration. The details: measure is finite, frequency is on a compact interval, every derivative of a Gaussian is bounded by a polynomial times a decaying Gaussian. So the exchanges are legitimate.

**Why two lists.** The point of the coupled form is symmetry on the direct/dual sides. A single list would produce f and widehatwidehat ff tied together by one function; two lists decouple them and let us impose independent interpolation conditions on H_{1} and H_{2} separately. This is the difference between the Cohn-Kumar scalar setting and the present coupled setting.

### 4.6 The jet operator and the distant-output bound

Everything we care about is *local data at nodes*: value and first derivative. Define the **jet projection**
(J u)_{c, n} = frac{u(n)}{Q_{n}},  (J u)_{d, n} = frac{u'(n) - D_{n} u(n)}{Q_{n}}  (n in N).
The normalizations Q_{n}, D_{n} undo the leading behavior of P at that node. With these factors,
J p_{a} = a^{+}
where a^{+} denotes the positive-node part of a, i.e. {c_{n}, d_{n}}_{n in N} with the "extra coordinates" (c_{0}, d_{0}, C) suppressed.

Define the **transformed jet operator**
S a = -J K_{p_{a}}.
Then the coupled pair H_{1}, H_{2} satisfies
J H_{1} = a_{1}^{+} - S a_{2},  J H_{2} = a_{2}^{+} - S a_{1}.

So the interpolation problem --- prescribe J H_{1} = J T_{k} and J H_{2} = 0 at every positive node --- becomes the *linear system*
a_{1}^{+} - S a_{2} = J T_{k},  a_{2}^{+} - S a_{1} = 0.

Whether this system has a solution depends on how small the off-diagonal operator S is. The construction is designed so S is not small everywhere, but is exponentially small *at distant nodes*, which is enough to fit into a fixed-point framework.

**Lemma (Kernel bounds).** For every t in [0, 1]:
lvert lambda(t) rvert < 3,  lvert z(t) rvert le (44)/(15) < 3,  pi Im z(t) > 0.187.
On 0 le t le 5/6 the last improves to pi Im z(t) > 0.70. On 5/6 le t le 1 the bounds are sharper:
lvert lambda(t) rvert < 1.25,  lvert z(t) rvert < 1.32,  pi Im z(t) > 0.187 + 2.48(1 - t).
In particular, lambda and z are uniformly bounded, and the damping exponent pi Im z(t) is uniformly positive on [0, 1], with Im z(t) ge 0.187 / pi in the worst case.

**Verification.** Direct calculation gives
lvert lambda(t) rvert = frac{1}{b sqrt{t^{2} + h^{2}}},  lvert z(t) rvert^{2} = h^{2} + frac{B^{2} - 2B h^{2}}{t^{2} + h^{2}},  pi Im z(t) = pi h ( frac{B}{t^{2} + h^{2}} - 1 ).
Since B > 2 h^{2}, all three decrease with t. Their largest / smallest values on [0, 1] occur at endpoints, giving the displayed numerics. The sharper endpoint 5/6 comes from plugging t = 5/6 into these formulas and using the rational bounds on pi and sqrt 3.

**Tail bound (distant outputs of S).** For every positive integer r, every pair (sum of absolute node coefficients ge w, absolute constant coordinate ge c) of inputs satisfies
sum_{n in N,; n ge r} ( lvert (Sa)_{c, n} rvert + lvert (Sa)_{d, n} rvert ) le U_{r}(w, c) := frac{800(w+c) e^{-0.70 r}}{1 - e^{-0.70}} + frac{5.3(2.54 w / r + 2c) e^{-0.187 r}}{1 - e^{-0.187}}.

**Why it works.** For n ge r ge 1, the integrand in K_{p_{a}} carries an explicit exponential e^{i pi z(t) n}; its real part is e^{-pi Im z(t) n} le e^{-0.187 n}. The frequency derivative from d/ds brings down a polynomial in lvert z(t) rvert < 3, controlled by the 2.54 w / r factor in the second term. On [0, 5/6] the stronger e^{-0.70 n} decay dominates, giving the first term; on [5/6, 1] the weaker but polynomial-decay factor dominates, giving the second. The constants 800 and 5.3 are the total-variation product 36 · pi · lvert dimension factor rvert and a uniform bound on the growth of lvert z rvert on that segment.

**Operator norm.** Collecting the two bounds over all node columns (summing the le U_{1}(1, 1) bound, which is about 2500) gives
lVert S rVert_{op (finite lists)} < 4000.
The specific bound 4000 is what shows up everywhere later. It is not sharp; it is what the elementary estimates give.

**Uniform derivative bounds.** For every integer j ge 0 and real s ge 0:
lvert p_{a}^{(j)}(s) rvert le pi^{j} V,  lvert K_{p_{a}}^{(j)}(s) rvert le 3(3pi)^{j} e^{-0.187 s} V,
where V = 25 lvert C rvert + 36 sum_{n} (lvert c_{n} rvert + lvert d_{n} rvert). The refined two-range bound separates the two frequency bands:
lvert K_{p_{a}}^{(j)}(s) rvert le 3(3pi)^{j}(36w + 25c) e^{-0.70 s} + 1.25 (1.32 pi)^{j} ( (2.54 w)/(s) + 2 c ) e^{-0.187 s}.

These bounds are uniform in the node locations and do not require finite support.

### 4.7 The coupled interpolation problem

The task of Stages A-C is:

**Find summable coefficient lists a_{1}, a_{2} with prescribed extra coordinates e_{i} = E_{i} q(k) (see Section 5) such that**
J H_{1} = J T_{k},  J H_{2} = 0  on all positive nodes  n in N.

Once found, the pair (H_{1}, H_{2}) satisfies all four contacts of Theorem 4.2. The next two sections
- split the infinite problem into a finite block (84 nodes at n le 168) and a tail;
- solve the finite block numerically, and prove the tail is contractive;
- bound the combined correction by 2 · 10^{-8} in ell^{1} norm.

The Gaussian pair's existence follows by this fixed-point argument. The signs H_{1} le T_{k} and H_{2} ge 0 (which are also part of Theorem 4.2) come later, in Section 7, from a Bernstein-polynomial certification.

---

## 5. Stage B: cutting to finite data

The interpolation problem of Section 4.7 lives in the infinite summable space. In practice it is solved in two steps:

1. **Truncate.** Restrict the positive nodes to those with n le 168 and the spectral variable s = b lvert x rvert^{2} to a finite quadrature grid. The truncated system is finite-dimensional; it is a 168 times 171 matrix equation.
2. **Correct for truncation.** Prove that the full infinite system differs from the finite one by a factor small enough that the finite solution, extended by zero, is within 2 · 10^{-8} of the full solution.

Step 2 is the harder of the two analytically, and uses the tail bound U_{r}(w, c) of Section 4.6.

### 5.1 Finite data

The finite truncation parameters are

M = 168,  F = N cap [1, M],  N = 384.

There are exactly 84 positive nodes in F:calF: each successive block of 12 integers contains 6 permitted residues {0, 1, 3, 4, 7, 9}, and 14 times 6 = 84.

A finite positive-node list has 168 coordinates: first all c_{n} in increasing n, then all d_{n} in increasing n. Three extra coordinates are appended: e = (c_{0}, d_{0}, C) in that order. So the complete finite input space has 171 coordinates.

### 5.2 The Fejér quadrature rule

Recall the folded spectral measure of Lemma 4.4(c): atoms C e_{j} P_{j} at t_{j} = j/6, and a density d_{j}(t) on each open interval ((j-1)/6, j/6) for 1 le j le 6. The atoms are kept exact (no quadrature error). The density on each interval is replaced by a finite quadrature sum.

The rule used is the **first Fejér rule**, split by interval. For each interval index j in {1, ..., 6} and each quadrature index v in {0, 1, ..., N - 1}:

alpha_{v} = (v + 1/2)/(N),  tau_{j v} = frac{2j - 1 + cos(pi alpha_{v})}{12},  o_{v} = (1)/(N) ( 1 - 2 sum_{a = 1}^{N/2 - 1} frac{cos(2 a pi alpha_{v})}{4 a^{2} - 1} ).

The points tau_{j v} lie in the *interior* of the interval: 0 < pi alpha_{v} < pi implies cos(pi alpha_{v}) in (-1, 1), hence (2j - 2)/12 < tau_{j v} < 2j / 12. The weights o_{v} are positive and sum to one on [0, 1]-normalized interval mass.

**What the quadrature replaces.** The density contribution to p_{a}(s) is
int_{(j-1)/6}^{j/6} d_{j}(t), e^{i pi t s}, dt.
The quadrature replacement is
sum_{v = 0}^{N-1} frac{o_{v}}{6}, d_{j}(tau_{j v}), e^{i pi tau_{j v} s}.
The factor of 1/6 is the physical length of the interval.

**Quadrature error bound.** For every integrand of the form d_{j}(t) h(t) with h entire and lvert h^{(m)}(t) / m! rvert le R^{m} for lvert t - u rvert le 1/12 around each midpoint u = (2j - 1)/12, the difference between the integral and the quadrature sum is bounded by 10^{-31}.

**Why this scale.** Fejér's first rule is exact for polynomials of degree le N - 1 = 383 on the standard interval. The Taylor series of h on a disc of radius 1/12 (smaller than the physical gap, so it stays inside the permitted region) converges fast enough that the tail after degree 383 is negligible. Cauchy's coefficient estimate gives lvert h^{(m)}(u)/m! rvert le bound, and the geometric tail is sum_{m ge 384} 2^{-m} < 2^{-383}. The result is dominated by 10^{83} · 2^{-384}, which is below 10^{-31}. The exact numerical allowance in the paper is 4 · 10^{83} · 2^{-384} < 10^{-31}; the integer inequality 4 · 10^{114} < 2^{384} is what makes it work.

10^{-31} is much smaller than every subsequent error budget (10^{-8}), so it will effectively be zero in the correction calculus.

### 5.3 The mass matrix W

The finite system is expressed as a matrix multiplication. Define the **mass matrix** W, with

rows = 6N + 7 = 2311,  columns = 171.

Its rows are indexed by (j, v) in {1, ..., 6} times {0, ..., N-1} (the density quadrature points, in lexicographic order) followed by j in {0, ..., 6} (the seven atoms, in increasing j order). Its columns use the finite-coordinate order specified in Section 5.1.

The entries are, for t = tau_{j v} and each finite node n in F cup {0}:

W_{(j, v), c_{n}} = -frac{2 pi^{2} o_{v}}{6}, e^{-i pi t n} sum_{ell = j}^{6} (t_{ell} - t), P_{ell}, e^{i pi t_{ell} n},
W_{(j, v), d_{n}} = frac{2 pi i o_{v}}{6}, e^{-i pi t n} sum_{ell = j}^{6} P_{ell}, e^{i pi t_{ell} n},
W_{(j, v), C} = 0,

and for atoms at j in {0, ..., 6}:

W_{atom(j), C} = e_{j} P_{j},  W_{atom(j), c_{n}} = W_{atom(j), d_{n}} = 0,

where e_{0} = 1 and e_{j} = 2 for j > 0.

**What Wa computes.** For any input column a (a 171-vector), the vector Wa in C^{2311} has entries equal to the finite quadrature values of p_{a} (or more precisely, of the folded density and atomic contributions) at the mass points. Multiplying by the moment factors (Section 5.4) recovers the derivatives of p_{a} at any real center.

Two notes.

1. The column ordering matters: the c_{n} and d_{n} blocks are grouped so that W can be split cleanly into the "positive coordinates" block (columns 1, ..., 168) and the "extras" block (columns 169, ..., 171).
2. All entries are exact rationals times exponentials e^{i pi t n}; later we will floor-round to rationals with a floor ε = 10^{-80} (Appendix B).

### 5.4 Moment factors and Taylor jets on the finite side

The mass matrix W produces the *zeroth* finite value of p_{a} at each mass point. To get higher derivatives, multiply by the **moment factors**. Define, for integers a ge 0, real numbers m, y, and each mass point t:

F_{a}(m, y; t) = e^{i pi t m} frac{(i pi y t)^{a}}{a!},  G_{a}(m, y; t) = lambda(t), e^{i pi z(t) m} frac{(i pi y z(t))^{a}}{a!}.

For a real coefficient column A in R^{171}, the Taylor coefficient of order a at m (scaled by y^{a}/a!) of p_{a} is
frac{y^{a}}{a!} p_{a}^{(a)}(m) = Re( F_{a}(m, y; t)^{top} W A ).
Similarly, for the transformed cardinal function,
frac{y^{a}}{a!} K_{p_{a}}^{(a)}(m) = Re( G_{a}(m, y; t)^{top} W A ).

**No conjugation.** Note that F_{a} and G_{a} are *vectors* (over the 2311 mass points), and the multiplication is an ordinary dot product (adjoint )^{top}, without taking the complex conjugate. This is important: the construction uses the linearity of the finite sum and differentiates it termwise; complex conjugation would convert the derivative to a conjugate derivative and give the wrong signs.

**Why this works on the finite side only.** The formulas use W A, which is the *finite* quadrature evaluation of p_{a}. This is not the same as the exact p_{a}(m); the two differ by the quadrature error, which is 10^{-31}. So the finite moment products approximate the exact Taylor coefficients to that precision.

### 5.5 The finite operator, the target, and the reference columns

We now assemble the finite interpolation problem.

**The finite operator S^{#}.** With W the mass matrix as above, the 168 times 171 matrix representing "the transformed jet operator on finite data" is:

S^{#}_{c(m), :} = -frac{1}{Q_{m}} Re( G_{0}(m, 1; t)^{top} W ),
S^{#}_{d(m), :} = frac{1}{Q_{m}} Re( D_{m} G_{0}(m, 1; t)^{top} W - G_{1}(m, 1; t)^{top} W ),

for each finite node m in F. The rows are indexed by the same 168 coordinates as before (all c_{m} in increasing m, then all d_{m}).

**Interpretation.** For a finite input a in R^{171}, the vector S^{#} a in R^{168} is the finite quadrature evaluation of J K_{p_{a}} at the finite nodes. The sign convention: S^{#} is the negative of J K_{p} (matching the definition of S in Section 4.6).

**The target column Y.** The right-hand side at five "fundamental nodes" n_{1}, ..., n_{5} = 1, 3, 4, 7, 9 is the target T_{k}. Because T_{k} is an exponential, its jet projection has a compact form. Let q(k) in R^{10} be the vector
q(k) = ( T_{k}(1), T_{k}(3), T_{k}(4), T_{k}(7), T_{k}(9), ; -T_{k}'(1), ..., -T_{k}'(9) ).

Define Y by
Y_{d(n_{j}), j} = -frac{1}{Q_{n_{j}}},  Y_{c(n_{j}), j + 5} = frac{1}{Q_{n_{j}}},  Y_{d(n_{j}), j + 5} = -frac{D_{n_{j}}}{Q_{n_{j}}},  j = 1, ..., 5,
with all other entries zero.

**What Y q(k) is.** At the nodes n in {1, 3, 4, 7, 9}, Y q(k) gives the target jet J T_{k}(n). At every other finite node, Y q(k) = 0. This is the finite target data, exact, no quadrature error, because T_{k} evaluated at integers is a finite expression.

The target at the remaining finite nodes (all those that are not in {1, 3, 4, 7, 9}) is small -- bounded by a geometric tail, < 10^{-11} in ell^{1} norm -- so it will be treated as an error term in the correction argument. This is the "5 nodes carry the data, the rest are handled as perturbations" structure.

**The extra coordinates.** The three extra coordinates e_{i} = (c_{0}, d_{0}, C) for list i are chosen explicitly, independent of k. Collect them into vectors:

- E_{1} in R^{3 times 10}: row of e coordinates for H_{1}'s list.
- E_{2} in R^{3 times 10}: row of e coordinates for H_{2}'s list.

Concretely, the first column of E (viewed as 6 times 10) is
(0.567, ; 0.250, ; -0.0074, ; 0, ; -0.209, ; 0.0096)^{top}
and the sixth column is
(-0.976, ; 0.408, ; 0.0115, ; 1, ; 0.938, ; -0.017)^{top}.
The first three entries of each column give E_{1}, the last three give E_{2}. So E_{i} q(k) =  (extras of list i for target data q(k)).

**Why this choice.** The extras matter because the coupled equations for H_{1}, H_{2} include cross-contributions S from the *other* list's extras. Without an E_{i} source, the constant-coordinate column would be zero and the finite system would be degenerate. The specific numeric values (.567, .250, ...) were found by solving for the extras that put the full infinite interpolation system into a well-posed regime; they are documented and verifiable, but the "why these specific numbers" question is really "why any numbers with these properties work" -- the answer is an interval of valid choices, of which these are a convenient point.

**The coupled finite system.** Splitting S^{#} into its finite-positive columns S^{#}_{F F} (168 times 168), finite extras columns S^{#}_{F e} (168 times 3), and writing D = -S^{#}_{F F} (so that D is the finite off-diagonal in the block equation), the finite version of the coupled interpolation system reads

A_{1} - D A_{2} = Y + S^{#}_{F e} E_{2},  A_{2} - D A_{1} = S^{#}_{F e} E_{1}.

Here A_{i} in R^{168 times 10} are the finite-positive parts of the two lists, one column per coordinate of q(k). The sign of D follows from the sign convention in S: -S on the diagonal becomes +D when the equation is written in this "shifted" form.

**The finite geometric-sum solution.** Summing and differencing the two equations,
(I - D)(A_{1} + A_{2}) = g_{1} + g_{2},  (I + D)(A_{1} - A_{2}) = g_{1} - g_{2},
where
g_{1} = Y + S^{#}_{F e} E_{2},  g_{2} = S^{#}_{F e} E_{1}.

For the finite system, we replace the exact inverses (I - D)^{-1} and (I + D)^{-1} by truncated geometric sums:
V_{+} = sum_{j = 0}^{511} D^{j},  V_{-} = sum_{j = 0}^{511} (-D)^{j}.

Then
X_{1 F} = (1)/(2)( V_{+}(g_{1} + g_{2}) + V_{-}(g_{1} - g_{2}) ),  X_{2 F} = (1)/(2)( V_{+}(g_{1} + g_{2}) - V_{-}(g_{1} - g_{2}) ),
and
X_{i} =  X_{i F}  E_{i}  in R^{171 times 10},  i = 1, 2.

The reference list for a given k is then X_{i} q(k), extended by zero on the positive nodes n > 168 (the "tail").

**Error from the truncation of the geometric sum.** This is where the tables in Section 5.6 come in. The identities
(I - D) V_{+} = I - D^{512},  (I + D) V_{-} = I - D^{512}
show that the finite geometric sums solve the equations up to D^{512}. If lVert D^{512} rVert is small enough, the difference between X_{i} q(k) and the true solution is small. Section 5.6 verifies D^{2^9} = D^{512} has norm below 10^{-12} (row 8 of the power table).

**What is left.** Given the reference lists X_{i} q(k), the full infinite solution differs from them by a correction that lives in the difference between the finite quadrature and the exact spectral integrals (bounded by 10^{-29}), the difference between the finite nodes and all nodes (bounded by 10^{-14} at 168 from the tail bound U_{168}), and by the difference between D^{512} and zero. Section 6 assembles these bounds into a 2 · 10^{-8} correction.

### 5.6 The power-sum computation and the two tables

The argument of Section 5.5 requires explicit bounds on the powers of D that enter the geometric sums. These are computed by **repeated squaring**, and the bounds are strict rationals at each step.

**Algorithm.** Start with U = D and V_{+} = V_{-} = I. Then for r = 0, ..., 8:

- Update V_{+} ← V_{+}(I + U).
- Update V_{-} ← V_{-}(I + ε_{r} U) where ε_{0} = -1 and ε_{r} = 1 for r > 0.
- Record the column-sum norms of U, V_{+}, V_{-}.
- Square U ← U^{2} (before the next step).

After step 8 has been executed, U = D^{512} is discarded, and V_{+} = sum_{j=0}^{511} D^{j}, V_{-} = sum_{j=0}^{511} (-D)^{j}, exactly the sums used in Section 5.5.

**Why this works.** Before step r (i.e. at the top of the loop with the current U), U = D^{2^{r}}. Then V_{+} is being multiplied by (I + D^{2^{r}}), and
prod_{r=0}^{9} (I + D^{2^{r}}) = sum_{j=0}^{511} D^{j},
the standard binary identity. The V_{-} product is analogous: it uses (I - D)(I + D)(I + D^{2}) ..., which is the alternating sum sum_{j=0}^{511} (-D)^{j}.

**Power table.** The bounds are verified numerically; the entries below are strict upper bounds on the column-sum norm at each stage.

| Step r | lVert U rVert | lVert V_{+} rVert | lVert V_{-} rVert |
|---|---|---|---|
| 0 | 1.74 | 2.74 | 2.1 |
| 1 | 2.09 | 6.94 | 2.2 |
| 2 | 2.27 | 15.32 | 2.13 |
| 3 | 1.75 | 25.3 | 1.97 |
| 4 | 0.75 | 31.1 | 1.88 |
| 5 | 0.12 | 32.09 | 1.85 |
| 6 | 0.0031 | 32.12 | 1.85 |
| 7 | 0.000002 | 32.12 | 1.85 |
| 8 | 0.000000000001 | 32.12 | 1.85 |

**Reading the table.**

- Row 6 gives lVert D^{64} rVert le 0.0031, a serious power contraction.
- Row 8 gives lVert D^{256} rVert le 10^{-12}.
- Therefore lVert D^{512} rVert le (10^{-12})^{2} = 10^{-24}: the geometric-sum truncation error is essentially zero compared to every other error budget.

This is the entire mechanism by which the finite geometric sums V_{pm} compute the reference lists. The bounds are discovered by running the algorithm and checking each column-sum in rational arithmetic; they are not derived from general principles. The construction is empirical at this step, and the resulting numerical table is what makes the proof go through.

**Column-norm table.** Each reference column has a bounded ell^{1} norm; the bounds are:

| Quantity | Columns 1-5 | Columns 6-10 |
|---|---|---|
| max_{j} lVert (X_{1})_{j} rVert_{1} | 1.29 | 2.65 |
| max_{j} lVert (X_{2})_{j} rVert_{1} | 1.01 | 3.03 |
| max_{j} sum_{i=1}^{2} lVert (X_{i})_{j} rVert_{1} | 2.20 | 5.40 |

The two right-hand columns correspond to the Y-target columns (indices 1, ..., 5) and the derivative-target columns (indices 6, ..., 10). The combined column norms are what appear in the reference-list norm bound:
sum_{i=1}^{2} lVert X_{i} q(k) rVert_{1} < 4.8  (for every  k ge 2.36).

### 5.7 Verification

Every entry of both tables is a strict rational inequality that can be checked by exact arithmetic on the finite matrix D. The two accompanying Python programs do this:

- `verification/arithmetic_bounds.py`: checks the scalar budgets and the numerical error-propagation majorants, using 512-bit Arb balls and exact rational arithmetic.
- `verification/numeric_balls.py`: encloses the exact quadrature arrays at 256-bit precision and tests every strict comparison in the power and column tables, plus the sign tables in Section 7.

Both are documented in `verification/README.md`. Together they certify the finite data. The passage back to the exact infinite spectral integrals is supplied by the quadrature-error lemma of Section 5.2.

---

## 6. Stage C: exact interpolation

We now prove that the finite reference lists of Section 5 can be corrected to *exact* solutions of the full infinite interpolation system, with a correction bounded by 2 · 10^{-8} in ell^{1}.

### 6.1 The proposition

**Proposition (Exact interpolation).** *For every real k ge 2.36, there exists a unique pair of summable positive-node coefficient lists a_{1}, a_{2} in E (where E is the space of lists with prescribed extra coordinates e_{i} = E_{i} q(k)), whose full lists ell_{1}, ell_{2} satisfy the exact jets*
H_{1}(n) = T_{k}(n),  H_{1}'(n) = T_{k}'(n),  H_{2}(n) = 0,  H_{2}'(n) = 0  (n in N),
*where H_{1} = p_{ell_{1}} + K_{p_{ell_{2}}}, H_{2} = p_{ell_{2}} + K_{p_{ell_{1}}}.*

*Let ell_{i}^{ref} = X_{i} q(k), extended by zero on the tail T (all positive nodes n > 168). Then*
sum_{i=1}^{2} lVert ell_{i}^{ref} rVert_{1} < 4.8,  sum_{i=1}^{2} lVert ell_{i} - ell_{i}^{ref} rVert_{1} < 2 · 10^{-8},  sum_{i=1}^{2} lVert ell_{i} rVert_{1} < 5.

*The functions H_{1}, H_{2} are entire and real on the real axis. In particular H_{1} - T_{k} and H_{2} have double zeros at every positive node, including the artificial nodes that are not actual lattice radii.*

The third inequality (total ell^{1} norm < 5) is the sum of the reference bound 4.8 and the correction bound 2 · 10^{-8}.

**Uniqueness.** The bounded inverse below gives uniqueness, within the prescribed coefficient system, for summable lists with fixed extras and jets prescribed on the enlarged periodic node set N.

This is worth contrasting with what is *not* claimed. Talebizadeh Sardari (2021, Theorem 1.7) proved that infinitely many linearly independent radial Schwartz functions have all their values and first derivatives (and those of their Fourier transforms) *vanish on the actual triangular shells* alone. So the interpolation problem that prescribes jets only on A ∖ {0} is *not* well-posed. The present paper enlarges the node set to N, which contains all the actual shells and many artificial extras. The well-posedness is on the larger set, within the fixed summable coefficient system.

### 6.2 The nonlinear block equation

For summable lists with fixed extras, the coupled interpolation system of Section 4.7 becomes a nonlinear equation in the unknown tail a_{T} = (a_{1 T}, a_{2 T}:

a_{F} = L_{F F}{-1} (r_{F} - L_{F T} a_{T},

a_{T} = L_{T T}{-1} (r_{T} - L_{T F} a_{F},

where

- a_{F} is the finite-positive part (the first 168 coordinates of each list);
- a_{T} is the tail (nodes n > 168);
- L is the 2 times 2 block system satisfying the coupled equation from Section 4.7, split into finite/finite, finite/tail, tail/finite, tail/tail blocks;
- r_{F}, r_{T} are the residual vectors (target minus the current approximation), bounded below.

The Schur complement of this system, with a_{F} eliminated, is a tail-space equation
Q a_{T} = psi_{T},  Q = L_{T T} - L_{T F} , L_{F F}{-1} L_{F T},
where psi_{T} is the effective right side.

### 6.3 The operator bounds

Four bounds establish that Q is a small perturbation of the identity.

**Bound on the finite inverse.** lVert L_{F F}{-1} rVert < 40. This follows from the finite column-sum bound: the finite operator, viewed as a 336 times 336 block matrix (two lists combined), has column sums bounded by the reference-list column norms.

**Bound on the tail/tail block.** lVert L_{T T} - I rVert < delta, where delta = 1.1 · 10^{-14}. The tail decay is controlled by U_{168}(w, c) from Section 4.6, with the specific values w = 5 and c = 0.02 from the reference-list bounds.

**Bound on the tail/finite block.** lVert L_{T F} rVert < delta. Same mechanism; the transformed contribution from a finite input vanishes rapidly at distance n > 168.

**Bound on the finite/tail block.** lVert L_{F T} rVert < 4000. This uses the operator norm of S (Section 4.6). It is not small; this is why one cannot just invert the finite block naively. But the Schur complement absorbs it.

### 6.4 The Schur complement is invertible

Combining the tail/finite, finite/finite, finite/tail, and tail/tail bounds gives
lVert Q - I rVert < delta(1 + 40 · 4000) = 1.760011 · 10^{-9} < 1.

Since this is < 1, Q has a convergent Neumann series Q^{-1} = sum_{k ge 0} (I - Q)^{k}, and Q is invertible with
lVert Q^{-1} rVert < frac{1}{1 - 1.76 · 10^{-9}} approx 1 + 1.76 · 10^{-9}.

Explicitly, for any right side psi = (psi_{F}, psi_{T}, the solution is reconstructed as
a_{T} = Q^{-1} ( psi_{T} - L_{T F} L_{F F}{-1} psi_{F} ),
a_{F} = L_{F F}{-1} ( psi_{F} - L_{F T} a_{T} ).

This is the exact full solution for any given right side psi. The question is now: how far is our reference list X q(k) from this exact solution?

### 6.5 The three sources of residual

The reference list ell^{ref} = (X_{1} q(k), X_{2} q(k)) satisfies the interpolation system approximately. Three sources of discrepancy:

1. **The target at the ten "carrier" nodes.** The five q(k)-target nodes {1, 3, 4, 7, 9} are hit exactly. But the target T_{k} is nonzero at the other 79 finite nodes; the reference list gives Y q(k) = 0 there. The total target loss from omitted finite nodes is bounded by
(1 + (1 + 2.12)/2.36)/(1.75) sum_{n = 12}^{∞} e^{-2.36(n - 1)} < 8 · 10^{-12}.

2. **The quadrature vs. exact measure.** Section 5.2's quadrature error is 10^{-31} per quadrature point; multiplied by 2311 entries it stays well below 10^{-24}.

3. **The geometric-sum truncation.** As computed in Section 5.6, lVert D^{512} rVert < 10^{-24}. Three-times the residual from omitted D powers, times the zero-tail norm, is negligible.

Adding these three bounds:
lVert r_{F} rVert_{1} < 8 · 10^{-12} + 10^{-24} + 1.4 · 10^{-15} < 10^{-11}.
(The middle number is a target-tail calculation; the last is the geometric-sum bound.)

**Tail residual.** At tail nodes n > 168, the reference list has zero positive-node coefficients, but the transformed operator still acts. Its total norm there is dominated by
lVert r_{T} rVert_{1} le U_{168}(5, 0.02) + 10^{-170} < 8.3 · 10^{-14}.
The .02 is the combined extras constant bound (Section 6.6 below); the 10^{-170} is the target T_{k} decayed to node n = 169 and beyond, via the local geometric majorant of Section 8.

### 6.6 The extras constant and reference corner

The three extras coordinates satisfy, for E_{i} q(k) with k ge 2.36,
lvert E_{i} q(k) rvert < .02  (each coordinate).
This is verified numerically: the specific values in Section 5.5 give lvert -.0074 + .0115/k rvert + lvert .0096 - .017/k rvert = .017 - .0285/k < .02 for k ge 2.36.

Remember that the sum of the three extra coordinates, plus the reference list's other coordinates, is < 4.8 in ell^{1} norm (Section 5.5). Adding the tail 10^{-170} and extras .02 gives lVert ell^{ref} rVert < 4.8 + 10^{-170} still < 5.

### 6.7 The correction size

Put u_{F} = lVert (a - a^{ref})_{F} rVert_{1} and u_{T} = lVert (a - a^{ref})_{T} rVert_{1}. The block equations for the correction (a - a^{ref}), together with the Schur bound, give the system

u_{F} le 40 ( 10^{-11} + 4000 , u_{T} ),
u_{T} le 8.3 · 10^{-14} + 1.1 · 10^{-14} (u_{F} + u_{T}

Substituting the first inequality into the second:
u_{T} le frac{8.3 · 10^{-14} + 4.4 · 10^{-24}}{1 - 1.760011 · 10^{-9}} < 8.31 · 10^{-14},
and from that
u_{F} < 1.37 · 10^{-8}.

Their sum is < 2 · 10^{-8}, which is the claimed correction bound.

Adding u_{F} + u_{T} to the reference norm lVert ell^{ref} rVert < 4.8 gives lVert ell rVert < 5 for the exact list.

The extras did not change in the correction (they are pinned by the prescribed E_{i} q(k)), so this is the full-list bound, not just the positive-node part.

### 6.8 Finishing the proof of Proposition 6.1

The exact solution a of the full interpolation system of Section 6.2 exists and is unique (the Schur complement is invertible). Its lists ell_{1}, ell_{2} satisfy
J H_{1} = J T_{k},  J H_{2} = 0
at every positive node, and hence by definition of J, the four contacts
H_{1}(n) = T_{k}(n),  H_{1}'(n) = T_{k}'(n),  H_{2}(n) = 0,  H_{2}'(n) = 0
at every n in N.

The H_{i} are entire and real on the real axis by construction (Section 4.5). Lemma 4.5 gives them as the two halves of a real radial Schwartz Fourier pair, which is f and widehatwidehat ff of the normalized Gaussian theorem.

This completes the existence half of Theorem 4.2. The signs H_{1} le T_{k} and H_{2} ge 0 are established in Section 7.

---

## 7. The signs at every radius and every parameter

We now prove the two sign conditions
H_{1}(s) le T_{k}(s),  H_{2}(s) ge 0  (s ge 0, ; k ge 2.36).
The contacts of Proposition 6.1 give equality at every node. So the sign conditions only need to be proved *strictly between* nodes.

### 7.1 The problem, stated

Recall:
H_{i}(s) = p_{ell_{i}}(s) + K_{p_{ell_{3-i}}}(s),  T_{k}(s) = k^{-1} e^{-k(s-1)},  i = 1, 2.
We want H_{1} - T_{k} le 0 and H_{2} ge 0. The signs -sigma_{i} with sigma_{1} = -1, sigma_{2} = 1 unify the two statements:
sigma_{1}(H_{1} - T_{k}) ge 0,  sigma_{2} H_{2} ge 0,  i.e.  sigma_{i}(H_{i} - delta_{i} T_{k}) ge 0
with delta_{1} = 1, delta_{2} = 0.

Since we have *exact* contacts at nodes, both sides vanish at every n in N. So the difference is *twice differentiable through zero* there. That is what makes the proof work.

### 7.2 The deleted-slack reduction

Define, for a twice-continuously-differentiable function u and a node m,
D_{m, 2} u(s) = frac{u(s) - u(m) - (s - m) u'(m)}{(s - m)^{2}}  (s ne m),
with removable value D_{m, 2} u(m) = u''(m)/2. At the origin, we set D_{0, 0} u = u. Define also the "slack" function
F_{i, slack}(s) = sigma_{i}(H_{i}(s) - delta_{i} T_{k}(s)),
so that the sign we want is F_{i, slack}(s) ge 0.

**Key identity.** At every positive node m in N, both H_{i} and delta_{i} T_{k} have the target jets (by the contacts), so:
F_{i, slack}(s) = (s - m)^{2}, D_{m, 2}(sigma_{i}(H_{i} - delta_{i} T_{k}))(s).
The sign F_{i, slack}(s) ge 0 is therefore equivalent to
sigma_{i}, D_{m, 2}( H_{i} - delta_{i} T_{k} )(s) ge 0.
(The sign of (s - m)^{2} is positive, so it drops out.)

The strategy is now clear:
- For each half-gap [m, m'] covered by s in [0, s_{0}] with s_{0} = 89.5, prove the quotient inequality by finite polynomial computation.
- For s ge s_{0}, prove the same inequality using a quadratic barrier for the sine product P(s).

### 7.3 Bernstein rows and the half-gap reduction

We construct polynomial rows to which we will apply a Bernstein-polynomial argument. Recall that on [0, 1], the *Bernstein basis functions* binom{40}{d} x^{d}(1 - x)^{40 - d} for d = 0, ..., 40 are nonnegative and sum to one. So if a polynomial has all Bernstein coefficient > c, then it is > c on the whole interval [0, 1].

For a fixed center m in N_{0} cap [0, 88] and a signed half-gap length y, the half-gap consists of s = m + yx, 0 le x le 1. The Taylor coefficients of the "deleted quotient" are
D_{m, nu} u(m + yx) = sum_{r = 0}^{∞} frac{y^{r}}{(r + nu)!} u^{(r + nu)}(m) x^{r},
where nu = 0 if m = 0 and nu = 2 if m > 0.

Define the **Bernstein Taylor row** at center m, signed half-gap y:
l_{i, r} = sigma_{i} frac{y^{r}}{(r + nu)!} ( widetilde p_{X_{i}}^{(r + nu)}(m) + widetilde K_{X_{3 - i}}^{(r + nu)}(m) ),  0 le r le 40,
where the tilde denotes the finite quadrature evaluation (an approximation to the exact p_{X_{i}}^{(r + nu)}(m) and K_{X_{3 - i}}^{(r + nu)}(m)).

And define the **Bernstein row**
L_{i, d} = sum_{r = 0}^{d} omega_{d, r}, l_{i, r},  omega_{d, r} = frac{binom{d}{r}}{binom{40}{r}}.

By design, L_{i, d} is the d-th Bernstein coefficient of the partial Taylor series sum_{r=0}^{40} l_{i, r} x^{r} of the deleted quotient, with the **partial Taylor series being of the reference lists** X_{i} q(k) rather than the exact lists ell_{i}.

**The correction from reference to exact.** By (5.4), the difference between reference and exact is controlled by lVert ell_{i} - X_{i} q(k) rVert_{1} < 2 · 10^{-8}. The derivative of a cardinal function is bounded by pi^{j} V, and the derivative of the transformed by 3(3pi)^{j} e^{-.187 s} V, with V le min(36 · 5, 180). The correct difference at every step is bounded by
sup_{s  in half-gap} | D_{m, nu} (H_{i} - H_{i}^{ref})(s) | < .00011.
This uses nu = 0 (zeroth derivative) or nu = 2 (with the additional 1/2 factor from int_0^1 (1-v) dv), and the maximum of pi^{nu}, 3(3pi)^{nu}.

Similarly, the finite-truncation error in the Bernstein rows (from truncating the Taylor series after 40 terms) is bounded by 10^{-6}. And the quadrature error is 41 · 4 · 5 · 10^{-31} = 8.2 · 10^{-29}.

### 7.4 Parameter boxes

The parameter k ranges over [2.36, ∞). Divide into five intervals:
[2.36, 2.65],  [2.65, 3.2],  [3.2, 4],  [4, 6],  [6, ∞).
On each, we replace the curve q(k) by a rectangle (containing box) and use the minimum over box vertices. This is valid because the expression below is *affine in each coordinate* of the box.

**Lemma (Containing-box bound).** *For consecutive endpoints K, J of the partition, real row l = (a_{1}, ..., a_{5}, b_{1}, ..., b_{5}), and K le k le J (no upper bound if J = ∞), one has*
l q(k) ge [l]_{K, J} := min_{x in {1/K, 1/J}} ( a_{1} + x b_{1} + sum_{j = 2}^{5} min_{H in {K, J}} { (a_{j} + x b_{j}) e^{-H(n_{j} - 1)} } ).
*At J = ∞ use 1/J = 0 and e^{-J(n_{j} - 1)} = 0 for j > 1.*

**Sketch.** The physical vector q(k) lies in the box x = 1/k in [1/J, 1/K], w_{j} = e^{-k(n_{j} - 1)} in [e^{-J(n_{j} - 1)}, e^{-K(n_{j} - 1)}]. On this larger box l q is affine in each coordinate while others are fixed, so its minimum occurs at a vertex. First minimizing each w_{j} and then the single shared x gives the displayed formula. This is a *containing-box bound*, not sampling of the curve q(k). QED

### 7.5 Target support on every half-gap

The right side of the Bernstein polynomial must dominate the deleted quotient of T_{k}, not just of H_{i}. The target's own deleted quotient has its own Taylor coefficients a_{r}^{*}, and these are bounded uniformly over k.

**Lemma (Exponential supports).** *For consecutive parameter endpoints K, J and K le k le J, the target's deleted quotient satisfies, on every prescribed half-gap,*
D_{m, nu} T_{k}(m + y x) ge sum_{r = 0}^{40} a_{r}^{*} x^{r} - 10^{-6}  (0 le x le 1),
*with coefficients:*

- *At the origin (m = 0, nu = 0): a_{r}^{*} = (e^{K} / K) · (-Ky)^{r} / r!.*
- *At node 1 (and other isolated nodes with m = 1): a_{r}^{*} = K · (-Ky)^{r} / (r + 2)!.*
- *At node 3 (handling signed half-gaps): a_{0}^{*} = (J/2) e^{-13 J/6} when J < ∞; more complex expression at other y.*

**Sketch.** At the origin, T_{k}(s) = e^{-k(s-1)} / k, and ∂ T_{k} / ∂ k = T_{k}(s)(1 - s - 1/k) > 0 for k ge 2.36 > 2 and s le 1/2. So replacing k by K lowers the target. The Taylor polynomial through degree 40 is exactly (e^{K}/K)(-Ky)^{r}/r!, giving case 1.

At m = 1, the deleted quotient is D_{1, 2} T_{k}(1 + u) = (e^{-ku} - 1 + ku)/(k u^{2}), with value k/2 at u = 0. The k-derivative is (1 - (1 + ku) e^{-ku})/(k^{2} u^{2}) ge 0 by e^{x} ge 1 + x, so replacing k by K again lowers the target. The Taylor coefficients are K(-Ky)^{r}/(r + 2)!, giving case 2.

The exact tail after degree 41 is bounded by < 10^{-6} using scalar estimates and K le 6 in every parameter box (including J = ∞). QED

### 7.6 The bounded-range certificate

Combine the Bernstein rows with the target support:
Phi_{i}(x; k) = sum_{r = 0}^{40} ( l_{i, r} q(k) + delta_{i} a_{r}^{*} ) x^{r}.
Its Bernstein coefficients are
L_{i, d} q(k) + delta_{i} sum_{r = 0}^{d} omega_{d, r} a_{r}^{*}.

**Certification condition.** For every parameter box, every half-gap center m in N_{0} cap [0, 88], both signed half-gaps at each positive center, and every d in {0, ..., 40}, one has
L_{i, d} q(k) + delta_{i} sum_{r = 0}^{d} omega_{d, r} a_{r}^{*} > 0.001.
Here q(k) is evaluated at the box lower bound [l]_{K, J} (Section 7.4).

This is a finite rational check: 37310 separate inequalities, one per combination of (parameter box, half-gap, signed direction, d). The verification script `verification/numeric_balls.py` runs them.

**Conclusion of bounded-range signs.** When the certification condition holds, the Bernstein polynomial is > 0.001 on [0, 1]. Apply the three small-error bounds:
D_{m, nu} F_{i, slack}(m + yx) > .001 - .00011 - .000001 - .000001 = .000888 > 0.
So F_{i, slack} > 0 on [0, 89.5] off nodes, and = 0 at nodes. Done for the bounded range.

### 7.7 The tail: quadratic barrier for the sine product

Beyond s_{0} = 89.5, the number of nodes is infinite, so finite-row Bernstein certification cannot work. Instead, the paper uses the following *barrier*.

**Lemma (Quadratic barrier).** For every real s and every nearest point m in Z of the full periodic zero set (so lvert s - m rvert = min_{n in Z} lvert s - n rvert),
P(s) ge 0.68 (s - m)^{2}.

**Sketch.** P is periodic with period 12, has double zeros at Z, and no other real zeros. So it suffices to check the inequality on half-gaps. At each half-gap midpoint the exact values are computable:

| gap mod 12 | P(s)/(s-m)^{2} at its midpoint |
|---|---|
| [0, 1], [3, 4] | 12 - 8sqrt 2 > .6862 > .68 |
| [1, 3] | 1 > .68 |
| [4, 7], [9, 12] | (8)/(9)(3 + sqrt 2 + sqrt 3 + sqrt 6) > .68 |
| [7, 9] | 9 > .68 |

Logarithmic concavity of s |-> P(s)^{1/2}/(s - m) interpolates the endpoint values at nodes. QED

### 7.8 The tail signs

For s ge s_{0}, split each H_{i}:
R_{i, 0}(s) = C_{i} + sum_{n in N_{0},; n le 88} ( frac{c_{in}}{(s - n)^{2}} + frac{d_{in}}{s - n} ),
p_{i}(s) = P(s) R_{i, 0}(s) + p_{i, high}(s),
where p_{i, high} collects nodes n > 88 only.

**Step 1: rational part is positive.** Combine the intermediate values from Section 7.6 (with min(0, ·) floors on the lower evaluations), together with the reference correction. One gets
sigma_{i} R_{i, 0}(s) > .002  (s ge s_{0}).
The denominators s - n and (s - n)^{2} for n le 88 and s ge 89.5 are all ge 1.5, so all reciprocals are le 1 in modulus.

**Step 2: remainder second derivative is small.** Define
Z_{i}(s) = sigma_{i} ( p_{i, high}(s) + K_{p_{3 - i}}(s) ) + delta_{i} T_{k}(s).
A four-term Taylor bound using the derivative estimates of Section 4.6 gives
sup_{s ge s_{0}} lvert Z_{i}''(s) rvert < .0002.
The four terms are: direct high contribution (36pi^{2} · 2.4 · 10^{-7}), low-frequency transformed (3(3pi)^{2} · 36 · 5 · e^{-.70 s_{0}}), refined transformed (1.25(1.32pi)^{2}(2.54 · 5 / s_{0} + .04) e^{-.187 s_{0}}), and the target (2.36 e^{-2.36(s_{0} - 1)}).

**Step 3: Zed has exact double zeros at high nodes.** By the exact contacts, at every m in N, m ge 91, the decomposition
F_{i, slack}(s) = sigma_{i} P(s) R_{i, 0}(s) + Z_{i}(s)
has Z_{i}(m) = Z_{i}'(m) = 0 (the P · R_{i, 0} term vanishes to order 2 at m, so it must absorb all of F_{i, slack}'s jet there, but F_{i, slack} also vanishes to order 2 at m, so Z_{i} does too).

**Step 4: Taylor remainder bound.** For s ge s_{0}, let m be the nearest point of Z with m ge 91 (always possible in the tail by a nearest-neighbor pairing). Taylor's integral formula gives
lvert Z_{i}(s) rvert le .0001 (s - m)^{2}.

**Step 5: sign.** Combine:
sigma_{i} P(s) R_{i, 0}(s) ge .68 · .002, (s - m)^{2} = .00136 (s - m)^{2}.
Since .00136 > .0001,
F_{i, slack}(s) ge .00136(s - m)^{2} - .0001(s - m)^{2} = .00126(s - m)^{2} > 0.
At a node, the exact zero gives F_{i, slack}(m) = 0 (by construction). Done.

### 7.9 Completing the proof of Theorem 4.2

Combining Sections 7.6 and 7.8:
- On [0, 89.5] off nodes: F_{i, slack} > 0 (Bernstein certificate).
- At nodes: F_{i, slack} = 0 (exact contact).
- On [89.5, ∞): F_{i, slack} ge 0 (quadratic barrier).

So H_{1} le T_{k} and H_{2} ge 0 at every s ge 0 and every k ge 2.36. The four contacts of Proposition 6.1 complete the proof of Theorem 4.2 (the normalized Gaussian pair).

---

## 8. From Gaussian pairs to energy comparisons

The normalized pair gave us one specific function f(x) = e^{-pi h b lvert x rvert^{2}} H_{1}(b lvert x rvert^{2}) with widehatwidehat ff ge 0 and f le (the Gaussian of scale k-dependent). We now show how *any* function f with f le G_{alpha} (some target Gaussian) and widehatwidehat ff ge 0 translates into an energy lower bound.

### 8.1 The density linear program

**Proposition (density LP).** *Let f : R^{2} to R be an integrable even Schwartz function with widehatwidehat ff ge 0 pointwise. For every locally finite configuration C of centered disk density one,*
liminf_{R to ∞} frac{1}{N_{R}} sum_{substack{x, y in C_{R}  x ne y}} f(x - y) ge widehatwidehat f(0)f(0) - f(0).
*If also f le g pointwise for a function g : (0, ∞) to [0, ∞), then the same bound holds with g on the left:*

E_{g}(C) ge widehatwidehat f(0)f(0) - f(0).

**Sketch.** The proof is a Cauchy-Schwarz argument plus a background comparison, due to Cohn and Zhao (2014) and Cohn and de Courcy-Ireland (2018).

*Notation.* Let mu_{R} = sum_{x in C_{R}} delta_{x} be the point measure, and nu_{R} be the area measure restricted to B_{(1 + ε)R}, of total area V_{R} = pi(1 + ε)^{2} R^{2}.

*Positivity.* widehatwidehat ff ge 0 implies the sesquilinear form Q_{f}(mu, nu) = iint f(x - y) dmu(x) dnu(y) is positive semi-definite. So Cauchy-Schwarz applies:
Q_{f}(mu_{R}, mu_{R}) · Q_{f}(nu_{R}, nu_{R}) ge lvert Q_{f}(mu_{R}, nu_{R}) rvert^{2}.

*Cross-term estimate.* For each x in C_{R}, int f(x - y) dv_{R}(y) equals int f plus an error bounded by int_{lvert z rvert ge ε R} lvert f rvert, because the disk of radius ε R around x is contained in B_{(1 + ε)R}. So
Q_{f}(mu_{R}, nu_{R}) = N_{R} (widehatwidehat f(0)f(0) + o(1))
as R to ∞, uniformly in C through the boundary behavior.

*Self-term.* Q_{f}(nu_{R}, nu_{R}) / V_{R} to widehatwidehat f(0),f(0), by dominated convergence.

*Diagonal.* Q_{f}(mu_{R}, mu_{R}) = N_{R} f(0) + sum_{x ne y in C_{R}} f(x - y).

*Combining.*
frac{Q_{f}(mu_{R}, mu_{R})}{N_{R}} ge frac{(N_{R} (widehatwidehat f(0)f(0) + o(1)))^{2}}{N_{R} · Q_{f}(nu_{R}, nu_{R})} = frac{N_{R}}{V_{R}} · frac{(widehatwidehat f(0)f(0) + o(1))^{2}}{widehatwidehat f(0)f(0) + o(1)}.
N_{R}/V_{R} to (1 + ε)^{-2}. Taking liminf in R and then ε ↓ 0:
liminf_{R} frac{Q_{f}(mu_{R}, mu_{R})}{N_{R}} ge widehatwidehat f(0).f(0).
Subtracting f(0) gives the claimed bound on the off-diagonal sum. QED

**The crucial role of "density one".** Without density approx 1, N_{R}/V_{R} can be 0 or ∞, and the bound is vacuous. Centered disk density one is what makes the LP close.

### 8.2 Sharp pairs for every Gaussian parameter

The theorem of Section 4.2 gives the normalized Gaussian pair for k ge 2.36. This corresponds to a specific Gaussian target:
k = pi(alpha / b - h)  ⇒  alpha = b(k / pi + h).
So the target T_{k} gives a "G_{alpha} minorant" for every alpha ge alpha_{min} = b(2.36/pi + 2/5) approx 1.

For 0 < alpha < alpha_{min}, the paper uses **Fourier complementation** (Cohn-Miller duality): the already-constructed large-beta witness F_{beta} at beta = 1/alpha > 1 supplies the small-alpha minorant f_{alpha} = G_{alpha} - widehatwidehat F_F_{beta}.

**Lemma (Gaussian duality).** *For every alpha > 0, there is a real radial Schwartz function f_{alpha} satisfying*
f_{alpha}(x) le e^{-pi alpha lvert x rvert^{2}}  pointwise,  widehatwidehat f_f_{alpha} ge 0  pointwise,
f_{alpha}(a) = G_{alpha}(a)  for every  a in A ∖ {0},  widehatwidehat f_f_{alpha}(w) = 0  for every  w in A^{*} ∖ {0}.

**Proof sketch for alpha ge alpha_{min}.** Let k = pi(alpha / b - h) (so k ge 2.36), let f be the function of Theorem 4.2, and set f_{alpha} = k e^{-k} f. This scales f by a positive factor, so widehatwidehat f_f_{alpha} ge 0 still. The pointwise bound: with s = b lvert x rvert^{2},
f_{alpha}(x) = k e^{-k} e^{-pi h s} H_{1}(s) le k e^{-k} e^{-pi h s} T_{k}(s) = e^{-(k + pi h) s} = e^{-pi alpha lvert x rvert^{2}} = G_{alpha}(x).

The direct contacts come from the exact jets in Theorem 4.2, applied at every n in N, which contains every nonzero shell of A. The dual contacts come from the same contacts applied at every n in N, using that A^{*} = JA has the same shells as A (Section 1.1).

**Proof sketch for 0 < alpha < alpha_{min} (Fourier complementation).** Set beta = 1/alpha > 1/alpha_{min}. Let F_{beta} be the large-beta witness constructed above:
F_{beta} le G_{beta},  widehatwidehat F_F_{beta} ge 0,  F_{beta}(a) = G_{beta}(a) ; (a ne 0),  widehatwidehat F_F_{beta}(w) = 0 ; (w ne 0).

Now define
f_{alpha} := G_{alpha} - widehatwidehat F_F_{beta}.

This is the Fourier complement: f_{alpha} is the Fourier transform of G_{beta} - F_{beta} (using widehatwidehat{G_G_{alpha}} = G_{1/alpha} = G_{beta} and widehatwidehat Fhat_Fhat_{beta} = F_{beta}). Verify the four required properties.

*f_{alpha} le G_{alpha}.* By widehatwidehat F_F_{beta} ge 0, pointwise G_{alpha} - widehatwidehat F_F_{beta} le G_{alpha}. Done.

*widehatwidehat f_f_{alpha} ge 0.* By Fourier duality, widehatwidehat f_f_{alpha} = widehatwidehat{G_G_{alpha}} - widehatwidehat Fhat_Fhat_{beta} = G_{beta} - F_{beta}. By F_{beta} le G_{beta}, this is nonnegative. Done.

*Direct contacts.* For a in A ∖ {0}:
f_{alpha}(a) = G_{alpha}(a) - widehatwidehat F_F_{beta}(a).
We need widehatwidehat F_F_{beta}(a) = 0. Since F_{beta} is radial, widehatwidehat F_F_{beta} is radial too, so widehatwidehat F_F_{beta}(a) depends only on lvert a rvert. By hypothesis, widehatwidehat F_F_{beta}(w) = 0 for every w in A^{*} ∖ {0}. Since A^{*} = JA and J is a rotation, A and A^{*} have the same set of radial shells. So widehatwidehat F_F_{beta} vanishes on all those shells, including every a in A ∖ {0}. Hence f_{alpha}(a) = G_{alpha}(a). Done.

*Dual contacts.* For w in A^{*} ∖ {0}:
widehatwidehat f_f_{alpha}(w) = G_{beta}(w) - F_{beta}(w).
By hypothesis, F_{beta}(w) = G_{beta}(w) for every w in A ∖ {0}. Since A and A^{*} have the same shells and F_{beta}, G_{beta} are radial, F_{beta}(w) = G_{beta}(w) holds for w in A^{*} ∖ {0} as well. So widehatwidehat f_f_{alpha}(w) = 0. Done.

So f_{alpha} = G_{alpha} - widehatwidehat F_F_{beta} is the small-alpha witness. This is the Cohn-Miller duality in dimension two, with the specific technical point that the direct/dual contact transfer between A and A^{*} depends on their having the same *radial* shells.

**Why the same-shells fact is crucial.** The complement trick above transfers contacts from A^{*} (dual, where F_{beta} has zero Fourier) to A (direct, where f_{alpha} needs direct contacts). This transfer only works because both lattices are radial (so widehatwidehat F_F_{beta} on A equals widehatwidehat F_F_{beta} on A^{*} at equal radii) and their shells coincide. In a hypothetical scenario where A and A^{*} had different shells, the small-alpha case would require a genuinely different construction. QED

**Consolidation.** The lemma shows that for every alpha > 0, there is a minorant f_{alpha} le G_{alpha} with widehatwidehat f_f_{alpha} ge 0, and with exact contacts. This gives the density-LP inequality.

### 8.3 Poisson summation and the lattice value

The contacts are what make the LP bound *sharp*. Without contacts, the LP bound widehatwidehat f(0)f(0) - f(0) would be strictly less than the target lattice sum. With contacts at both A (direct) and A^{*} (dual), the LP bound equals the target.

**Lemma (Poisson and lattice value).** *For any even Schwartz f : R^{2} to R with widehatwidehat ff absolutely integrable (automatic on Schwartz functions),*
sum_{a in A} f(a) = sum_{w in A^{*}} widehatwidehat f(w).f(w).

**Consequence for f_{alpha}.** Write the sum splitting off the origin:
f_{alpha}(0) + sum_{a ne 0} f_{alpha}(a) = widehatwidehat f_f_{alpha}(0) + sum_{w ne 0} widehatwidehat f_f_{alpha}(w).
The dual contacts give sum_{w ne 0} widehatwidehat f_f_{alpha}(w) = 0. So
sum_{a ne 0} f_{alpha}(a) = widehatwidehat f_f_{alpha}(0) - f_{alpha}(0).
By the direct contacts f_{alpha}(a) = G_{alpha}(a) at every nonzero a in A:
sum_{a ne 0} G_{alpha}(a) = widehatwidehat f_f_{alpha}(0) - f_{alpha}(0).

So the LP bound is exactly the lattice sum. QED

The proof of the paper's Poisson summation lemma: the periodization p_{f}(x) = sum_{a in A} f(x + a) is well-defined and smooth by Schwartz decay. Its Fourier coefficients over the torus R^{2}/A are exactly widehatwidehat f(w)f(w) for w in A^{*}. Evaluating the Fourier series at x = 0 gives the identity.

**Only the lattice is periodized.** No periodicity is assumed of the *competitor* C; the LP applies to arbitrary locally finite configurations. This is central to the significance of the result. Periodization is used only to establish the target (LHS vs RHS) identity, not to restrict the competitor.

### 8.4 The Gaussian energy comparison

Combining:
1. The LP with f_{alpha} (Section 8.1),
2. The minorant f_{alpha} le G_{alpha} (Section 8.2),
3. The lattice value identity widehatwidehat f_f_{alpha}(0) - f_{alpha}(0) = sum_{a ne 0} G_{alpha}(a) (Section 8.3),

yields the following.

**Proposition (Gaussian energy comparison).** *For every alpha > 0 and every locally finite configuration C of centered disk density one,*
liminf_{R to ∞} frac{1}{N_{R}} sum_{substack{x, y in C_{R}  x ne y}} e^{-pi alpha lvert x - y rvert^{2}} ge sum_{a in A ∖ {0}} e^{-pi alpha lvert a rvert^{2}}.

Equivalently, E_{G_{alpha}}(C) ge E_{G_{alpha}}(A) for every alpha > 0.

This is the **Gaussian case** of the main theorem, proved in full. The extension to all completely monotone kernels takes the rest of Section 9.

### 8.5 The role of each ingredient

The four ingredients of the Gaussian case, and what each contributes:

| Step | Uses | Gives |
|---|---|---|
| LP (8.1) | widehatwidehat ff ge 0, density one | Energy lower bound via widehatwidehat f(0)f(0) - f(0) |
| Minorant (8.2) | Exact contacts of f_{alpha} | widehatwidehat f_f_{alpha}(0) - f_{alpha}(0) matches lattice sum |
| Poisson (8.3) | Periodization of lattice | Identity (target-side equality) |
| Contacts (8.2, 6.1) | Interpolation proof | Makes the LP bound *sharp*, not just a lower bound |

None of these pieces is optional. The LP would give a lower bound without contacts; the contacts make the bound tight. The interpolation machinery (Sections 4-7) is what enables the contacts. The Fourier structure (Section 4) is what enables the interpolation. Each is necessary.

---

## 9. From Gaussian to universal: positive mixtures

We now reduce the general completely monotone kernel g to the Gaussian case, by representing g (or rather g(ε + ·)) as a positive mixture of Gaussians. This is the classical Hausdorff-Bernstein-Widder theorem.

### 9.1 The positive mixture representation

**Theorem (Hausdorff-Bernstein-Widder).** *Let F : [0, ∞) to [0, ∞) be smooth and completely monotone: (-1)^{j} F^{(j)}(t) ge 0 for every j ge 0 and every t ge 0. Then there is a finite positive Borel measure rho on [0, 1] such that*
F(t) = int_{0}^{1} v^{t}, drho(v)  (t ge 0),
*with rho({0}) = 0 (no atom at the excluded endpoint v=0).*

The intuition: each v^{t} = e^{t log v} is itself completely monotone, and the completely monotone cone is exactly the positive cone spanned by exponentials with negative rate (i.e. by v^{t} for 0 < v le 1).

**Proof sketch.** Three steps.

*Step 1: discrete moments.* Fix d > 0. Repeated integration by parts gives (-1)^{r} Delta^{r} F(j d) ge 0, where Delta is the forward difference operator. Define a linear functional Λ on polynomials by Λ(x^{j}) = F(j d); positivity gives Λ a representing measure eta_{d} on [0, 1] with eta_{d}([0, 1]) = F(0).

*Step 2: refine to positive real exponents.* For l in ℕ, push eta_{d} forward by x |-> x^{l} (with d = 1/l). The result rho_{l} has mass F(0) and int v^{j/l} drho_{l} = F(j/l) for j ge 0. Take a weakly convergent subsequence; the limit rho = lim_{n} rho_{l_{n}} exists by compactness.

*Step 3: continuity in the exponent.* For fixed t > 0, let t_{n} = lceil l_{n} t rceil / l_{n}. Then lvert v^{t_{n}} - v^{t} rvert le (t_{n} - t)/(e t) to 0 uniformly in v in [0, 1]. So continuity of F forces int v^{t} drho = F(t).

Finally, the limit is finite on (0, 1] but the mass at 0 is zero (because v^{t} to 1_{(0, 1]} as t ↓ 0, and F(0^+) = F(0) gives rho((0, 1]) = F(0)). QED

**Conversion to Gaussians.** Each v^{t} for v in (0, 1] is
v^{t} = e^{t log v} = e^{-pi alpha t},  alpha = -(log v)/(pi),
which is exactly the Gaussian kernel G_{alpha}. So
F(t) = int_{0}^{1} G_{-log(v)/pi}(t), drho(v)
is a positive mixture of Gaussians.

### 9.2 Why we shift: singular kernels

The mixture above requires the representing measure rho to be finite, and this finite mass equals F(0). For a completely monotone kernel that is singular at 0, e.g. g(t) = t^{-s/2} with 0 < s < 2, F(0) = ∞ and the representing measure is infinite. So the direct mixture representation breaks down.

The paper circumvents this by *shifting* away from the singularity. Define, for ε > 0,
F_{ε}(t) = g(ε + t),  t ge 0.
This is smooth and completely monotone on [0, ∞) (being the composition of a completely monotone function with a shift), and F_{ε}(0) = g(ε) < ∞ for every singular kernel of interest. So F_{ε} has a finite positive representing measure rho_{ε} on [0, 1], with total mass g(ε).

The shift is removed as ε ↓ 0, in the limit of the energy argument. This is critical: without the shift, we cannot represent t^{-s/2} as a finite positive mixture of Gaussians. With the shift, we can, and the ε ↓ 0 limit recovers the singular behavior.

### 9.3 The mixture energy argument

Fix a kernel g : (0, ∞) to [0, ∞) that is smooth and completely monotone. For ε > 0, let rho_{ε} be the representing measure of F_{ε}(t) = g(ε + t), mass g(ε).

For a configuration C of centered disk density one, consider the shifted energy
E_{g(ε + ·)}(C) = liminf_{R} frac{1}{N_{R}} sum_{substack{x, y in C_{R}  x ne y}} g(ε + lvert x - y rvert^{2}).

A subtlety: the shift moves lvert x - y rvert^{2} up by ε, which for a pair at distance zero would give ε > 0 instead of 0. But zero-distance pairs are excluded from the off-diagonal sum; the shift is only felt at pairs with lvert x - y rvert^{2} > 0. For pairs at distance at least delta > 0, the shifted energy is close to the unshifted one as ε to 0. But pairs closer than delta can have large differences; that is why the limit ε ↓ 0 must be taken *after* taking R to ∞ and the lower limit in liminf.

**Key inequality (Fatou's lemma).** Write E_{g(ε + ·)}(C) as an inner product against the positive mixture. Fatou's lemma makes the liminf of a positive mixture ge the mixture of liminfs:
E_{g(ε + ·)}(C) ge int_{0}^{1} E_{G_{alpha(v)}}(C), drho_{ε}(v)
where alpha(v) = -log(v)/pi. The Gaussian case (Section 8) applies to each E_{G_{alpha}}, giving
E_{G_{alpha(v)}}(C) ge E_{G_{alpha(v)}}(A) = sum_{a ne 0} G_{alpha(v)}(lvert a rvert^{2}).
So
E_{g(ε + ·)}(C) ge int_{0}^{1} sum_{a ne 0} G_{alpha(v)}(lvert a rvert^{2}), drho_{ε}(v).

Tonelli's theorem moves the sum out:
= sum_{a ne 0} int_{0}^{1} G_{alpha(v)}(lvert a rvert^{2}), drho_{ε}(v).
Recognizing the inner integral as F_{ε}(lvert a rvert^{2}) = g(ε + lvert a rvert^{2}):
E_{g(ε + ·)}(C) ge sum_{a ne 0} g(ε + lvert a rvert^{2}).

This is the shifted version of the main theorem.

### 9.4 Removing the shift

As ε ↓ 0, the RHS sum_{a ne 0} g(ε + lvert a rvert^{2}) increases monotonically to sum_{a ne 0} g(lvert a rvert^{2}). The LHS E_{g(ε + ·)}(C) requires care, but the paper proves that
lim_{ε ↓ 0} E_{g(ε + ·)}(C) = E_{g}(C)
for every locally finite C of centered disk density one. The argument uses the singularity structure: for smooth completely monotone g, the singular behavior of g at zero is dominated by the ε ↓ 0 limit and can be controlled by the integral int_{0}^{1} v^{t} drho_{ε} against the pair density.

**Main Theorem (restated).** *Let g : (0, ∞) to [0, ∞) be smooth and completely monotone. For every locally finite configuration C of centered disk density one,*
E_{g}(C) ge E_{g}(A) = sum_{a in A ∖ {0}} g(lvert a rvert^{2}),
*interpreted in the extended nonnegative reals [0, ∞].*

### 9.5 The infinite case

When g(lvert a rvert^{2}) is infinite for the first shell a (e.g. g(t) = t^{-1}), the RHS is infinite. The theorem says: E_{g}(C) = ∞ for every density-one C. This is vacuous from the inequality direction but non-vacuous as a statement: it says that no configuration has finite energy for a strongly singular kernel, at density one.

The proof of this non-vacuous statement works by taking the limit ε ↓ 0 explicitly: for singular g, the right side becomes infinite as ε ↓ 0, and the LHS (which is ge the RHS by the mixture bound) is therefore also infinite.

### 9.6 Conclusion

We have established the full main theorem. To summarize the four-stage structure:

1. **Stage A (Section 4):** construct the entire family of candidate functions p_{a}, K_{p_{a}}, with explicit jet control.
2. **Stage B (Section 5):** reduce to finite data via M = 168, N = 384 quadrature points, and compute reference lists X_{i} q(k).
3. **Stage C (Sections 6-7):** correct to exact interpolation and prove the two sign conditions.
4. **Transfer (Sections 8-9):** turn the Gaussian pair into an LP lower bound, then mixture into universal bounds.

Each stage individually is elementary or numerically hardened; their combination proves the main theorem.

---

## 10. Renormalized Riesz and jellium energies

The main theorem's Gaussian comparison extends to the logarithmic and Riesz potentials by *renormalization*, a systematic procedure that neutralizes the divergent parts of the raw energy sum. This section develops the renormalized theory and proves Corollary 2.3.

### 10.1 The field energy setup

Fix 0 le s < 2 once and for all. The interaction is g(x) = lvert x rvert^{-s} for 0 < s < 2 and g(x) = -log lvert x rvert at s = 0. The basic normalizations:
kappa_{0} = 2pi,  kappa_{s} = 4pi  (0 < s < 2),
g_{0}(x) = -log lvert x rvert,  g_{s}(x) = lvert x rvert^{-s},  q_{0} = 2pi,  q_{s} = frac{2^{2-s}piGamma(1-s/2)}{Gamma(s/2)}.

**The weighted extension.** The potential g_{s} in the plane lifts to a *weighted extension problem* in R^{3}:
w_{s}(y) = lvert y rvert^{gamma},  gamma = s - 1  for  0 < s < 2,  gamma = 0  (so  w_{0} = 1)  at  s = 0.
Points (a, 0) in R^{2} times {0} are identified with their base-plane position a; the "charges" live on the plane y = 0.

**Admissible fields.** A field is a gradient E = nabla h in L^{p}_{loc}(R^{3}; R^{3}) for some p in (1, 2) (or more generally p in the range fixed at the start; see the paper), with w_{s} E in L^{1}_{loc}, satisfying the divergence equation
-div(w_{s} E) = kappa_{s} ( sum_{p in C} delta_{(p, 0)} - delta_{R^{2}} ).
Here the right side is the charge distribution: point charges at the particles, minus a uniform neutralizing background on the plane y = 0.

**Cutoff and truncated field.** For a self-energy cutoff eta > 0, define
f_{s, eta} = (g_{s} - g_{s}(eta))_{+},  E_{eta} = E - sum_{p in C} nabla f_{s, eta}(X - (p, 0)).
The function f_{s, eta} is the "capped fundamental solution": it equals g_{s} outside the cap and is constantly g_{s}(eta) inside. Subtracting it makes E_{eta} finite-energy (removes the singular part of E at each particle).

**Field energy.** For a centered square K_{R} = [-R/2, R/2]^{2}:
W_{s, eta}(E) = limsup_{R to ∞} [ frac{1}{R^{2}} int_{K_{R} times R^{k}} w_{s} lvert E_{eta} rvert^{2} - kappa_{s} g_{s}(eta) ].
A word about order of limits: for each fixed cutoff eta, we take R to ∞ first, then eta ↓ 0. The constant kappa_{s} g_{s}(eta) subtracts the self-energy of one particle (of a specific cutoff).

W_{s}(E) = lim_{eta ↓ 0} W_{s, eta}(E),  W_{s}(C) = inf_{E  compatible} W_{s}(E).

### 10.2 The periodic field and the Green function

Everything turns on the periodic picture. Fix a torus T = R^{2} / L of area N in ℕ (where N = lvert T rvert is an integer), and N distinct points a_{1}, ..., a_{N} in T.

**Canonical potential.** There is a mean-zero trace potential H on T times R^{k} satisfying
-div(w_{s} nabla H) = kappa_{s} ( sum_{i = 1}^{N} delta_{(a_{i}, 0)} - delta_{T} ).
Here delta_{T} denotes the uniform area measure on the base T (the neutralizing background). The identity understood in the weak sense: test against gradients nabla phi with phi compactly supported.

**Periodic energy.** Define
W_{s, L}(a_{1}, ..., a_{N}) = lim_{eta ↓ 0} [ (1)/(N) int_{T times R^{k}} w_{s} lvert nabla H_{eta} rvert^{2} - kappa_{s} g_{s}(eta) ],
with H_{eta} the truncated version of H (subtracting f_{s, eta} at each particle).

**Green function representation.** Let G_{L} be the mean-zero unit-source extension Green function:
-div(w_{s} nabla G_{L}) = delta_{(0, 0)} - N^{-1} delta_{T}.
Integration by parts gives the explicit formula
W_{s, L} = kappa_{s}^{2} [ (1)/(N) sum_{i ne j} G_{L}(a_{i} - a_{j}, 0) + lim_{x to 0} ( G_{L}(x, 0) - frac{g_{s}(x)}{kappa_{s}} ) ].
The first sum is off-diagonal (all pairs of distinct particles); the second is a limit that removes the diagonal singularity from the Green function's own self-contribution.

This follows the integration-by-parts argument of Petrache and Serfaty (2017, Proposition 1.5). In the log case (s = 0), G_{L} reduces to the ordinary 2D torus Green function.

**Comparison across period lattices.** A key issue for the corollary is: two periodic configurations C and D may not share a lattice, so their tori may have different sizes. The next lemma gives a comparison identity that works across arbitrary period lattices.

### 10.3 Heat-kernel comparison

Define Psi_{t}(x) = (4pi t)^{-1} e^{-lvert x rvert^{2}/(4t)}, the two-dimensional heat kernel. For a periodic configuration C = ∪_{i} (a_{i} + L) with N = lvert T rvert classes per unit cell, set
P_{t, L}(x) = sum_{v in L} Psi_{t}(x - v),  H_{C} = (1)/(N) sum_{i, j = 1}^{N} P_{t, L}(a_{i} - a_{j}) - Psi_{t}(0).

P_{t, L} is the heat-kernel Green function on T; H_{C} is (up to constants) the "self-energy at heat time t."

**Lemma (Heat comparison across period lattices).** *Let 0 le s < 2, and let C and D be simple periodic configurations of density one. Then*
V_{s}(C) - V_{s}(D) = frac{kappa_{s} q_{s}}{Gamma(1 - s/2)} int_{0}^{∞} ( H_{C} - H_{D} ) t^{-s/2}, dt,
*where V_{s}(C) = W_{s, L}(a_{1}, ..., a_{N}) is the canonical periodic value of C, and similarly for D.*calD.*

**Proof sketch.** The heat representation of the Green function on any flat torus is
G_{L}(x) = (1)/(Gamma(alpha)) int_{0}^{∞} ( P_{t, L}(x) - (1)/(N) ) t^{alpha - 1}, dt,
where alpha = 1 - s/2 for 0 < s < 2.

This is a spectral identity on the torus T; it applies to any flat torus. The individual H_{C} are uniformly bounded: near t = 0 by Ct^{-1} e^{-c/t} (from the Gaussian decay of Psi_{t} at nonzero image displacements), and near t = ∞ by Ce^{-lambda t} where lambda is the spectral gap of the torus Laplacian.

So the difference H_{C} - H_{D} is absolutely integrable against t^{-s/2} on (0, ∞), and the identity holds. QED

**Why the comparison is useful.** The key point: H_{C} is directly comparable to Gaussian energies! Setting g_{t}(u) = e^{-u/(4t)} at heat time t, one has
H_{C} = (1)/(4 pi t) E_{g_{t}}(C).
So the Gaussian energy comparison (Section 8) applies: H_{C} ge H_{A}(t) for every t > 0 and every periodic C of density one. Apply the lemma:

V_{s}(C) ge V_{s}(A),  for every periodic  C  of density one.

This is the periodic version of Corollary 2.3(i).

### 10.4 The infinite-system comparison

For a *nonperiodic* configuration C of density one, the periodic heat comparison does not directly apply: there is no period lattice. The paper extends the comparison via a *cubic approximation*: replace C by a square-periodic configuration on large cubes, prove the comparison there, and let the cube grow.

**Sketch.** Fix a large cube Λ_{R} = [-R, R]^{2}. Let T_{R} = R^{2} / (2R Z)^{2} be the torus of area (2R)^{2}. Replace C by the periodic configuration C_{R} obtained by wrapping points into the torus. This introduces boundary errors of size O(R) (from points near ∂ Λ_{R}), but these vanish relative to R^{2} in the field-energy normalization.

For each R, the periodic comparison gives V_{s}(C_{R}) ge V_{s}(A). Take R to ∞: on the left side, V_{s}(C_{R}) to W_{s}(E_{C} for a canonical field E_{C} obtained from the periodic canonical potentials H_{R} by a limit. On the right side, V_{s}(A/R Z) to W_{s}(A) trivially (the lattice repeats).

Combining: W_{s}(E_{C} ge W_{s}(A). Infimizing over E_{C} yields
W_{s}(C) ge W_{s}(A).

**Screening.** The field-energy lower bound ignores the possibility that C has "screening" structure: negative charge redistributions on the field side that improve the energy. The paper handles this by *screening*: construct a sequence of fields E_{j} compatible with a density-one configuration C_{j} whose renormalized energies approach W_{s}(A) from above.

The mechanism: take the canonical field E_{C} from the periodic limit, and modify it in shells of thickness ε R around each particle and near the cube boundary, to correct for the "screening defect"
V_{s}(C) - W_{eta}(E^{can}_{C} = (kappa)/(lvert T rvert) sum_{p in T} sum_{q ne p} q_{eta}(p - q) - 2kappa F_{eta},  F_{eta} = int_{R^{2}} f_{eta}(x, 0), dx.

The defect is nonnegative (stationarity), and vanishes as eta ↓ 0 for minimizers of W. Screening provides an explicit approximating field whose energy is V_{s}(C) up to o(1).

**Corollary (from Section 10.4).** *Under these definitions,*
W_{s}(A) = inf_{C  density one} W_{s}(C),
*attained at the triangular lattice A.*

The full argument for screening (Petrache-Serfaty 2017, Section 6) is technical: it constructs local correction fields in each shell, uses weighted trace and Poincaré inequalities, and proves that the sum of shell contributions is o(R^{2}) in the field-energy normalization. The paper writes out the construction in full.

### 10.5 The periodic torus inequality

**Corollary (Torus inequality).** *For every integer n ge 1, and every configuration of n^{2} distinct points in R^{2}/(nA),*
V_{s}(a_{1}, ..., a_{n^{2}}) ge V_{s}(a_{1}^{△}, ..., a_{n^{2}}^{△}),
*where a_{i}^{△} are the n^{2} classes of A/(nA) in the torus.*

**Proof.** Apply the heat comparison lemma (Section 10.3) with D = A and C the given torus configuration embedded periodically. The heat-kernel identity gives
V_{s}(C) - V_{s}(A) = frac{kappa_{s} q_{s}}{Gamma(1 - s/2)} int_{0}^{∞} (H_{C} - H_{A}(t)) t^{-s/2}, dt.
The Gaussian energy comparison (Section 8) gives H_{C} ge H_{A}(t) for every t > 0, so the integral is ge 0. QED

Note: repeating the argument with nA instead of A (and H_{A} unchanged because A is nA-periodic) shows that the inequality holds on the nA torus as well. So the triangular lattice minimizes W_{s} on every larger torus, not just the unit torus.

### 10.6 The jellium thermodynamic limit

The **jellium** is a finite system of N ordered pairs of charges (points plus neutralizing background) on √(N) times √(N) squares, with the *ordered-pair* interaction
J_{s, R}(a_{1}, ..., a_{N}) = sum_{i ne j} g_{s}(a_{i} - a_{j}) - 2 sum_{i = 1}^{N} int_{K_{R}} g_{s}(a_{i} - x), dx + int_{K_{R}} int_{K_{R}} g_{s}(x - x'), dx, dx'.

The three terms are: point-point interaction (excluding self), point-background interaction, and background-background interaction. All at scale √(N) (so N points in N area, density one).

The jellium thermodynamic minimum is
e_{Jel}^{(s)} = lim_{N to ∞} (1)/(N) min_{a_{1}, ..., a_{N} in K_{√(N)}} J_{s, √(N)}(a_{1}, ..., a_{N}).

**Corollary (Jellium minimum).** *For every 0 le s < 2,*
e_{Jel}^{(s)} = frac{W_{s}(A)}{kappa_{s}}.

The proof has three parts.

**Part 1: Periodic comparison.** For each N, minimize J over N points in K_{√(N)}. Confinement: points near the boundary can be moved inward without increasing the energy, because the fundamental solution g_{s} is monotone (attractive / monotone-positive). So the minimum is attained in a compact region and is finite.

**Part 2: Limit exists.** Using the periodic comparison V_{s} ge V_{s}(A) on the nA torus, one bounds the min-energy from below by N · W_{s}(A) / kappa_{s} - o(N). Using the periodic comparison plus a "widely spaced" approximation, one bounds it from above by the same thing. So the limit exists and equals W_{s}(A)/kappa_{s}.

**Part 3: Identification.** The classical thermodynamic results of Lewin-Lieb-Seiringer (2019, Section VI, Theorem 2) and Lauritsen (2021, Theorem II.1) supply the necessary existence and equality statements for the thermodynamic limit of the jellium. Combined with the periodic comparison, they identify the limit as W_{s}(A)/kappa_{s}.

This completes Corollary 2.3(ii).

### 10.7 Summary of Section 10

The renormalized/jellium results reduce to the Gaussian energy comparison via three moves:

1. **Heat representation.** Turn the renormalized energy V_{s} into an integral over heat time of Gaussian energies at rate 1/(4pi t), where the Gaussian case (Section 8) is available.
2. **Period lattice transfer.** Cross period lattices via the heat-kernel comparison lemma, using the fact that the heat kernel on any torus has the same spectrum but with different gaps.
3. **Screening + confinement.** Handle infinite configurations via a cubic approximation, and handle the jellium via confinement + periodic comparison + classical thermodynamic identification.

Each of these moves is standard in the crystallization literature; the paper's novelty is that the Gaussian comparison underlying move 1 (Section 8) is now *sharp*, thanks to the construction in Sections 4-7.

### 10.8 The shell-field decomposition

The "screening" move of Section 10.4 is a field-theoretic construction that must be fully specified in any complete proof. Here is the shape of the argument; the paper's Section ~6 gives the full details.

Fix the cutoff eta and a compatible field E_eta with W_ le m + delta (where m = W_s(A)). The target is: replace E_eta by a screened field widehatwidehat E_E_eta compatible with a *new* configuration with the same background and with energy no larger (up to o(R^2)).

The construction splits the ambient space R^{2+k} into five regions.

*Inner region.* D_0 = K'_R times [-ell, ell]^k, where K'_R is a concentric square inside K_R, its faces at distance ε R from the outer boundary. Inside D_0, the old field E_eta is retained.

*Cap regions* (only k = 1). Above D_0: K_R times (ell, R); below: K_R times (-R, -ell).

*Shell region.* The lateral sides of the shell between K'_R and K_R, subdivided into rectangular cells H_i ordered so each has area |H_i| in [1/m_i, 1] for integer-labeled m_i. On H_i, integer n_i = floor( m_i |H_i| ) particles are placed.

*Outside shell.* Subdivided further into auxiliary "rooted tree" cells, each of side a fixed small width b, so that eta-balls around retained centers distribute across these cells in a tree structure.

**The five component fields.**

- E_cross: completes the retained smeared charges across the inner boundary. On each auxiliary tree, the assigned outward flux through the face leading to the root is -kappa times the tree's charge mass. This makes the cell problems locally compatible.
- E_vert: a harmonic vertical correction h_2(y) = (ell^gamma)/(1-gamma) C^{pm} |y|^{1-gamma} in the caps, with source -ell^gamma (C^+ + C^-) delta_{R^2}
- E_side: on each shell cell H_i, a constant-outward-flux solution with source kappa(m_i - 1) + ell^gamma(C^+ + C^-) and datum -E_eta · nu_0 + j on the inner face. The kappa(m_i - 1) corrects for the "missing" background of the shell-cell.
- E_round: a rounding correction with source kappa sum_i delta_i |H_i|^{-1} delta_{R^2} where delta_i = n_i - m_i |H_i| (the rounding error at each cell). Its sum delta_i = 0, so it vanishes as a global source.
- E_new: the contributions of the N_new new particles placed at the barycenters of the n_i-decomposed rectangles, each with the truncated fundamental solution field around it.

Each component has an energy bound of the form

int w |E_cross|^2 le C(1+g(eta))(1+M_{R,eta'}) ε^{-1} R,
int w |E_vert|^2 le C e_{ε,R}, ε^3 R^2,
int w |E_side|^2 le C_{s,eta'} ε (M_{R,eta} + 1 + M_{R,eta'}) R^2,
int w |E_round|^2 le o(R^2),
int w |E_new|^2 le C(1 + g(eta)) N_new,  N_new = O(ε R^2) + o(R^2).

Here M_{R,eta} = R^{-2} Q_eta(E,R) is the truncated field energy per unit area. The constants depend on the fixed auxiliary cutoff eta' but not on eta, ε, R, or the realization. Summing these inequalities gives the combined bound

R^{-2} int w |widehatwidehat E_E_eta|^2 le M_{R,eta} + C_{s,eta'} ε(1 + g(eta))(1 + M_{R,eta} + M_{R,eta'}) + o_R(1).

The right side can be made as close to M_{R,eta} as desired (by choosing ε small), and then M_{R,eta} is itself close to m + delta by assumption. Letting eta, delta tend to zero gives W_s( E)E) = m.

### 10.9 The two-cutoff defect

The elimination of "coincident particles" (two points at the same location) is a crucial step for making the minimizer a genuine point set. The paper does it via a two-cutoff defect.

Let alpha < eta be two cutoffs. Since E_eta and E_alpha are both compatible fields (with the same source), they differ by a "cutoff difference":
E_eta - E_alpha = sum_p nabla(f_eta - f_alpha)(X - (p, 0)).

Stationarity of the natural field-energy functional under translations annihilates the divergence terms in a local energy identity. After separating self, distinct-labeled-pair, and background terms, one obtains

D_{alpha,eta} := a_alpha - a_eta + 2kappa (F_eta - F_alpha) ge 0,

where

- a_eta is the "mean truncated energy" at cutoff eta, morally E_P[|E_bbE_P[|E_eta|^2 dR^2];bbR^2];
- F_eta = int_{R^2} f_eta(x, 0), dx is the self-energy of one charged particle at cutoff eta;
- D_{alpha,eta} counts the energy of the ordered pairs of truncated kernels q_eta(p - q) for p != q.

**Collision detection.** Suppose a coincident pair exists at some location, in some configuration contributing to the ergodic average with positive probability. Let c_coll = E_P sum_{p} N_p(N_p - 1) count the coincidence intensity (number of colliding pairs per unit area, in expectation). Then the coincident-pair part of D_{alpha,eta} is kappa[g_s(alpha) - g_s(eta)] c_coll. Since g_s(alpha) to ∞ as alpha ↓ 0 for 0 < s < 2, and the right side of D_{alpha,eta} ge 0 stays bounded as alpha ↓ 0 (with only eta fixed), one gets c_coll = 0.

**Simplicity.** So the minimizer is *simple*: no two particles coincide. This is exactly the property one wants to reason about point sets, not "clouds of charges at shared locations."

**Identifiability.** Once simplicity is established, the field minimizer E_eta and its "collision-free" limit lim_{eta ↓ 0} E_eta can be treated as an actual gradient field compatible with a genuine point configuration. The minimizer is attained and its energy equals m = W_s(A). The selection of a specific realization uses the ergodic theorem: the defect density D_{alpha,eta} is a translation-invariant, integer-valued, mean-D_{alpha,eta} spatial density; the ergodic theorem gives almost-sure convergence to its mean, and a countable selection of "vanishing-defect" cutoffs gives a single realization with the desired properties.

---

## 11. The arithmetic verification

The paper's numerical inequalities are not folklore; they are certified. Every strict constant comparison has either a rational proof or a machine-checked enclosure. This section describes both.

### 11.1 Rational enclosures of elementary constants

Every terminating decimal is exact rational. Set ε = 10^{-80}. Let
p_{*} = 3.14159... (80 digits),  b_{*} = 0.8660254... (80 digits).

**Lemma (Decimal enclosures).** *For b = √(3)/2 and pi:*
p_{*} < pi < p_{*} + ε,  b_{*} < b < b_{*} + ε.

**Proof sketch for pi.** Machin's identity: pi/4 = 4arctan(1/5) - arctan(1/239). For 0 < x < 1, the alternating arctangent series has the finite-polynomial bound
A_{N}(x) < arctan x < A_{N}(x) + x^{2N+1}/(2N+1),
where A_{N}(x) = sum_{j=0}^{N} (-1)^{j} x^{2j+1} / (2j+1).

Taking N = 65 gives two explicit rational inequalities that enclose pi within p_{*} + ε from above and p_{*} from below. They are verified by clearing positive denominators in rational arithmetic.

**Proof sketch for b.** b = √(3)/2 iff b^{2} = 3/4. Squaring the rational endpoints gives the pair of rational comparisons b_{*}^{2} < 3/4 < (b_{*} + ε)^{2}. QED

For the Gaussian duality threshold k ge 2.36, a shorter rational chain suffices:
pi(2/√(3) - 2/5) > (157)/(50) · (1025838)/(1732500) = (1025838)/(433250) > (59)/(25) = 2.36.

**Exponential enclosure.** For rational a ge 0 and integer q ge 0, let P_{q}(a) = sum_{j=0}^{q} a^{j}/j!. If a > 0 and q + 2 > a, then
P_{q}(a) < e^{a} < P_{q}(a) + frac{a^{q+1}}{(q+1)!} · (1)/(1 - a/(q+2)).
This is used with q = 2000 for scalar comparisons and with a = 1, q = 6 for e < 2719/1000.

### 11.2 The scalar error budgets

Eleven strict inequalities are the "arithmetic backbone" of the sign and correction bounds. Each has a rational proof via Section 11.1.

| Bound | Value |
|---|---|
| U_{168}(5, .02) + 10^{-170} | < 8.3 · 10^{-14} |
| U_{91}(5, .02) + 10^{-89} | < 2.4 · 10^{-7} |
| Target tail c_{jet} e^{-2.36 · 168}/(1 - e^{-2.36}) | < 10^{-170} |
| Target tail c_{jet} e^{-2.36 · 90}/(1 - e^{-2.36}) | < 10^{-89} |
| Half-gap Taylor tails R(m, Y, nu) at four tuples | < 10^{-8} each |
| Target-support Taylor tails (two cases) | < 10^{-6} each |
| Second-derivative sup_{s ge s_{0}} lvert Z_{i}''(s) rvert | < .0002 |
| Product endpoint 12 - 8√(2) | > .68 |
| Bernstein margin .001 - .00011 - .000001 - .000001 | = .000888 > 0 |
| Low-node rational margin .0021 - 2 · 10^{-8} | > .002 |
| Tail-barrier product .68 · .002 | > .0001 |

Each of these is proved by a chain of rational inequalities, with pi and √(3) replaced by the rational enclosures p_{*}, b_{*} and, where needed, by 157/50 < pi < 22/7 and √(3) < 1733/1000.

### 11.3 The floor-rounded alternative

The paper also supplies a completely automatic rational procedure for verifying every finite matrix comparison, at the cost of introducing a rounding floor ε = 10^{-80} at every step of the arithmetic.

**Prescription (Appendix B, in brief).**

- `R(x)` = ε · floor( x/ε ), computed separately on real and imaginary parts.
- `e_star(w)` computes e^{w} via 12 squarings of the Taylor polynomial at w/4096, with `R(x)` applied at every arithmetic step.
- `W` (2311 × 171): mass matrix, exact rational + `R` floors. Computed once.
- `Ssharp`, `Ysharp`: 168 × 171 and 168 × 10, from `W` and jet projection, with `R` floors.
- `E1, E2`: 3 × 10, from the explicit constants of Section 5.5.
- `U, vp, vm`: power-sum doubling (9 stages), with `R` floors and strict bounds checked against the tables of Section 5.6.
- `X1, X2`: assembly into 171 × 10 reference lists.

The row-comparison procedure for signs (Section 7.6) becomes:

- For each parameter box, each half-gap center m in N_{0} cap [0, 88], both signed half-gaps at m > 0, both lists i = 1, 2, and d in {0, ..., 40}:
- Compute `L_{i, d}` as a rational row-vector `lsum = sum_{r=0}^d w_dr l_r` with w_{dr} = binom{d}{r}/binom{40}{r}, then evaluate `lsum` at the containing-box lower bound `[l]_{K, J}`.
- Compute `tsum` (target support) and add.
- Test `value > .001`.

Similarly, the tail-region check (Section 7.8) becomes: for each parameter box, each i, sum the sign-flipped low-node contributions and the reciprocal-scaled d_{n}, c_{n} rows, and test `value > .002`.

**Total comparisons.** 37310 half-gap checks + 10 tail checks. All strict and rational.

**Roundoff budget.** From Appendix B:

- Exponential error: < 10^{-68} (via a 12-stage squaring budget with amplification factor 3^{12} e^{7}).
- Single-moment error: < 10^{-48} after 41-term composition.
- Matrix power errors e_{r}: < 10^{-36} after 9 stages.
- Partial-sum error v_{9}: < 10^{-34}.
- Column error (X1, X2 assembled): < 10^{-29}.
- Bernstein row entry error: < 10^{-16}.
- Lower-evaluation error: < 10^{-6} (94 terms × 2 · 10^{-28}).

Combined budget: the floor-rounded verification, if it passes, transfers to the looser exact-array bounds with margin. For example, .00216 - 10^{-6} > .0021 and .00116 - 10^{-6} > .001.

### 11.4 The Arb verifier

Two Python programs verify the finite matrix data at high precision:

- `verification/arithmetic_bounds.py`: uses 512-bit Arb balls and exact rational arithmetic, checks scalar budgets and error-propagation majorants.
- `verification/numeric_balls.py`: uses 256-bit precision, encloses quadrature arrays, checks the power table and the sign tables.

Both check every strict comparison in every table. If all pass, the finite construction is certified.

**What the verifier does not do.** It does not implement the floor-rounded procedure of Section 11.3 end-to-end; the floor-rounding is a separate (redundant) rational proof. The user can run either verifier, and independently run the floor-rounded procedure via a short implementation described in Appendix B.

### 11.5 What certification actually proves

The certifier checks finitely many strict comparisons. Combined with Sections 5.2 (quadrature error < 10^{-31}), 4.6 (tail bound U_{r}), and 6.3 (operator bounds), this recovers the exact infinite-frequency statement of Proposition 6.1 with the 2 · 10^{-8} correction bound.

The four error sources in moving from finite to exact:

| Source | Bound |
|---|---|
| Quadrature → exact spectral | 10^{-31} |
| Power-sum truncation D^{512} | < 10^{-24} |
| Omitted finite nodes (target T_{k} loss) | < 8 · 10^{-12} |
| Tail (nodes n > 168) | < 8.3 · 10^{-14} |

The largest is 8 · 10^{-12}; multiplied through the Schur complement (lVert Q^{-1} rVert approx 1), the finite solution extends to the exact infinite solution with correction lVert ell - ell^{ref} rVert < 2 · 10^{-8}.

---

## Appendix A: Where the numbers come from

The specific values 168, 384, 89.5, and the constants in the tables are not chosen by magic; each is a threshold at which a rational tail estimate crosses a strict bound. The pattern is:

- M = 168: the point where U_{168} (tail error at node 168) crosses 8.3 · 10^{-14} for the reference list size used. Any larger M works too; 168 is a convenient round choice.
- N = 384 = 168 + 216: the Fejér quadrature order at which e^{.083 pi M} < 10^{83} (the Cauchy-derivative bound). Roughly half of M; not tight.
- s_{0} = 89.5: the smallest s where the "low-node" vs "high-node" cutoff is consistent with the reference-list bound w < 5. Determined by maximizing the test ratio mu · lambda / sup lvert Z_{i}'' rvert > 1 on the interval.
- Correction 2 · 10^{-8}: the sum of the finite-correction bound 1.37 · 10^{-8} (Section 6.7) and the tail-correction bound 8.31 · 10^{-14}, rounded up.

Every choice is a matter of tradeoff between the analytic simplicity of the bound and the numerical tightness of the tables. None is sharp.

## Appendix B: Reference for the floor-rounded algorithm

The floor-rounded procedure (Section 11.3) is spelled out in full in the paper's Appendix B. Its relation to the exact finite construction is: Appendix B is a redundant rational path that does not go through Arb; a reader who runs it and the Arb verifier gets an independent double-check.

The floor-rounded procedure uses 37310 half-gap comparisons + 10 tail comparisons. Its roundoff budgets are as in Section 11.3. The floor-rounded algorithm is fully written out in mathematical pseudo-code; a machine implementation in any language (Python, Sage, Julia, Magma) should reproduce the comparison outputs within the rigorous bounds.

## Appendix C: Statement of every hypothesis used

For ease of reference, every numerical hypothesis invoked anywhere in the paper:

1. b_{*} < √(3)/2 < b_{*} + 10^{-80}, p_{*} < pi < p_{*} + 10^{-80} (Lemma 11.1).
2. 157/50 < pi < 22/7 (for shorter numerical checks).
3. √(3) < 1733/1000.
4. e < 2719/1000.
5. Table values: power bounds (Section 5.6), column bounds, sign bounds, half-gap bounds.
6. The E_{1}, E_{2} specific values from Section 5.5 (extras).
7. k ge 2.36 throughout the construction (from the Gaussian duality threshold).
8. Reference list size lVert X_{1} q rVert + lVert X_{2} q rVert < 4.8.
9. Exact list correction bound lVert ell - ell^{ref} rVert < 2 · 10^{-8}.
10. Extras constant bound < .02.
11. s_{0} = 89.5.

Everything else is proved from these.