Home·The theorem·The proof in one reading·Architecture·Units·Manuscript·Glossary·FAQ·Work items·Data·Method·Provenance·Changes
← III.3 Free lattice kernels and covariances (notes)Part III·Unit outlineIII.5 Cluster expansion of the polymer gas (notes) →

III.4. Gaussian integration: change of Gaussian sector, determinants, window law — detailed notes

Contents

1. What this unit proves 2

2. What it uses 12

3. Strategy 13

4. Steps 13

4.1 Exact change of variables 13

4.2 Gap, decay and windows 15

4.3 Determinants 18

4.4 Gaussian moments and insertions 19

4.5 Jets 21

4.6 Large fields and constants 22

5. Where the difficulty lies 23

6. Notation 24

7. What remains unexplained 25

The full account of unit III.4: every proposition with its proof sketch, checked against the Lean in two rounds and edited. The condensed outline and the map of the Lean are on the unit page.

Unit page · All units · Overview

Part: III, Ultraviolet construction (renormalization group). Lean: about 40 work packages. The main ones are YM/SectorNext/GAUSSIAN_SECTOR_CHANGE/ (exact change of variables, gap, windows), YM/AnalyticBridge/RELATIVE_LOGDET_TRACE/ and YM/PhysicalNext/AFTERNOON_INTERACTING_RANK_COMPARISON/ (determinants and the polymer form of a sector), YM/AnalyticBridge/ANNULUS_RATIO_INSERTIONS/, YM/SectorTools/PHYSICAL_GAUSSIAN_INSERTION/ and YM/Primitives/CONNECTED_GAUSSIAN_CUMULANTS/ (Gaussian moments), SECTOR_AFFINE_JETS/, SECTOR_COVARIANCE_JETS/ and SELECTED_ORDERED_KERNELS/ (jets), LOCAL_GAUSSIAN_LAWS/, GAUSSIAN_FIELD_ALIGNMENT/, YM/Push/FREE_RESOLVENT/ and YM/RG08/Gaussian/ (locality, infinite-volume kernels), YM/SectorNext/LARGE_FIELD_REMAINDERS/, and the constant-producing statements in Next/STATIONARY, MergeProof and PhysicalNext/V9_JOINT_ENDPOINT_LOAD. Size: 37,956 lines in 4,737 declarations of the main-theorem closure.

Conventions. The parameter \(x\) is the coarse inverse-coupling parameter: in applications \(x = \hat{\beta}_m\). Throughout:

1. What this unit proves

Setting. One block-spin step \(S\) maps a fine torus \(\Lambda_f\) (side \(f = Lc\), block size \(L\) odd) to a coarse torus \(\Lambda_c\). The push-forward of the density at fixed coarse field \(V\) is decomposed (III.1, III.9) into sectors. A sector is given by:

The integration variables are the free links of \(F\), with product Haar measure. In each sector the unit splits the free links as follows (chill):

We write \(n_a = |\mathcal{A}_A | \dim \mathfrak{g}\).

On selected stencils, the step energy contains a sum of affine squares \(Q_Y (a) = \tfrac{1}{2} \lvert Ma + s (Y) \rvert^2\) (guide):

From these the unit defines three objects:

The unit proves four things:

Proposition III.4.1 (exact Gaussian change of variables in one sector). Fix the following data:

Let \(C\) be a matrix on the active coordinates and \(m\) a map from frozen arrays to active coordinates. Assume they are permitted (Permitted):

Let \(\Phi_{A, k} = \text{bill}_A \cdot \text{exposed}_R \cdot \text{enters}_{T, \beta, k, V}\) be the sector integrand. Here \(\text{exposed}_R\) is the interacting density of \(R\), and \(\text{enters}\) is the tagged block-spin kernel times \(e^{- \beta \cdot \text{gap energy}}\). Then the following hold (Licenses):

\(\displaystyle \int \Phi_{A, k} \hspace{0.17em} d \text{Haar}^F = \text{pref}_S (\beta) \hspace{0.17em} J_{\text{sc}}^{\hspace{0.17em} \lvert \mathcal{A}_A \rvert} \hspace{0.17em} (2 \pi g^2)^{n_a / 2} (\det C)^{1 / 2} \int \text{Haar}^{\mathcal{F}_A} (dY) \int \gamma_C (dz) \hspace{0.17em} \Psi (Y, z) .\)

Here \(\gamma_C = N (0, C)\). Set \(a = m (Y) + gz\), and let \(U\) be the configuration with frozen links \(Y\), active links \(e^{a_e} \nu_e\) and tree links \(1\). Then (dyed)

\(\displaystyle \Psi (Y, z) = \tilde{J}_d (a) \hspace{0.17em} \chi_d (a) \hspace{0.17em} \text{bill}_A \hspace{0.17em} \text{lik}_k (U) \hspace{0.17em} \text{exposed}_R (U) \hspace{0.17em} \exp \hspace{-0.17em} \left( - \beta \hspace{0.17em} \overline{\text{res}} (Y, U, a) - \beta Q_Y (m (Y)) + \eta \cdot z + \tfrac{1}{2} z \cdot (C^{- 1} - W) z \right) .\)

In this formula:

Equivalently, \(\text{pref} \cdot \Psi (Y, z) = e^{\frac{1}{2} z \cdot C^{- 1} z} \hspace{0.17em} r (Y, m (Y) + gz)\), where \(r\) is the integrand in Lebesgue log-coordinates (recolored). The identity holds for every permitted \((C, m)\). No quantitative locality assumption is needed.

Status: theorem.

Lean: authorized_change.

Proposition III.4.2 (licence of every committed RG step; the transfer identity). Let \(\mathscr{R}\) be a directed RG run (III.9) whose delays satisfy the rate conditions and the per-stage chart conditions (Starcash). Let \(m < N\), let \(p\) be a period, and let \(R\) be a representation at level \(m + 1\) on its selected history with \(\beta_R \in [\hat{\beta}_{m + 1} / 2, \hspace{0.17em} 2 \hat{\beta}_{m + 1}]\). Write:

Then:

\(\displaystyle \mathcal{S}_{\beta} (\rho_R) (V) = \sum_A \text{sieve} (V) \sum_k \text{pref} \hspace{0.17em} J_{\text{sc}}^{\lvert \mathcal{A}_A \rvert} (2 \pi g^2)^{n_a / 2} (\det C_{A, k, V})^{1 / 2} \iint \Psi \hspace{0.17em} d \gamma_C \hspace{0.17em} d \text{Haar} .\)

Here \(\mathcal{S}_{\beta}\) is the block-spin push-forward, \(\rho_R\) is the fine density, and \(\text{sieve} (V) = \exp (- \beta \cdot \text{orientation-averaged coarse energy of } V)\).

Status: theorem, conditional on the run, history and band hypotheses. III.4.12 supplies the rate and chart conditions; III.9 supplies the selected history.

Lean: committed, displaced_wrong.

Proposition III.4.3 (gap and locality of the actual precision and covariance). Fix \(G\), a patch \(P = (L, J)\) and a teller whose kernel satisfies the branch condition. There are \(v, \eta, b, E > 0\) and \(K \geqslant 0\) (post-\(L\) constants) with the following property. For large \(x\), take any step with patch \(P\) and coarse side \(c \geqslant 100 (1 + R (x))\), any funded tube, and any \(A, F, k, V\) with \(\text{flags} (A) = \text{flags}_{t (x)} (V)\). Then:

The constants do not depend on the volume, the store, the forest or the field.

Status: theorem.

Lean: opaqueCost.

Proposition III.4.4 (window law). Let \(\sigma, \eta > 0\), \(K \geqslant 0\), and fix a nonnegative function \(B : \mathbb{R} \to \mathbb{R}\). There are \(u, E, g > 0\) such that for every sufficiently large integer \(r\) the following holds, uniformly in the torus and the finite coordinate set. Let the coordinate positions \(p_j\) satisfy \(\sum_j e^{- v \hspace{0.17em} \text{dist} (i, p_j)} \leqslant B (v)\) for every coarse site \(i\) and every \(v > 0\) (Revenue). Let \(A\) be a finite symmetric matrix with:

Let \(L\) be the windowed inverse, \(L_{\cdot i} = (A|_{\mathcal{W}_i})^{- 1} e_i\) extended by zero, and \(A_{\mathcal{W}} = \tfrac{1}{2} (L + L^{\top})\) (clouded). Then:

A strengthened form (same_purchase) gives a rate \(s \in (0, \sigma]\) and a constant \(K_1\) such that \(A\), \(A^{- 1}\), \(L\) and \((A_{\mathcal{W}})^{- 1}\) all have \(s\)-weighted Schur norm \(\leqslant K_1\), and \(\lVert e^{s \hspace{0.17em} \text{dist}} (L - A^{- 1}) \rVert_{\text{Schur}} \leqslant K_1 e^{- sr}\).

Status: theorem. The constants depend only on \((\sigma, \eta, K, B)\).

Lean: Schur.windows.

Proposition III.4.5 (log-determinants as local path series). Let \(A \succ 0\) with windows \(\mathcal{W}\) (\(i \in \mathcal{W}_i\)). Let \(P_u\) be the column-wise windowed inverse of \(A + uI\), and \(\Delta_u = (A + uI) P_u - I\). Assume the following bounds (Priced), at a fixed nonnegative Schur rate, with \(C \geqslant 1\) and \(0 \leqslant q < 1\), for every \(u \geqslant 0\):

Assume also \(A_{\mathcal{W}} \succ 0\). Then

\(\displaystyle \log \det A = \sum_i \left( \ell_i + \sum_{n \geqslant 0} r_{i, n} \right), \qquad \log \det (1 + S) = \sum_i \sum_{n \geqslant 0} \frac{(- 1)^n (S^{n + 1})_{ii}}{n + 1}, \qquad \log \det A_{\mathcal{W}} = - \log \det A + \log \det (1 + S),\)

with:

All series converge absolutely. Moreover (labeled), the matrix products expand into closed index paths. A term \(r_{i, n}\) has one \(P_u\) factor and \(n + 1\) defect factors, hence \(n + 2\) jumps; the corresponding trace-series term has \(n + 1\) jumps. The absolute values, summed over the intermediate indices, are bounded by \(Cq^{n + 1}\) and \(q^{n + 1} / (n + 1)\), respectively.

Status: theorem.

Lean: Priced.chains.

For the relative expansions, the sector regularity conditions are: a lawful representation; matching fine paths and cutoff; a local chart with \(d > 0\), \(4 d < \rho_J\) and \(2 d < \rho_{\text{first}}\); positive coupling and sector width; \(0 < t_L (x) \leqslant 1\); output threshold \(t (x)\); and the tube's geometric, local-readability, predictor-derivative, covariance-positivity and gap conditions (Chips, Boosted).

Proposition III.4.6 (relative normalisation of a sector and its localisation). Let \(K_F\) be the flat kinetic precision: the Gram matrix of the move at \(V = 1\) with empty store, on all free links of \(F\).

(i) If the fine path family satisfies its path bounds and the kernel satisfies the patch's branch condition, \(\det K_F\) does not depend on the forest or on the gauge tags (Native.shared).

(ii) Assume:

Then on every sector point \(p = (A, F, k, V, Y, z)\) with matching flags, the positive density factor, with the sector masks and polymer product kept outside it, factorises as

\(\displaystyle \text{freight} (p) = \text{loading} (V) \hspace{0.17em} e^{\Lambda (p)}, \qquad \text{loading} (V) = \text{banker} (V) \hspace{0.17em} \text{pref} (\beta) \hspace{0.17em} e^{v_R \lvert \Lambda_f \rvert} \hspace{0.17em} \text{spare} (\beta)^{\lvert F^{\text{free}} \rvert} (\det K_F)^{- 1 / 2} > 0.\)

Here \(\text{freight} = J_{\text{sc}}^{\lvert \mathcal{A} \rvert} (2 \pi g^2)^{n_a / 2} (\det C)^{1 / 2} e^{\frac{1}{2} z \cdot C^{- 1} z} \hspace{0.17em} \tilde{J} \cdot (\text{noise}) \cdot e^{\text{exponent}_R}\), and \(\text{spare} (\beta) = (2 \pi g^2)^{\dim \mathfrak{g}/ 2} J_{\text{sc}}\) is the Gaussian normalisation per link. The exponent is

\(\displaystyle \Lambda (p) = - \beta \hspace{0.17em} \mathcal{E} (p) + \eta \cdot z + \tfrac{1}{2} z \cdot (C^{- 1} - W) z + \kappa (p) + \sum_e \log \tilde{J} (a_e) + (\text{tails}),\)
\(\displaystyle \kappa (p) = - \lvert \mathcal{F}_A \rvert \log \text{spare} (\beta) - \tfrac{1}{2} (\log \det W - \log \det K_F) + \tfrac{1}{2} \log \det (WC) .\)

Here \(\mathcal{E}\) is the orientation-averaged row energy, including the subtraction of the affine energy at \(a\) and its addition at \(m\):

\(\displaystyle \mathcal{E}= E_{\text{noise}} (U, V) + E_{\text{fine}}^{\text{cut}} (U) - E_{\text{coarse}}^{\text{cut}} (V) - Q_Y (a) + Q_Y (m) .\)

The tails are the remaining fine exponent after removing its volume constant and restoring its cut energy, together with the coarse cutoff correction. Thus \(\mathcal{E}\) is not merely a sum of affine squares.

(iii) Assume in addition the sector regularity conditions below, a nonnegative polymer exponent, and both path-counting bounds of Lemma 4.21. Then \(\kappa (p) = - \lvert \mathcal{F}_A \rvert \log \text{spare} + \sum_e \text{sect}_e + I + \text{Sh}\) (book_price), where:

(iv) Regrouping (rate_pass) gives \(\kappa = \sum_o \text{rent}_o + \text{far} + I + \text{Sh}\):

Status: theorem.

Lean: Dyed.cardinal_spin, Dyed.keptEarly, Dyed.spoolLift.

Proposition III.4.7 (polymer form of a sector, with large-field suppression). Fix:

Then for large \(x\) the following holds. Let \(y = \text{retime} (x)\), \(z = \beta_F (x, D.L)\) and \(\text{hot} (s) = e^{(\log \log s)^{3 / 2}}\). Eventually \(x \leqslant y \leqslant x^2\) and \(t (y) = x^{- 1 / 2} \text{hot} (x)\). Take a step with patch \(D\), a sector datum with chart \(J\), and positive cutoff parameter \(W_0\), subject to the following assumptions:

For such data there is a decomposition (Plate)

\(\displaystyle \Lambda (p) = \sum_{o \in \text{cores} (A)} \text{base}_o (p) + \sum_X \text{aux}_X (p)\)

whenever the flags match and the product of the fine-field, block, gauge-tag and chart cutoffs is nonzero. The bounds below hold for every flag-matched \(p\), including points outside that support. In it:

Status: theorem.

Lean: foreshadows, with hypotheses Helps.

Proposition III.4.8 (moments under the sector law). Fix \(G\). For every covariance cap \(L > 0\) and \(k \in \mathbb{N}\) there is \(D \geqslant 1\) with the following property. Take:

Then \(f = v \prod_{i \in Y} r_i \in L^1 (\mu \otimes \gamma_C)\) and

\(\displaystyle \int \lvert f \rvert \hspace{0.17em} d (\mu \otimes \gamma_C) \leqslant H \hspace{0.17em} D \hspace{0.17em} (1 + \lvert P \rvert \dim \mathfrak{g})^{2 k + 1} \hspace{0.17em} e^{- c \lvert J \rvert \rho^2} \hspace{0.17em} 2^{\dim \mathfrak{g} \hspace{0.17em} \lvert J \rvert} \prod_{i \in Y} \left( e^c \hspace{0.17em} 4^{\dim \mathfrak{g}} w_i \right) .\)

In a sector, take \(\mu = \text{Haar}^{\mathcal{F}_A}\) and \(C = W_{\mathcal{W}}\); III.4.3 supplies a uniform covariance cap. Moreover:

Status: theorem.

Lean: same_law.

Proposition III.4.9 (Gaussian insertions supported on selected coordinates). Fix \(p \geqslant 1\), a finite jet order, \(G\), a patch and a teller satisfying the branch condition. There are post-\(L\) constants \(K, M, s, b > 0\) and \(B \geqslant 0\) such that, for large \(x\), the following holds. Take a funded step with that patch, \(c \geqslant 100 (1 + R (x))\) and matching flags. Fix any \(g, d, \theta > 0\), frozen array and slot set \(U\). Restrict the linear and quadratic correction to \(U\), obtaining \(t \cdot z_U + \tfrac{1}{2} z_U \cdot \mathcal{J}z_U\), where \(t = \eta\) and \(\mathcal{J}= C^{- 1} - W\). Its tilted law is \(N (\mathcal{A}t_U, \mathcal{A})\), with \(\mathcal{A}= (C^{- 1} -\mathcal{J}_U)^{- 1} \succ 0\). The likelihood ratio of the \(U\)-marginals satisfies

\(\displaystyle \mathbb{E}_{\gamma_{C, U}} [(dN_U / d \gamma_{C, U})^p] \leqslant \exp \hspace{-0.17em} (2 p (\lvert U \rvert \ell^2 / (4 \theta) + 2 K \lvert U \rvert (\theta + \varepsilon / 2)))\)

whenever \(Kp (\theta + \varepsilon / 2) \leqslant 1 / 4\). The parameters are:

The support restriction concerns the Gaussian variable \(z\); the coefficients may depend on all the background and frozen data. The bound retains its dependence on \(g\), \(d\) and \(|U|\).

Status: theorem.

Lean: Physical.blocks, built on insert_letter.

Proposition III.4.10 (jets of the block coefficients and of the covariance). Fix \(G\), a patch \(P\), a teller satisfying the branch condition, and an order \(n\). There are \(r, d, v, \eta, \lambda_0, C_{\ast} > 0\) such that for large \(x\) the following holds. Take every funded step with patch \(P\) and \(c \geqslant 100 (1 + R (x))\), every \(A, F, V\) with matching flags, every \(k\), and every logarithmic perturbation \(B\) of the background with \(\lVert B \rVert \leqslant r\). The store, forest and tags remain fixed as \(B\) varies. Then:

\(\displaystyle \prod_{b \in \ell} e^{d \hspace{0.17em} \text{dist} (p_b, p_i)} \sum_j e^{d \hspace{0.17em} \text{dist} (p_i, p_j)} \lvert \partial_{\ell} X_{ij} (B) \rvert \leqslant C_{\ast},\)

and the analogous column bound holds with the external weights anchored at \(p_j\);

The selected predictor rows are real-analytic, with \(k\)-jets bounded by constants depending only on \((G, P, K, k)\), on a ball of fixed radius (sounds, chosen_branch). The precision \(W\) is also real-analytic (Lemma 4.27); the covariance and resolvent conclusions used here are finite-order jet bounds.

Status: theorem.

Lean: assembly, fromTeller, coverall, stores.

Proposition III.4.11 (affine squares in the large-field annulus). Fix a teller, a block size \(L\), \(W_0 > 0\), \(K \in \mathbb{R}\) and an integer halo \(D \geqslant 0\). There are \(C \geqslant 1\) and an eventually nonnegative function \(e (x)\) with \((1 + R (x))^n \beta_F (x, L) e (x) \to 0\) for every \(n \in \mathbb{N}\), such that for large \(x\) the following holds. Take a funded step of block size \(L\) with \(c \geqslant 100 (1 + R (x))\), stores \(O, A\), a forest \(F\), fields \(U, V\) and tags \(k\). Assume \(U\) is in tree gauge. In the \((D + 5) R (x)\)-halo of \(O\), assume the following local sector conditions: fine plaquettes whose support meets that halo and which belong to \(A\) have toll at least \(t_L (x) / 4\), those outside \(A\) have toll at most \(t_L (x) / 2\); block bonds whose support meets that halo and which belong to \(A\) have discrepancy at least \(W_0 t_L (x)\), those outside have discrepancy at most \(2 W_0 t_L (x)\); the coarse flags match \(V\); and the chosen frames have positive mass. Then there is a direction \(a\), vanishing on the tree links, such that for every pair of windows \(H, Q\) with \(H_{\text{face}} \subseteq Q_{\text{face}}\) and \(H_{\text{noise}} \subseteq Q_{\text{noise}}\), and with the face, noise and coarse rows of \(Q\) in \(\{p \not\in A^{(2 R)}, p \in O^{(DR)} \}\):

Here \(|O|\) is the number of cost cells, aisles and flags. The comparison with \(a'\) uses uniqueness of the logarithm on the selected stencils.

Status: theorem.

Lean: annulus_pay, with cleans_at and reacts.

Proposition III.4.12 (production of the step constants). Assume:

Then there are pre-\(L\) data with kernel \(K\) and the prescribed orders, two families of path prices, free-field and slope bounds, and a cost bound, with the following property. For every tariff and scope there is a sequence of block-size floors. After any increase of those floors and any choice of block sizes above them, there is a stagewise Hessian gap \(\eta\). For every compatible post-\(L\) cost choice (gap \(\eta\), activity cap at least \(1\), and slope bound below the chosen cap), protection and initial delays, the delays can be enlarged so that the rate and chart conditions hold for all further enlargements. On every paid run with these data there are stationary path families, including the full stationary estimates in the retained family. For each adjacent stage pair \((j, i)\) above the initial stage, the per-step chart conditions hold with \(a_{ji}, E_{ji}\) chosen after \(L_j\) (sponsored_permit). issue_cosmos replaces each entry value by the maximum of its old value and the thresholds for \((j, j)\) and \((j, j - 1)\), preserving the rate conditions.

Status: theorem under these kernel and order assumptions.

Lean: sailsV, using geared.

2. What it uses

The analytic tools are Gaussian affine changes of variables, the Haar Jacobian in exponential coordinates, Combes–Thomas estimates, the resolvent integral for \(\log\), Neumann series, Gaussian integration by parts and Fourier decay from analytic strips. General references are Helgason, Differential Geometry, Lie Groups and Symmetric Spaces; Glimm–Jaffe, Quantum Physics; and Katznelson, An Introduction to Harmonic Analysis. Balaban's fluctuation-integral papers and Dimock's The renormalization group according to Balaban provide related RG constructions; precise bibliographic locations remain open (§7).

The order of constants is pre-\(L\) data, block sizes, post-\(L\) bounds, then thresholds in \(x\). Statements with additional parameters retain those dependencies. The precision and path-counting certificates come from late_price, and the determinant-jet fund from Lemma 4.21. The packet-energy, chart, noise and representation hypotheses in III.4.7 are separate assumptions. The branch and order assumptions of III.4.12, and the rate assumptions of the scheduling statements in Lemma 4.33, remain explicit.

3. Strategy

Keep links within \(10 R (x)\) of the large-field store as Haar variables. Write the others in exponential coordinates about the coarse-field copy \(\iota V\), and separate the affine squares \(Q_Y (a) = \tfrac{1}{2} |Ma + s (Y) |^2\) from the remaining energy. The substitution \(a = m (Y) + gz\) and division by an arbitrary Gaussian density give the exact identity of III.4.1. Choose the symmetrised windowed inverse for \(C\) and the windowed mean for \(m\). The endpoint windows give finite range, at most \(\lfloor R (x) / 20 \rfloor\); the linear correction and \(C^{- 1} - W\) remain in the integrand.

A local Poincaré inequality in tree gauge gives the gap of \(W\). Predictor-derivative comparison permits absorption of the gauge errors at small threshold. Path decay gives a uniform weighted Schur bound; Combes–Thomas and the window law then give \(C = W^{- 1} + O (e^{- cR (x)})\) and positivity. The resolvent integral for \(\log\), followed by a Neumann expansion, expresses the relative Gaussian normalisation as single-window and path terms. Regroup these near the large-field cores. Their logarithmic cost is absorbed by the packet energy \(\beta s^2 \gtrsim \text{hot} (z)^2\), leaving exponentially suppressed core factors and a small polymer remainder on the cutoff support.

Gaussian moment bounds control the exponential corrections, marked jets control background derivatives, and the annulus estimates recover affine squares away from the store. The construction follows the fluctuation-integral method, with window errors retained explicitly. III.4.12 chooses the constants in the order needed along the run.

4. Steps

4.1 Exact change of variables

Lemma 4.1 (Gaussian density and scaling). For \(C \succ 0\), \(g > 0\), \(m \in \mathbb{R}^n\), and any Banach-valued \(f\) on \(\mathbb{R}^n\), \(f \in L^1\) iff \(z \mapsto e^{\frac{1}{2} z \cdot C^{- 1} z} f (m + gz) \in L^1 (\gamma_C)\), and

\(\displaystyle \int f = g^n (2 \pi)^{n / 2} (\det C)^{1 / 2} \int e^{\frac{1}{2} z \cdot C^{- 1} z} f (m + gz) \hspace{0.17em} d \gamma_C .\)

Lean's Bochner integral assigns value zero to a nonintegrable function, so the identity also holds under that convention without an integrability premise. Sketch: the density of \(\gamma_C\) follows by pushing \(N (0, 1)^{\otimes n}\) forward by \(C^{1 / 2}\) (cloud_formula); then translate and dilate.

Lean: Gaussian.gauged, at_price.

Lemma 4.2 (exponential chart on links). For a local chart ratio \(J\) and base links \(\nu\), the map \(a \mapsto (e^{a_e} \nu_e)_e\) sends \(\prod_e J_{\text{sc}} J (a_e) \hspace{0.17em} da_e\) on \(B_{\rho_J}^{\mathcal{A}}\) onto product Haar measure restricted to the image. Hence \(\int g \hspace{0.17em} d \text{Haar} = \int_{B_{\rho_J}^{\mathcal{A}}} \prod_e J_{\text{sc}} J (a_e) \hspace{0.17em} g (e^a \nu) \hspace{0.17em} da\) for strongly measurable \(g\) vanishing off the image, with equivalence of integrability. In the iterated form on the slice (frozen links Haar, active links in the chart), assume also that \(g\) is bounded; then every chart section and the outer integral are integrable.

Lean: log_transport, release_integral, thaw_integral.

Lemma 4.3 (regularised integrand). Assume the branch condition, a local chart ratio, the standing condition, \(d > 0\) and \(4 d < \rho_J\). For a multiplier with a bounded Borel envelope, the chart integrand equals \(J_{\text{sc}}^{\lvert \mathcal{A} \rvert} \tilde{J}_d (a) \chi_d (a)\) times the sector multiplier everywhere. Its sector integral equals \(J_{\text{sc}}^{\lvert \mathcal{A} \rvert} \iint r \hspace{0.17em} da \hspace{0.17em} d \text{Haar} (Y)\), with section-wise, outer and joint integrability. Sketch: on the support, \(\lvert a_e \rvert \leqslant d\) forces \(\chi_d = 1\) and each extended Jacobian to equal \(J\); off the support both sides vanish (protected_pay). The envelope supplies boundedness and measurability.

Lean: sheltered, regularized_price.

Lemma 4.4 (affine-square splitting). Under the fine path bounds, the averaged step energy equals \(\overline{\text{res}} + Q_Y (a)\) (compose_residual). For any \(C, m, z\) and \(g > 0\), with \(\beta = g^{- 2}\) and \(\eta = - g^{- 1} (Wm + M^{\top} s (Y))\):

\(\displaystyle - \beta Q_Y (m) + \eta \cdot z + \tfrac{1}{2} z \cdot (C^{- 1} - W) z = \tfrac{1}{2} z \cdot C^{- 1} z - \beta Q_Y (m + gz) .\)

Sketch: expand \(\tfrac{1}{2} \lvert M (m + gz) + s \rvert^2\) and use \(g^{- 2} = \beta\).

Lean: Rows.adjusted, recolored.

Lemma 4.5 (proof of III.4.1). Apply Lemma 4.1 to \(r (Y, \cdot)\) for each \(Y\), then Lemma 4.4 to identify the Gaussian integrand with \(\text{pref} \cdot \Psi (Y, \cdot)\). The same change applied to norms multiplies their Gaussian integral by the positive constant \(\text{pref} \hspace{0.17em} (2 \pi g^2)^{n_a / 2} (\det C)^{1 / 2}\). Lemma 4.3 and the Fubini criterion therefore give joint integrability. Inserting the section identity into the outer Haar integral proves the formula.

Lean: authorized_change.

Lemma 4.6 (admission of the chart). Fix a teller, local chart ratio, block size \(L\) and allocation width \(w > 0\). There are \(D, E > 0\) such that for large \(x\), with \(d = D \hspace{0.17em} t_L (x)\):

Sketch: the allocation gives toll \(\leqslant t_L (x)\) off the cost cells and block discrepancy \(\leqslant 2 w \hspace{0.17em} t_L (x)\) off the aisles. The axial-gauge estimate bounds the chord distance to \(\nu_e\) by \(m \hspace{0.17em} t_L (x)\). The logarithmic chart then gives \(\lvert a \rvert \leqslant Cm \hspace{0.17em} t_L (x)\).

Lean: matinee.

Lemma 4.7 (permits and the licence; proof of III.4.2). Under the fixed patch, branch, funding and volume hypotheses, combine Lemma 4.6, III.4.3 and continuity of the windowed mean. For large \(x\), matching flags and every \(\beta > 0\), \((W_{\mathcal{W}}, m)\) is permitted, \(W \succ 0\), and the selected predictor derivatives are bounded. Enlarge the entry thresholds to cover every admissible stage pair. Then:

The step size is \(d = a_{ji} t (\hat{\beta}_{m + 1})\), because \(t_L (\hat{\beta}_m) = t (\hat{\beta}_{m + 1})\).

Lean: sponsored_permit, committed.

Lemma 4.8 (identity completion). Assume \(W \geqslant \eta I\) with \(\eta > 0\) and \(R (x) \geqslant 120\). Complete \(W\) by the identity on the inactive positions (frozen and tree links), with the corresponding endpoint windows. The windowed inverse and its symmetrisation are the identity completions of \(W^{\text{per}}\) and \(W_{\mathcal{W}}\); the mean is extended by zero. The determinant is unchanged. This puts the actual and flat log-determinants on the same full link set.

Lean: Completion.stocked.

4.2 Gap, decay and windows

Lemma 4.9 (gap of the active precision). Under the patch, branch, funding and volume hypotheses of III.4.3, there is \(\eta = \eta (G, P, \text{teller}) > 0\) with \(\eta \lvert a \rvert^2 \leqslant a \cdot Wa\) for large \(x\), every flag-matched sector and every active vector \(a\). Sketch:

Lean: Gap.not_flat.

the derivative-comparison theorem itself states two things. For links within the predictor's reach that are \(\delta\)-close to \(1\), the derivative of \(a \mapsto\) curvature of \(\text{pred} (e^a U) \text{pred} (U)^{- 1}\) at \(0\) is \(\leqslant B \lVert a \rVert_{\ell^1 (\text{stencil})}\). It also differs from its value at \(U = 1\) by \(\leqslant \varepsilon \lVert a \rVert_{\ell^1}\). The constant \(\delta (\varepsilon)\) comes from continuity and is not explicit.

Lemma 4.10 (Combes–Thomas). Given \(\sigma, \varepsilon > 0\) and \(M \geqslant 0\), there are \(v \in (0, \sigma]\) and \(C = 2 / \varepsilon\) such that every symmetric \(A\) with \(\lVert e^{\sigma \hspace{0.17em} \text{dist}} A \rVert_{\text{Schur}} \leqslant M\) and gap \(\varepsilon\) satisfies \(e^{v \hspace{0.17em} \text{dist} (i, j)} \lvert A^{- 1}_{ij} \rvert \leqslant C\), in any volume. The same holds for all principal submatrices, since they inherit the gap. Sketch:

Lean: Schur.inverse_tilt, Schur.provides.

Lemma 4.11 (proof of the window law, III.4.4).

  1. \((AL)_{ji} = \delta_{ji}\) for \(j \in \mathcal{W}_i\), because column \(i\) of \(L\) solves the window system. So \(AL - 1\) is supported at distance \(> r\) and has unweighted Schur norm \(\leqslant e^{- ur} \cdot\)const.

  2. Hence \(L - A^{- 1} = A^{- 1} (AL - 1)\) has the same size, and so does its symmetrisation.

  3. The gap of \(A^{- 1}\) is at least \(\eta / (K + 1)^2\), because \(\lVert A \rVert \leqslant K + 1\). For \(r\) large the error is below half of it.

Lean: Schur.windows.

Lemma 4.12 (proof of III.4.3).

  1. Lemma 4.9 gives the gap.

  2. The path-decay and predictor-derivative bounds give weighted row and column bounds for \(M\); multiplying \(M^{\top} M\) gives the weighted Schur bound on \(W\).

  3. The counting function is \(B (v) = \dim \mathfrak{g} \cdot L^4 \cdot 4 \cdot \sum_{z \in \mathbb{Z}^4} e^{- v |z|_1}\) for \(v > 0\): each coarse site carries at most \(4 L^4 \dim \mathfrak{g}\) coordinates.

  4. The endpoint windows contain \(r_x\)-balls.

  5. III.4.4 gives (iii) and \(C \succ 0\).

Lean: opaqueCost. The decay certificate holds adds bounds for \(C^{- 1}\) and \(\lVert C^{- 1} - W \rVert \leqslant Me^{- aR (x)}\).

Lemma 4.13 (perturbation of inverses). Let \(s \geqslant 0\) and \(A, D \succ 0\) with \(\lVert D^{- 1} \rVert_s \leqslant K\), \(\lVert A - D \rVert_s \leqslant \delta\) and \(K \delta \leqslant \frac{1}{2}\). Then \(\lVert A^{- 1} \rVert_s \leqslant 2 K\). Sketch: from \(A^{- 1} = D^{- 1} - D^{- 1} (A - D) A^{- 1}\), a bound \(t\) on \(A^{- 1}\) improves to \(K + t / 2\); iterate from any finite-volume bound and pass to the limit.

Lean: inverse_perturb.

Lemma 4.14 (infinite-volume kernels).

(proof read for periodic extraction and the uniform flat inverse; statement only for the other assertions)

Lean: publish.

Lemma 4.15 (tree inverse and the free covariance).

Lemma 4.16 (covariance and locality of entries).

4.3 Determinants

Lemma 4.17 (proof of III.4.5).

  1. On the spectrum of \(A \succ 0\), \(\log \lambda = \int_0^{\infty} [(1 + u)^{- 1} - (\lambda + u)^{- 1}] du\), so \(\log \det A = \sum_i \int_0^{\infty} [(1 + u)^{- 1} - (A + u)^{- 1}_{ii}] du\).

  2. Split \((A + uI)^{- 1} = P_u + ((A + uI)^{- 1} - P_u)\).

  3. The first part gives \(\ell_i\), bounded by \((C + 1) C\) through \(\int (1 + u)^{- 2} = 1\).

  4. Expand the second part as \(P_u \sum_{n \geqslant 1} (- \Delta_u)^n\). Its negative enters the logarithm. The \(n\)th correction in III.4.5 has absolute integral at most \(Cq^{n + 1}\).

  5. For \(\log \det (1 + S)\) use the trace series; \(1 + S = AA_{\mathcal{W}}\).

Lean: Priced.chains, letterBase.

Lemma 4.18 (proof of III.4.6 (ii)–(iv)).

Lean: book_price, rate_pass.

Lemma 4.19 (local bounds on the pieces). Call a function \(\varepsilon\)-small if it is a sum of local terms \(\phi_X\) on flag-matched sectors where the cutoff product is nonzero, with \(\lvert \phi_X \rvert \leqslant b_X\) on every flag-matched sector and \(\sum_{X \ni y'} b_X e^{eN_B (X)} \leqslant \varepsilon\) at every site \(y'\). Here \(B\) is the specified cover radius. Under the corresponding hypotheses of III.4.7, for large \(x\):

(proof read for (a), (b), (d); statement and outline for (c), (e))

Lemma 4.20 (proof of III.4.7).

  1. On flag-matched sectors, \(\Lambda = \sum_o (\text{center}_o + \text{rent}_o) + \text{tenders}\) (issued). The center is \(- \beta\) times the row energy of III.4.6 summed over row positions within \(200 R (y)\) of \(o\) (center). The remainder consists of the terms of Lemma 4.19 (b)–(e), the row energy outside these halos, and the quadratic window correction.

  2. On the nonzero cutoff support, the packet-energy hypothesis gives \(\text{center}_o \leqslant - \beta as^2 \lvert o \rvert\) for each core of the store, with \(s = z^{- 1 / 2} \text{hot} (z)\). Since \(zs^2 = \text{hot} (z)^2\) and \(\beta \geqslant z / 2\), the condition \(2 \alpha \leqslant a / 2\) gives \(2 \alpha \hspace{0.17em} \text{hot} (z)^2 \leqslant \beta as^2\). The bound is ranked. Apply Lemma 4.19 (a) with \(\delta = \alpha\) to obtain \(\text{center}_o + \text{rent}_o \leqslant - \alpha \hspace{0.17em} \text{hot} (z)^2 \lvert o \rvert\).

  3. paltry uses the lower bound minimum to clamp the core term between fixed bounds, preserving its value on the nonzero cutoff support and its upper bound on all matching sectors.

  4. The tenders are \(\ell\)-small with \(\ell = 4 q + y^4 \nu + 2 \nu + a'\), where \(q = y^{- 1 / 8}\), \(\nu = y^{- 5}\) (from \(e^{- cR (y)} \leqslant y^{- 5}\)) and \(a' = y^{- 1 / 24}\). Then \(2 \ell \leqslant 16 y^{- 1 / 24} \leqslant y^{- 1 / 48}\) for large \(y\) (the power comparison).

  5. issue_invoice symmetrises over lattice motions and converts the small sum into \(\lvert e^{\text{aux}_X} - 1 \rvert \leqslant f_X\).

Lemma 4.21 (precision, counting and jet certificates). Fix \(G\), a patch and a teller satisfying the branch condition, a halo multiplier \(l \geqslant 1\) and a polymer exponent \(e \geqslant 0\).

4.4 Gaussian moments and insertions

Lemma 4.22 (exponentials of quadratic forms). Let \(x \sim N (0, I)\) and \(\mathcal{Q}= \sum_i w_i \langle q_i, x \rangle^2\) with \(w_i \geqslant 0\) and \(\sum_j w_j \lvert \langle q_i, q_j \rangle \rvert \leqslant 1 / 8\) for each \(i\). Then \(\mathbb{E}e^{\mathcal{Q}} \leqslant e^{2 \sum_i w_i \lvert q_i \rvert^2}\). Sketch: Gaussian integration by parts gives \(F' (t) =\mathbb{E} [\mathcal{Q}e^{t\mathcal{Q}}] \leqslant TF + \tfrac{1}{2} F'\), so \((e^{- 2 Tt} F)' \leqslant 0\).

Lean: authentic, amount.

Lemma 4.23 (polynomial moments). For \(L > 0\) and \(k \in \mathbb{N}\) there is \(D (L, k) > 0\), independent of \(n\), such that every \(0 \preceq C \leqslant LI\) on \(\mathbb{R}^n\) satisfies integrability and \(\mathbb{E}_{\gamma_C} (1 + \lvert z \rvert^2)^k \leqslant D (L, k) (1 + n)^{k + 1}\). Sketch: dominate a power by an exponential with \(t = 1 / (4 L)\), bound the sum of squared coordinates by \(n\) times its largest term, and use \(\mathbb{E}e^{tz_x^2} \leqslant 2\).

Lean: many.

Lemma 4.24 (proof of III.4.8).

  1. On \(\{v \ne 0\}\), \(1 \leqslant e^{- c \lvert J \rvert \rho^2} e^{c \sum_J \lvert z_i \rvert^2}\).

  2. The site budget gives \(\sum d_i (1 + \lvert z_{a_i} \rvert^2 + \lvert z_{b_i} \rvert^2) \leqslant \lvert Y \rvert c + 2 c \sum_X \lvert z \rvert^2\).

  3. Use \(ab \leqslant a^2 + b^2\) to bound the product of the polynomial score and \(e^{3 c \sum_U |z|^2}\). The exponential square has expectation at most \(2^{\dim \mathfrak{g} |U|}\), since \(6 c = 1 / (4 L)\) and \(|U| \leqslant |J| + 2 |Y|\); Lemma 4.23 bounds the squared polynomial score.

For the fourth-moment bound: \(\lvert e^q - 1 \rvert \leqslant xe^{2 x (1 + \lvert z_a \rvert^2 + \lvert z_b \rvert^2)}\), then apply III.4.8 with \(k = 0\). For the coupling-band estimate: the mean-shift equation writes \(\eta\) through \(W^{\text{per}} - W^{- 1} = O (e^{- uR (x) / 60})\), and \(C^{- 1} - W\) is controlled by the inverse-window estimate.

Lean: same_law, r4, oasis.

Lemma 4.25 (Gaussian likelihood ratios).

\(\displaystyle | \log T (z) | \leqslant \tfrac{\delta}{2} (n + |z|^2) + \tfrac{M}{2} (h^2 |u|^2 + |h| (|u|^2 + |z|^2)) .\)

Expand the shifted quadratic form and apply the Schur bounds (priceShift).

(proof read for the weighted moment and insertion bounds; statement only for the other assertions)

4.5 Jets

Lemma 4.26 (analytic predictor rows). Fix \(G\), a patch \(P\), a kernel \(K\) satisfying the branch condition and an order \(k\). There are \(C \geqslant 0\) and \(r, \delta > 0\) such that the following holds. Take a step with that patch and kernel, matching flags, stencil span at most \(R (x) / 50\), \(\text{allot} (P) + 1 \leqslant 9 R (x)\), \(2 (\text{allot} (P) + 1) < c\), and \(\text{permit} (P) t (x) \leqslant \delta\). On every selected block stencil, the predicted group element and its relative curvature are real-analytic on the \(r\)-ball of mixed coarse and fine logarithmic perturbations; the curvature has \(k\)-jets bounded by \(C\). Here \(\text{allot}\) and \(\text{permit}\) are fixed patch-dependent geometric constants. Sketch:

Lean: sounds, chosen_branch.

Lemma 4.27 (precision near the background). Under the fixed patch, branch, funding, volume and flag hypotheses of III.4.10, constants \(r, s, C_{\ast}, \eta > 0\) precede the threshold in \(x\). On \(\lVert B \rVert \leqslant r\):

For \(d \geqslant 0\) with \((1 + 2 k) d \leqslant \text{width} / 2\), the move-matrix entries have marked \(k\)-jets with the corresponding off-diagonal exponential weight. (proof read for the precision bound; statement only for the entry bound)

Lean: precision, entries.

Lemma 4.28 (transport along walks). Fix the jet order and a bound on the marked jets of the link field. For each \(\kappa > 0\) and endpoint jet cap \(M > 0\), there are a positive mark radius and \(N > 0\), uniform in the lattice, walk and scaling parameter \(s \in [0, 1]\), such that adjoint parallel transport moves a marked jet bound from the end of a walk of length \(\ell\) to its start, at base point \(0\), with cap multiplied by \(Ne^{\kappa \ell}\) (carried). For a strip-decaying kernel of rate \(\gamma > 0\), walks with length at most \(B (1 + |z|_1)\), \(B > 0\), finite sets of walks below every length cutoff, and the required endpoint bounds, choosing \(\kappa = \gamma / B\) gives a uniform marked bound on the truncated sum (follows). If the path family has strip bound \((\sigma, M)\) with \(M \geqslant 0\), \(d \geqslant 0\), \(\gamma > 0\) and \(2 \gamma + 2 kd \leqslant \sigma / 2\), selecting only paths of length at least \(R\) gives marked \(k\)-jets bounded by a constant times \(Me^{- \sigma R / 4}\) (transport_rows).

Lemma 4.29 (cut and window errors). Assume fine path bounds with positive rate \(\sigma\) and nonnegative amplitude, and the branch condition. Near the flat background \(B = 0\), the actual sector Gram matrix equals the reindexed path-cut affine block. If no free links are frozen, the tree-coordinate identification also identifies their geometric windowed resolvents (notarized). With a large step, \(R (x) \geqslant 120\) and the marked gap and counting bounds needed for the inverses, the following bounds hold at \(B = 0\) for \(u \geqslant 0\):

4.6 Large fields and constants

Lemma 4.30 (clean fields).

(proof read for the response estimate; statement only for the clean-field construction)

Lemma 4.31 (affine squares along paths). Assume fine path bounds with amplitude \(C\) and cutoff direct coefficient \(d_0\). Fix a path cutoff \(R_0 \in \mathbb{N}\), thresholds \(t, s\), \(B > 0\) and \(h \geqslant 0\). Suppose the face under consideration has toll at most \(t\) and satisfies the path-retention condition, its two coarse path rows have norm at most \(Bh\), and its first-order curvature and curvature error are at most \(Bh\) and \(Bh^2\). Require the same two curvature bounds at every face reached by a retained path of length at most \(R_0\), with chord deviation from the reference field at most \(h\) along every such path. Then the affine path series converges and work minus the extracted square is at least \(- 4 (mB)^2 (1 + 2 c_G R_0) h^3\), where \(m = \sqrt{d_0} + C + 1\) (fine_squares). Under the clean-field and annulus hypotheses of Lemma 4.30, each fine-face, noise and coarse row at a swept site has error at least \(- e (x)\), where \(e\) is eventually nonnegative and \((1 + R (x))^n \beta_F (x, L) e (x) \to 0\) for every \(n\) (coat).

Lemma 4.32 (proof of III.4.11).

  1. Sum the row bounds of Lemma 4.31 over \(Q\). The counted face, noise and coarse rows number at most \(X (1 + R (x))^4 |O|\), so the total error is \(e' (x) |O|\), where \(e' (x) = X (1 + R (x))^4 e (x)\) still satisfies \((1 + R (x))^n \beta_F (x, L) e' (x) \to 0\) for every \(n\).

  2. Retain the affine squares selected by \(H\) and discard the other squares by positivity. The additional coarse rows in the work are nonnegative.

  3. The extracted squares are unique, because the chart logarithm is unique once \(\max (K, C) t_L (x)\) is below the chart radius.

Lean: annulus_pay.

Lemma 4.33 (proof of III.4.12).

  1. Choose the free-field bounds.

  2. Choose the stationary width caps and path amplitudes, build a sheet within them, then choose drift bounds, slope caps and block-size floors. The stationary theorem gives a Hessian gap and full stationary path estimates only after the free-field, retained-family, paid-run, chart and pre-\(L\) compatibility hypotheses have been supplied. Its construction combines the one-loop analytic estimate with downward induction over levels.

  3. Build the startup and free-field data for that sheet, preserving the required orders and bounds.

  4. For each \(L\)-sequence the delays are enlarged and the chart-threshold enlargement is applied.

  5. Assemble the stationary path families and their estimates.

Related constant-producing statements:

(proof read for the main constant selection and stationary theorem; statement only for the ancillary scheduling and comparison results)

Lean: sailsV, geared, book, endures, post, takings.

5. Where the difficulty lies

The gap is a gauge-theoretic local Poincaré estimate: the predictor derivative must be compared with its flat-field value after a different gauge choice for each active link. Its smallness radius comes from continuity rather than an explicit formula. Inverse estimates then spend decay rate, while every bound must remain uniform in the volume, store, forest, tags and field.

The determinant comparison must accommodate different active dimensions, hence identity completion and the frozen-link normalisation. The resulting core costs grow polynomially in \(\log x\); the retimed large-field threshold supplies the stronger suppression \(e^{- \alpha \text{hot} (z)^2 |o|}\). The final powers \(x^{- 1 / 8}\), \(x^{- 1 / 24}\) and \(y^{- 1 / 48}\) allow the separate polymer budgets to be added. Gaussian transport and the affine-square identity are finite-dimensional; their application still requires the chart support and integrability conditions.

6. Notation

Written proof Lean (glossary) Meaning
\(\gamma_C\) Gaussian.cloud centred Gaussian with covariance \(C\)
\(z \cdot C^{- 1} z\) force precision quadratic form
\((2 \pi g^2)^{n_a / 2} (\det C)^{1 / 2}\) chartCost Gaussian normalisation
\(g = \beta^{- 1 / 2}\) coupling coupling of the representation
\(J\), \(J_{\text{sc}}\), \(\rho_J\) ChartRatio, LocalRatio normalised Haar Jacobian of exp, scale, radius
\(\mathcal{F}_A\), \(\mathcal{A}_A\) chill (Frozen, Active) free links within \(10 R (x)\) of \(A\) / the others
\(\iota V\) replica coarse field copied to the fine lattice
\(Q_Y (a)\), \(W\) Rows.guide, Rows.gram affine squares; precision \(M^{\top} M\)
\(C = W_{\mathcal{W}}\), \(m (Y)\) fog, meanLog windowed covariance; windowed mean
\(L\), \(A_{\mathcal{W}}\) Schur.per, Schur.clouded column-wise window inverse; its symmetrisation
Schur decay, gap Pours, Gapped weighted row and column sums; \(a \cdot Aa \geqslant \eta \lVert a \rVert^2\)
\(\Psi\) dyed sector integrand in Gaussian coordinates
permitted, licence Permitted, Licenses hypotheses and conclusions of III.4.1
\(\ell_i\), path letters, \(S\) letterBase, page, errata, excess window log letter; path terms; \(AA_{\mathcal{W}} - 1\)
\(\kappa (p)\), \(\Lambda (p)\) keptEarly, spoolLift relative log-normalisation; full log-integrand
\(\text{spare} (\beta)\) spare \((2 \pi g^2)^{\dim \mathfrak{g}/ 2} J_{\text{sc}}\)
\(\text{rent}_o\), far term, \(I\), \(\text{Sh}\) rents, nightly, introTax, showTax pieces of \(\kappa\)
\(\varepsilon\)-small Petty sum of local terms with polymer norm \(\leqslant \varepsilon\)
\(y^{- 1 / 48}\) garlic final polymer budget
\(R (x)\), \(t (x)\), \(t_L (x)\) radiusB, temp, fiery cluster radius, small-field threshold, fine-level threshold
\(M\), \(s (Y)\) move, source linear row map and frozen-link source
\(\mathcal{E}\), \(\text{center}_o\) Milk.levy, Coffer.center row energy remainder; its \(- \beta\)-weighted core sum
sector support Dyed.Playing nonzero product of field, block, tag and chart cutoffs
matching flags Matched \(\text{flags} (A) = \text{flags}_{t (x)} (V)\)
branch condition Patch.Branches uniqueness of selected minimisers and eligible axial gauge
sector regularity, noise bounds Chips, noises assumptions in III.4.6–III.4.7
\(c_G\) Gauge.colorScale matrix norm factor in the path estimate
price, bill, tax, rent, teller numerical bounds; Teller the step-estimator interface

7. What remains unexplained

  1. Proof depth. The supporting proofs marked “statement only” remain incomplete readings: the strip-resolvent and Fourier-inverse results in Lemma 4.14; the flat comparison in 4.15; 4.16; the path-counting and interaction estimates in 4.19(c),(e); 4.21; parts of 4.25; the entry bound in 4.27; 4.28–4.29; the clean-field construction in 4.30; 4.31; and the ancillary scheduling statements in 4.33. Their statements and the stated mechanisms have been compared with the cited source; these labels do not certify their dependency proofs.

  2. Run-level hypotheses. The packet-energy and support-exponent assumptions of III.4.7 are substantive inputs. Their simultaneous supply on the selected run, and the transfer from stationary path families to the tube's work rows, still need a complete account. The precision, determinant-jet and noise bounds have local suppliers, but their use does not establish these other premises.

  3. Free-kernel comparison. Lemma 4.15 compares the flat sector under the no-frozen-link and free-metric compatibility conditions. A direct statement comparing an arbitrary interacting sector covariance with the infinite-volume free operator was not located. The full expansion of the free-metric compatibility conditions and the nested marked inverse bounds in Lemma 4.29 remains to be written.

  4. Consumers and coverage. The downstream reset and repeatable-step statements have not been checked here. The additional assigned packages for Wilson bulk terms, vacuum stability, proof interfaces, multisource terms and flat RG data are not covered by these steps. The closure-size figures in the opening metadata were retained without recounting.

  5. References. The precise results and bibliographic locations in Helgason, Glimm–Jaffe, Katznelson, Balaban and Dimock remain to be selected and verified. The cited Lean declarations are the evidence for the statements above.

← III.3 Free lattice kernels and covariances (notes)Part III·Unit outlineIII.5 Cluster expansion of the polymer gas (notes) →