Home·The theorem·The proof in one reading·Architecture·Units·Manuscript·Glossary·FAQ·Work items·Data·Method·Provenance·Changes
← III.4 Gaussian integration: change of Gaussian sector, determinants, window law (notes)Part III·Unit outlineIII.6 The stationary branch: background field of the effective action (notes) →

III.5. Cluster expansion of the polymer gas — detailed notes

Contents

1. What this unit proves 2

2. What it uses 11

3. Strategy 13

4. Steps 14

4.1 Algebra of the finite gas 14

4.2 Tree-graph bound 15

4.3 Sums over trees 15

4.4 Cluster sums and the logarithm 16

4.5 Deleted regions and forced polymers 16

4.6 Packets and the large-coupling numerics 17

4.7 The logarithm by deletion 19

4.8 The sector decomposition 21

4.9 Matrices 22

5. Where the difficulty lies 24

6. Notation 25

7. What remains unexplained 26

The full account of unit III.5: 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: work packages YM/Push/ABSOLUTE_SUMS/ (3,624 lines), YM/AnalyticNext/CONNECTED_ACTIVITY_BOUNDS/ (3,589), YM/Push/SECTOR_MEASURE/ (1,153), YM/SectorNext/VACUUM_NORMALIZATION/ (935), YM/Progress/STATIONARY_SCENES/ (445). Size: 9,746 lines in 1,097 declarations used by the main theorem.

Conventions.

1. What this unit proves

The unit consists of five packages.

III.5.1–III.5.7, III.5.9, III.5.10, III.5.12 and III.5.13 are general implications. III.5.8(c), III.5.11 and III.5.14 concern the block-spin step under the funding hypotheses of III.1; III.5.14 also fixes a patch satisfying its branch conditions.

Proposition III.5.1 (Mayer identity for a finite gas).

Status: theorem. Lean: Gas.exact_polynomial; coefficient identities Coef.recurrence, Partition.partition.

Proposition III.5.2 (tree-graph bound). Let \(T\) be a graph on a finite vertex set with a distinguished vertex. Then \(| \sum_G (- 1)^{|G|} |\), summed over the connected spanning subgraphs \(G \subseteq T\), is at most the number of spanning trees of \(T\). In particular \(| \phi^T_R (z) |\) is at most the number of spanning trees of \(G_R (z)\). Status: theorem. Spanning trees are counted as acyclic parent maps towards the distinguished vertex. Lean: Graph.pay.

Proposition III.5.3 (pinned cluster bounds).

\(\displaystyle \text{Ready} (w ; a, \gamma, \delta, q) : \qquad \gamma > 0, \quad \delta > 0, \quad 4 \hspace{0.17em} \text{ball}_B \hspace{0.17em} \gamma / \delta \leqslant q < 1, \quad \text{Hit}_B (w ; \hspace{0.17em} a + \delta, \hspace{0.17em} \gamma) .\)

Then, for every site \(x\) and every integer \(m \geqslant 0\):

Each displayed bound is a summability statement of the form \(\text{SB}\). Status: theorem. Form (a) is used in III.2, III.10, III.11, III.12, III.13 and III.15. Lean: Ursell.log_hit_tail, Trees.touch_tail, Ursell.cluster_at, Ursell.cluster_tail; hypothesis Trees.Ready.

Proposition III.5.4 (convergence and logarithm of the gas). Let \(X, B\) be as in III.5.3, \(R\) reflexive and symmetric with \(R (i, j) \Rightarrow X_i \cap X_j \neq \emptyset\), \(w : \iota \to \mathbb{R}\) and \(a \geqslant 0\). Assume \(\text{Ready} (|w| ; a, \gamma, \delta, q)\). Then:

Status: theorem. Used in III.10. Lean: Gas.price_log, Gas.absolute, Gas.absolute_at.

Proposition III.5.5 (removing the polymers that meet a region). In III.5.4, take for \(R\) the overlap relation (\(R (i, j)\) iff \(i = j\) or \(X_i \cap X_j \neq \emptyset\)), with nonempty supports. For \(Y \subseteq \Lambda_n\) let \(w^Y_i = w_i \hspace{0.17em} \boldsymbol{1} [X_i \cap Y = \emptyset]\). Then

\(\displaystyle \Xi (w^Y) = \Xi (w) \hspace{0.17em} e^{- p}, \qquad |p| \leqslant \frac{|Y| \hspace{0.17em} \gamma}{1 - q},\)

where \(p\) is the sum of the cluster terms with \(X (z) \cap Y \neq \emptyset\).

Status: theorem. Used in III.10. Lean: Gas.restores_log, Gas.missing_tail.

Proposition III.5.6 (re-exponentiation with forced polymers). Fix \(B \in \mathbb{N}\) and a common lattice \(\Lambda_n\).

\(\displaystyle \frac{t \hspace{0.17em} e^{\mu} c}{1 - \sigma} + \frac{t \hspace{0.17em} \gamma}{(1 - \sigma) (1 - q)} \leqslant \eta \leqslant \sigma, \qquad 4 \hspace{0.17em} \text{ball}_B \hspace{0.17em} \eta / \delta' \leqslant q' .\)

Fix a site \(x\) and \(m \geqslant 0\). Consider finite sets \(s\) of atoms with multiplicities \(n_{\alpha} \geqslant 1\) such that:

Summed over all such pairs \((s, (n_{\alpha})_{\alpha \in s})\), the quantity

\(\displaystyle e^{A \hspace{0.17em} \text{cov}_B (\bigcup_{\alpha \in s} \text{supp} \hspace{0.17em} \alpha)} \prod_{\alpha \in s} \frac{\text{wt} (\alpha)^{n_{\alpha}}}{n_{\alpha} !}\)

is at most \(t^{- m} e^{- \mu} \hspace{0.17em} \eta / (1 - q')\).

Status: theorem. The general form, with an arbitrary selection predicate, is Multiple.estimates. Used in III.2, III.10, III.11 and III.13. Lean: OmitClusters.restore.

Proposition III.5.7 (gases of seed packets).

Setting. Let \(S\) be a block-spin step (III.1) from \(\Lambda_f\) to \(\Lambda_c\) with block side \(L\), and let \(l, B \in \mathbb{N}\).

\(\displaystyle \tau_L = L^4 \hspace{0.17em} \text{card} \{ \text{plaquette orientations} \} + 5, \qquad K_{\text{hit}} = (2 lB + 3)^4 \hspace{0.17em} \tau_L, \qquad K_{\text{step}} = (4 lB + 5)^4 \hspace{0.17em} \tau_L .\)

(a) Packet sums. Assume \(a, C \geqslant 0\), \(0 \leqslant p_A \leqslant C \hspace{0.17em} e^{(s - g) |A|}\) for every connected \(A\), and \(\rho := K_{\text{step}}^2 \hspace{0.17em} e^{\hspace{0.17em} s - g + a (2 l + 3)^4} \leqslant q < 1\). Then for every coarse site \(x\) and every \(m \geqslant 0\),

\(\displaystyle \sum_{A \text{connected}, x \in D (A), |A| \geqslant m} p_A \hspace{0.17em} e^{a \hspace{0.17em} \text{cov}_B (D (A))} \leqslant K_{\text{hit}} \hspace{0.17em} C \hspace{0.17em} \frac{q^{\max (1, m)}}{1 - q} .\)

The same bound holds for an arbitrary family of variants with nonnegative weights, each nonzero variant assigned to a connected packet and supported in its dome, provided the sum of the variant weights of each packet is at most \(Ce^{(s - g) |A|}\). For variants, the sum is pinned by the actual support and uses its covering number.

(b) Mixed words.

\(\displaystyle \sum_s \left( \prod_{\alpha \in s} \text{wt} (\alpha) \right) (1 + \text{ch} (s))^k \hspace{0.17em} e^{a \hspace{0.17em} \text{cov}_B (D (s))} \leqslant \frac{k! \hspace{0.17em} e^{\varepsilon}}{\varepsilon^k} \hspace{0.27em} e^{- \varepsilon T - \mu_s r} \hspace{0.27em} \frac{\gamma_{\text{root}} \hspace{0.17em} q^m}{1 - q},\)

the sum over connected \(s\) with \(x \in D (s)\), \(|s| \geqslant m + 1\), at least \(r\) seeds in total, and \(\text{ch} (s) \geqslant T\). Packets that are not connected have weight \(0\).

Status: theorem. The hypotheses of (b) are the records Mixed.Costs.Buy and Mixed.Accounts. (b) is used in III.2 and III.10–III.13. Lean: Packets.packet_income, Packets.packet_variants_tail, Mixed.mixed_hit, Mixed.ready_vertices.

Proposition III.5.8 (the numerical conditions hold at large coupling).

Notation. Fix the gauge group. In (a) and (b), also fix the block side \(L\) and \(l \in \mathbb{N}\). For the coarse coupling \(x\) put:

(a) Let \(a, D \geqslant 0\), \(d \in \mathbb{N}\) and \(z < 2\). For all sufficiently large \(x\) the following holds. Whenever \(s \leqslant D \hspace{0.17em} \text{bill}^d P (\beta_F)^z\), \(g \geqslant M\) and \(0 \leqslant C \leqslant \exp (D \hspace{0.17em} \text{bill}^d P (\beta_F)^z)\):

If \(\beta \geqslant u \beta_F\), then \(\beta \mu t^2 \geqslant M\).

(b) Let in addition \(\varepsilon, \delta > 0\). For all large \(x\), all \(\beta \geqslant u \beta_F\), and all \(C, e_s \geqslant 0\), \(s \in \mathbb{R}\), and \(\gamma_{\text{opt}}\) with

\(\displaystyle C \leqslant e^{\text{fee}}, \qquad s + 2 \varepsilon e_s \leqslant \text{fee} := D \hspace{0.17em} \text{bill}^d P (\beta_F)^z, \qquad 0 \leqslant \gamma_{\text{opt}} \leqslant \delta / (16 \hspace{0.17em} \text{ball}_B),\)

the numerical inequalities of III.5.7(b) hold with:

The packet envelope and optional site-sum estimate in III.5.7(b) remain hypotheses on the actual weights.

(c) For a step-estimator \(\mathcal{T}\) (III.1), there is \(n\) such that for every \(L \geqslant n\) and all sufficiently large \(W\) there is \(\mu > 0\), chosen before \(l, d, u, a, D, z\), with the following property for every choice of \(l, d, u, a, D, z\) satisfying (a). For all sufficiently large \(x\), every funded step with block side \(L\) and coarse side \(\geqslant 100 (1 + B)\) satisfies the point charge of III.5.11 with constant \(\mu\). For every \(\beta \geqslant u \beta_F\), the numerical conclusions of (a) hold with \(g = \beta \mu t^2\). The threshold in \(x\) may depend on the fixed numerical parameters but not on the volume.

Status: theorem. (c) takes the energy grant per seed from III.10. Used in III.10. Lean: Delay.fits, Delay.comparison, Buy.quoted, Grant.order.

Proposition III.5.9 (logarithm by deletion).

(a) Abstract form.

\(\displaystyle \sum_{i : \hspace{0.17em} x \in X_i} w (X_i) \hspace{0.17em} v_i \hspace{0.17em} e^{a |X_i |} \leqslant \delta .\)

(b) At a coarse field.

Status: theorem. (b) is used in III.11. Lean: ReLog.catalog, ReLog.omitted, ReLog.swept, ReLog.intake, ReLog.retest, ReLog.retain_mask.

Proposition III.5.10 (trace-log of a near-identity product). Let \(A\) be positive definite and \(C\) symmetric (real matrices), and \(S = AC - 1\) with \(0 \leqslant q < 1\) and \(\sum_j |S_{ij} | \leqslant q\) for every \(i\). Then:

Status: theorem. Used in III.4. Lean: Round.exact_trace, Round.twin_log.

Proposition III.5.11 (sector decomposition of a block-spin step; gain per seed).

Setting. Fix a step \(S\) with block side \(L\), a step-estimator \(\mathcal{T}\), a tube \(T\) with \(\text{Funds} (S, T)\) (III.1), a coarse coupling \(x\) with threshold \(t\), and constants \(W, \kappa > 0\).

Sector of a packet. Say \((U, k, V)\) satisfies the sector prerequisites of \(A\) when:

Point charge. For every tree-gauge forest \(F\), every fine field \(U\) equal to \(1\) on the trunks of \(F\), every frame tag \(k\), every coarse field \(V\) and every packet \(A\) such that \((U, k, V)\) satisfies these prerequisites:

\(\displaystyle \mathcal{E}_A (U, V) \geqslant \kappa \hspace{0.17em} t^2 \hspace{0.17em} |A| .\)

Here \(\mathcal{E}_A\) is the assessed energy of \(A\): the sum of the fine, block-constraint, coarse and cross terms of III.1 with anchors in \(D_{2 B} (A)\), using walks truncated at \(\lfloor B / 100000 \rfloor\).

Claim. Let \(R\) be a lawful representation of the fine density (III.1) adapted to \(T\): its path data and clip are those of the upper stage of \(T\), and its threshold is \(t\). Let its coupling satisfy \(\beta \geqslant 0\). Let \(F\) be any tree-gauge forest and \(D = (D_A)_A\) any allocation: \(D_A (k, U, V) \geqslant 0\) is Borel, \(\sum_A D_A = 1\), and \(D_A \ne 0\) implies the two seed tests above for every fine cell and coarse bond, together with equality of the flag sets on the whole coarse lattice. Positive likelihood then implies the local sector prerequisites. This global support condition is needed to apply the point charge in every lattice orientation. Then for every coarse field \(V\)

\(\displaystyle (\text{push}_{\beta} \hspace{0.17em} \rho_R) (V) = \sum_A c_A (V), \qquad |c_A (V) | \leqslant \pi_{\beta} \hspace{0.27em} e^{- \beta \kappa t^2 |A|} \hspace{0.27em} \ell_A (V),\)

where:

Uniformity.

Status: theorem, conditional on \(\text{Funds}\) and the point charge. The point charge is discharged in two ways:

Used in III.4 and III.10. Lean: Presented.portionCost (conclusion PortionCost), Consumer.bought, points_grant; definitions Seeded, Alive, PointCharge, earn, Allocation.

Proposition III.5.12 (Combes–Thomas bound). Let \(M\) be a real symmetric matrix on a finite index set, with positions \(a_i \in \Lambda_n\). Assume \(M \geqslant r > 0\), \(\eta \geqslant 0\), and

\(\displaystyle \sup_i \sum_j e^{\eta d (a_i, a_j)} |M_{ij} | \leqslant C, \qquad \sup_j \sum_i e^{\eta d (a_i, a_j)} |M_{ij} | \leqslant C.\)

If \(0 \leqslant \theta \leqslant 1\) and \(\theta C \leqslant r / 2\), then for every \(u \geqslant 0\) the matrix \(M + uI\) is invertible and

\(\displaystyle e^{\theta \eta \hspace{0.17em} d (a_i, a_j)} \hspace{0.17em} | (M + uI)^{- 1}_{ij} | \leqslant \left( \min (1, r / 2) \hspace{0.17em} (1 + u) \right)^{- 1} \quad \text{for all } i, j.\)

The hypotheses pass to principal submatrices. Status: theorem. Lean: Green.inverse, Green.Bottom.restrict, Green.Rows.restrict.

Proposition III.5.13 (windowed inverses).

Setting.

Objects.

(a) Weighted row and column sums at rate \(\mu\) satisfy:

At any rate \(0 \leqslant \xi \leqslant \eta\), if \(\text{within} - A^{- 1}\) has row and column bounds \(s' \geqslant 0\) and \(\text{within}\) has bounds \(Y' \geqslant 0\), then \(\text{cov}\), \(\text{cov} - A^{- 1}\) and \(A \hspace{0.17em} \text{cov} - 1\) have bounds \(Y'\), \(s'\) and \(Cs'\), respectively. If also \(C, Y' > 0\) and \(s' \leqslant 1 / (2 C)\), then \(\text{cov} \geqslant 1 / (2 C)\) and \(\text{cov}^{- 1} \geqslant (Y')^{- 1}\).

(b) Fix positive bounds on the weighted row and column sums, spectral gap, rate \(\alpha\), and counting constant. There are \(0 < e \leqslant \alpha\) and \(K, c, c_1, c_2 > 0\), depending only on these data, such that for all sufficiently large integer \(R\) the following holds uniformly for all such matrices and windows with room \(R\). At rate \(e\), the full and windowed inverses of \(tA + (1 - t) I\) have bounds \(K\), and their residual and inverse difference have bounds \(Ke^{- cR} < 1\), for every \(t \in [0, 1]\). At \(t = 1\), the covariance has bound \(K\), its difference from \(A^{- 1}\) and the residual \(A \hspace{0.17em} \text{cov} - I\) have bounds \(Ke^{- cR}\), and \(\text{cov} \geqslant c_1\), \(\text{cov}^{- 1} \geqslant c_2\).

Status: theorem. Lean: Win.Windows.window_price, covariance_price, covariance_floor, Win.Promise.receive.

Proposition III.5.14 (precision matrix of the small-field Gaussian of a step).

Setting. Fix a step-estimator \(\mathcal{T}\) and a patch \(\mathcal{D}\) (block side \(L\), order \(J\)) satisfying the branch conditions for its kernel and gauge group. For a step with patch \(\mathcal{D}\), tube \(T\), coupling \(x\), tree-gauge forest \(F\), packet \(A\) and coarse field \(V\), define \(D\) as follows:

This is a transported flat linearization. Its equality to the derivative of the nonlinear cut rows at \(V\) is not asserted.

Let \(\text{gram} = D^T D + \text{diag} (\boldsymbol{1}_{\text{inactive}})\), with positions the coarse parents of the bonds.

Claim. There are constants \(r, K, U, \alpha > 0\), with \(\alpha\) half the decay rate of the upper path data, such that the following holds for all large \(x\). For every step with patch \(\mathcal{D}\) and \(\text{Funds}\), coarse side \(\geqslant 100 (1 + B)\), every \(F\), and every \(A, V\) for which the flags of \(A\) are the flagged sites of \(V\):

Consequently, for large \(x\), III.5.13(b) applies to the following windows: a fine bond belongs to the window of an addressed bond exactly when at least one of its projected endpoints is within distance \(\lfloor B / 4 \rfloor\) of at least one projected endpoint of the addressed bond. Include every color coordinate of each such bond. Their room is \(\lfloor B / 4 \rfloor\), and the error is \(K_{\text{win}} e^{- c_{\text{win}} \lfloor B / 4 \rfloor} < 1\).

Status: theorem, conditional on \(\text{Funds}\) and \(\text{Branches}\). Used in III.4 and III.11 through the window estimates below. Lean: Aff.bedrock, Aff.surplus, Aff.gram_budget, Aff.install_promises; definitions Aff.Dmat, Aff.gram, Promise.Served.

Further outputs.

2. What it uses

From other units.

Standard mathematics. The formal proofs establish the estimates directly. References for the methods are:

Imported constants and assumptions.

3. Strategy

For a finite gas, inclusion–exclusion writes the partition function as a sum over tuples. Splitting off the connected component of a distinguished occurrence gives the coefficient recurrence \((m + 1) g_{m + 1} = \sum_{k = 0}^m (k + 1) \ell_k g_{m - k}\). For the gas polynomial \(P (t) = \sum_n g_n t^n\) and cluster series \(L (t) = \sum_m \ell_m t^{m + 1}\), this becomes \(P' = L' P\). A geometric bound on the cluster coefficients justifies differentiation on an interval containing \([0, 1]\); then \(Pe^{- L} = 1\). Absolute bounds allow passage to arbitrary label sets through the directed family of finite subsets.

The bounds combine the tree-graph inequality with a sum over trees. A parent support \(Y\) meets labels of total inflated weight at most \(|Y| \gamma \leqslant \text{ball}_B \gamma \hspace{0.17em} \text{cov}_B (Y)\). The reserve exponent \(\delta\) absorbs each parent’s number of children through \(x^k \leqslant k!e^x\). Substochastic kernels sum the labels; the sum of the resulting child factorials over parent maps is at most \(4^m m!\). After cancellation of the occurrence factorial, the ratio per added vertex is \(4 \hspace{0.17em} \text{ball}_B \gamma / \delta\). Subadditivity of the covering number transfers these bounds to cluster unions. The resulting estimates retain site pins, tails, deleted regions and repeated forced polymers.

For seed packets, a spanning-walk count gives an entropy factor \(K_{\text{step}}^2\) per seed. The point charge supplies an energy gain \(\beta \kappa t^2\) per seed. Since \(t = \beta_F^{- 1 / 2} P (\beta_F)\) and \(\beta \geqslant u \beta_F\), the guaranteed gain is \(\kappa uP (\beta_F)^2\). At fixed \(L\), this dominates every cost \(\text{bill}^d P (\beta_F)^z\) with \(z < 2\). Half the gain can be retained as a mark for later estimates. The partition of unity over global seed records then splits the block-spin pushforward while preserving a common field, tag array and tree-gauge slice.

The deletion construction obtains another logarithm from the region recurrence. Its negative-logarithm pieces form a union-convolution series pinned at a chosen site. A weighted \(\ell^1\) bound closes induction on the region and compares signed activities with the gas of negative envelope activities. The final matrix estimates use a transported flat linearization: a gauge-fixed Poincaré inequality and a small-background comparison give a uniform spectral gap, while path bounds give weighted row and column estimates. Exponential conjugation yields inverse decay, and principal window inverses approximate the full inverse with error \(O (e^{- cB})\).

4. Steps

4.1 Algebra of the finite gas

In Lemmas 4.1–4.4 the label set \(\iota\) is finite, the weights are real, and \(R\) is reflexive and symmetric. All vertex sets are finite.

Lemma 4.1 (gas as a sum over tuples). Let \(R\) be reflexive and symmetric, and \(g_n = \frac{1}{n!} \sum_{z \in \iota^n} \text{all} (z) \prod_a w_{z_a}\), where \(\text{all} (z) = \sum_{G \subseteq G_R (z)} (- 1)^{|G|}\). Then \(g_n = \sum_{|s| = n, s \text{compatible}} \prod_{i \in s} w_i\), and \(\sum_n g_n = \Xi (w)\). Sketch:

Lean: Partition.allMass_eq, Gas.row, Gas.sums.

Lemma 4.2 (splitting off a component). For any relation \(R\), any tuple \(z\) on a finite vertex set \(\alpha\), and \(o \in \alpha\):

\(\displaystyle \text{all} (z) = \sum_{s \ni o} \phi^T_R (z|_s) \hspace{0.17em} \text{all} (z|_{\alpha \setminus s}) .\)

Sketch: a subgraph is the same thing as three pieces of data:

The signs multiply. Lean: Partition.partition. (proof read; the bijection lemmas statement only)

Lemma 4.3 (coefficient recurrence). \((m + 1) \hspace{0.17em} g_{m + 1} = \sum_{k = 0}^m (k + 1) \hspace{0.17em} \ell_k \hspace{0.17em} g_{m - k}\). Sketch:

Lean: Coef.unbased, Coef.discrete, Coef.recurrence.

Lemma 4.4 (III.5.1). Under the geometric coefficient bound of III.5.1, \(\exp (\sum_m \ell_m) = \Xi (w)\). Sketch:

Lean: Gas.exact_polynomial.

4.2 Tree-graph bound

Lemma 4.5 (Penrose partition; III.5.2). The absolute connected-graph sum is bounded by the number of spanning trees. Sketch:

The bound follows without identifying which trees have nonzero fibres. Lean: Graph.pay; breadth-first construction in Breadth.lean. (proof read; breadth-first construction statement only)

4.3 Sums over trees

Lemma 4.6 (substochastic sums along a forest). Let \(\alpha, \beta\) be finite, \(p : \alpha \to \alpha \sqcup \{o\}\) an acyclic parent map, \(Y\) a fixed root state, and \(X_j\) the state associated with \(j \in \beta\). If \(K (Y', j) \geqslant 0\) and \(\sum_j K (Y', j) \leqslant 1\) for every state \(Y'\), then \(\sum_{z : \alpha \to \beta} \prod_i K (X_{z_{p (i)}}, z_i) \leqslant 1\), with the first argument interpreted as \(Y\) when \(p (i) = o\). Sketch: induction on the set of summed vertices. Sum out a vertex of maximal rank first: it is a leaf, and its row sum is \(\leqslant 1\). Lean: Trees.substochastic.

Lemma 4.7 (fan-out factorials). \(\sum_{p : [m] \to [m] \sqcup \{o\}} \prod_v (\text{card} \{i : p (i) = v\}) ! = \binom{2 m}{m} m! \leqslant 4^m m!\). Sketch: with \(m + 1\) possible parents fixed, extending a map on \(n\) vertices multiplies the total factorial weight by \(n + m + 1\). Iterating from \(n = 0\) to \(m\) gives \((2 m) ! / m!\), and \(\binom{2 m}{m} \leqslant 4^m\). Lean: Trees.parent_fanout, fanout_bound.

Lemma 4.8 (trees rooted at a support). Assume \(w \geqslant 0\), \(\gamma, \delta > 0\), \(a \in \mathbb{R}\), and \(\text{Hit}_B (w ; a + \delta, \gamma)\). For any support \(Y\) and any finite \(D \subseteq \iota\),

\(\displaystyle \sum_p \sum_{z \in D^m} \boldsymbol{1} [X_{z_i} \cap X_{\text{parent} (i)} \neq \emptyset \forall i] \frac{1}{m!} \prod_i \left( w_{z_i} e^{a \hspace{0.17em} \text{cov}_B (X_{z_i})} \right) \leqslant e^{\delta \hspace{0.17em} \text{cov}_B (Y)} \hspace{0.17em} \left( 4 \hspace{0.17em} \text{ball}_B \gamma / \delta \right)^m,\)

where \(p\) runs over acyclic parent maps on the \(m\) labelled vertices with external root of support \(Y\). Sketch:

Lean: Trees.trees_afford, Trees.forest_at.

4.4 Cluster sums and the logarithm

Lemma 4.9 (rooted and marked cluster sums; III.5.3(c)). Under III.5.3’s hypotheses, the marked sum at order \(m\) is at most \(bq^m\). Sketch:

Lean: Ursell.cluster_at.

Lemma 4.10 (pinned bounds; III.5.3(a),(b)). Under III.5.3’s hypotheses, both pinned tails are at most \(\gamma q^m / (1 - q)\). Sketch:

Lean: Ursell.connected_pay, Ursell.log_hit_tail, Trees.touch_tail. (proof read; distinct-label forest count statement only)

Lemma 4.11 (from finite to arbitrary label sets; III.5.4). Under III.5.4’s hypotheses, the gas and cluster series converge absolutely, and \(\Xi = \exp (\text{cluster series}) > 0\). Sketch:

Lean: Gas.finite_log, Gas.price_log, Product.everyone_ne, Product.set_products.

4.5 Deleted regions and forced polymers

Lemma 4.12 (III.5.5). Under III.5.5’s hypotheses, deleting polymers meeting \(Y\) multiplies the gas by \(e^{- p}\), with \(|p| \leqslant |Y| \gamma / (1 - q)\) and the stated cluster tails. Sketch:

Lean: Gas.restores_log, Gas.missing_tail.

Lemma 4.13 (multiplicities on a fixed set). For a finite set \(s\), if \(0 \leqslant \sigma < 1\) and \(0 \leqslant w_i \leqslant \sigma\) for \(i \in s\), then \(\sum_{(n_i) \in \mathbb{N}_{\geqslant 1}^s} \prod_{i \in s} w_i^{n_i} / n_i ! \leqslant \prod_{i \in s} (w_i / (1 - \sigma))\). Sketch: bound \(w_i^{n_i} / n_i !\) by \(w_i \sigma^{n_i - 1}\) and multiply the geometric sums. Lean: Multiple.buy_at.

Lemma 4.14 (multiset sums). Let \(X_i\) be nonempty, \(w_i \geqslant 0\), \(a \geqslant 0\), \(t \geqslant 1\), \(0 \leqslant \sigma < 1\), and \(e_i \in \mathbb{R}\). Suppose \(\text{Ready} \left( t \hspace{0.17em} w \hspace{0.17em} e^e / (1 - \sigma) ; a, \gamma, \delta, q \right)\) and \(\gamma \leqslant \sigma\). Then the sum over multisets with connected support through \(x\), satisfying a predicate \(Q\) on which \(\sum n_i e_i \geqslant v\), with total multiplicity \(\geqslant m\), of \(e^{a \hspace{0.17em} \text{cov}_B (\cup)} \prod w_i^{n_i} / n_i !\), is at most \(t^{- m} e^{- v} \gamma / (1 - q)\). Sketch:

Lean: Multiple.estimates.

Lemma 4.15 (III.5.6). Under III.5.6’s hypotheses, the pinned multiset sum containing a forced label is at most \(t^{- m} e^{- \mu} \eta / (1 - q')\). Sketch:

Lean: OmitClusters.restore, Labels.together.

4.6 Packets and the large-coupling numerics

Lemma 4.16 (animals and covers). In the setting of III.5.7:

Sketch:

Lean: Packets.stores_count, Animals.animals, Packets.cover_support. (statement only for the animal count; covering proof read)

Lemma 4.17 (III.5.7(a)). Under III.5.7(a)’s hypotheses, the packet tail is at most \(K_{\text{hit}} Cq^{\max (1, m)} / (1 - q)\). Sketch:

Lean: Packets.packet_income, Packets.cover_income.

Lemma 4.18 (III.5.7(b)). Under III.5.7(b)’s hypotheses, the mixed-word sum has the bound stated there. Sketch:

Lean: Mixed.ready_vertices, Mixed.mixed_hit, Reserves.sets, Reserves.reserve_tail.

Lemma 4.19 (III.5.8). For fixed numerical parameters satisfying III.5.8, all its packet and tree conditions hold for sufficiently large \(x\), uniformly in volume. Sketch:

Lean: Delay.monitor_pays, Delay.step_delay, Delay.fits, Buy.quoted, Grant.order. (proof read; the grant of III.10 statement only)

4.7 The logarithm by deletion

Lemma 4.20 (deletion identity). For the gas \(Z_u\) of III.5.9(a) and \(x \in B\):

\(\displaystyle Z_u (B) = Z_u (B \setminus \{x\}) + \sum_{P \subseteq B, P \ni x} \varphi_u (P) \hspace{0.17em} Z_u (B \setminus P), \qquad \varphi_u (P) = \sum_{i : \hspace{0.17em} X_i = P} u_i,\)

and \(Z_u (\emptyset) = 1\). Sketch:

Lean: ReLog.stock_deletion, ReLog.shelf.

Lemma 4.21 (the incidence algebra). Let \(J\) be finite and let \(w\) satisfy the weight assumptions of III.5.9(a). For set functions on \(J\) define:

Then:

Sketch: cumulative sums turn union convolution into multiplication. Two nonzero factors pinned at \(x\) have intersecting supports, so the weight inequality gives submultiplicativity. Apply it to convolution powers and sum \(\sum_{n \geqslant 1} q^n / n = - \log (1 - q)\) for \(q < 1\); termwise absolute domination gives the last assertion.

Lean: ReLog.crop_glue, ReLog.lamp_shed, ReLog.lamp_cropped, ReLog.lamp_mono.

Lemma 4.22 (the induction). Let \(J\) be finite, \(w\) satisfy III.5.9(a)’s weight assumptions, and \(a \geqslant 0\). Let \(U, z\) be functions of supports with \(U \geqslant 0\), \(|z| \leqslant U\), and \(\sum_{P \ni x} w (P) U (P) e^{a |P|} \leqslant 1 - e^{- a}\) for every \(x\). Suppose \(Z, Y\) satisfy the deletion identities with activities \(- U, - z\), respectively, and \(Z (\emptyset) = Y (\emptyset) = 1\). Then for every \(B\):

Sketch: strong induction on \(B\). The empty region is immediate. For nonempty \(B\), pick \(x \in B\).

\(\displaystyle \hat{f}_x (A) = \sum_{P \subseteq A, P \ni x} U (P) \exp \hspace{-0.17em} \left( \sum_{T \subseteq A \setminus \{x\}, T \cap P \neq \emptyset} \lambda_Z (T) \right), \qquad A \subseteq B.\)

The conditions \(x \in P \subseteq A\) make \(f_x\) supported on sets containing \(x\). The deletion identity gives \(Z (A) = Z (A \setminus \{x\}) (1 - \hat{f}_x (A))\).

Lean: ReLog.leash, ReLog.matures, ReLog.swept, ReLog.catalog.

Lemma 4.23 (deleting sites). Under III.5.9(a)’s hypotheses, with \(Z = Z_u\) and \(\lambda = \lambda_u\), \(- \log Z (J \setminus Q) = \sum_{A \cap Q = \emptyset} \lambda (A)\), hence \(l = \sum_{A \cap Q \neq \emptyset} \lambda (A)\), and \(|l| \leqslant \sum_{x \in Q} \sum_{A \ni x} w (A) \lambda_{- v} (A) \leqslant a |Q|\) because \(w \geqslant 1\). Lean: ReLog.omitted.

Lemma 4.24 (at a coarse field; III.5.9(b)). Under III.5.9(b)’s hypotheses, its masked logarithm coefficients satisfy the common envelope and deletion-ratio bounds. Sketch:

Lean: ReLog.intake, ReLog.native_rate, ReLog.permits, ReLog.retain_mask, ReLog.retest_meas.

Lemma 4.25 (trace-log; III.5.10). Under III.5.10’s hypotheses, the trace-log series converges absolutely to \(\log \det (I + S)\), and both determinants are positive. Sketch:

Lean: Round.twins, Round.twin_log.

4.8 The sector decomposition

Lemma 4.26 (allocations and columns). An allocation is the partition of unity with global seed tests specified in III.5.11. It satisfies \(0 \leqslant D_A \leqslant 1\); where the likelihood is positive, \(D_A \ne 0\) implies the local sector prerequisites.

Sketch: sum the allocation before integrating, then apply the tree-gauge integration identity for \(g\) and the gauge-covariant block kernel.

Lean: Allocation, Allocation.hard, Chi.canonical, push_columns.

Lemma 4.27 (cut-out identity). For a representation adapted to the tube as in III.5.11, let \(\Delta_{\text{cut}} (U, V)\) be the orientation-averaged fine cut-row action minus its coarse counterpart. Then

\(\displaystyle \text{kernel} \times \rho_R = \text{sieve} (V) \cdot \text{exposed} (U) \cdot \text{kernel} \cdot e^{- \beta \hspace{0.17em} \Delta_{\text{cut}} (U, V)} .\)

Lean: Presented.expel, Presented.use_gap.

Lemma 4.28 (energy factorization). For a step satisfying its branch conditions,

\(\displaystyle \text{kernel} \cdot e^{- \beta \hspace{0.17em} \Delta_{\text{cut}}} = \pi_{\beta} \hspace{0.17em} e^{- \beta \hspace{0.17em} \bar{\mathcal{E}}_A} \cdot (\text{likelihood}) \hspace{0.17em} e^{- \beta \hspace{0.17em} \bar{\mathcal{E}}^{\text{rest}}_A},\)

where bars denote averages over the lattice symmetries, applied to the rotated packet. On the tree-gauge slice, the global seed tests and positive likelihood imply the local prerequisites in every orientation. The point charge therefore gives \(\bar{\mathcal{E}}_A \geqslant \kappa t^2 |A|\) wherever \(D_A \ne 0\) and the likelihood is positive; zero-likelihood terms vanish. Lean: Oriented.enters_eq, Oriented.mean_pays.

Lemma 4.29 (III.5.11). Under III.5.11’s funding, adaptation and point-charge assumptions, the sector outputs sum to the pushforward and satisfy its per-seed bound. Sketch:

Lean: Oriented.turn_charge, Presented.portionCost.

Lemma 4.30 (one \(\kappa\)). The uniform point-charge conclusions of III.5.11 hold. Sketch: the general conclusion uses the step-estimator’s energy grant (III.10). At a committed pair of scales, use the run’s energy hypothesis (III.1); its compatibility conditions identify the representation’s paths, clip and threshold with those of the tube. The constant is chosen before the volume, step and representation.

Lean: points_grant, Consumer.bought, charged_at. (proof read; the inputs from III.1 and III.10 statement only)

4.9 Matrices

Lemma 4.31 (III.5.12). Under III.5.12’s hypotheses, the shifted inverse has the stated pointwise exponential bound. Sketch: fix \(j\) and put \(s_k = e^{\theta \eta d (a_k, a_j)}\) and \(w = s \cdot (M + uI)^{- 1} e_j\).

Lean: Green.inverse, Green.diminishing, Green.schur.

Lemma 4.32 (III.5.13). Under III.5.13’s hypotheses, the window residual, inverse difference and covariance satisfy its weighted bounds. Sketch:

Lean: Windows.stitched, Windows.inverse_price, Windows.window_price, Win.planePrice.

Lemma 4.33 (spectral gap). Fix \(\mathcal{T}, \mathcal{D}\) as in III.5.14. There is \(0 < r \leqslant 1\) such that, for all sufficiently large \(x\), all funded steps with this patch, coarse side \(\geqslant 100 (1 + B)\) and the stated flag equality satisfy \(r |v|^2 \leqslant |Dv|^2\) for every \(v\) supported on active coordinates. Sketch:

Lean: Aff.inflow, Aff.surplus, Aff.decomposes. (proof read; per-bond comparison and III.6 inequality statement only)

Lemma 4.34 (decay). Fix \(\mathcal{T}, \mathcal{D}\) as in III.5.14. There is \(K > 0\), uniform over all funded steps with that patch, all \(V, A, F\) and every real \(x\), such that \(D\) and its unmasked version have weighted row and column bounds \(K\) at rate \(\alpha\) equal to half the upper path-decay rate, and \(\text{gram}\) has bounds \(K^2 + 1\). Moreover \(\text{gram}_{ij} = 0\) at distances greater than \(2 (\lfloor B / 100000 \rfloor + R_{\text{pred}} (\mathcal{D}) + 2)\). Sketch:

Lean: Aff.gram_budget, Aff.stencil, Aff.received.

Lemma 4.35 (III.5.14). The uniform matrix and window conclusions of III.5.14 hold. Sketch: combine Lemmas 4.33 and 4.34 with the index-counting bound, then apply III.5.13(b). The windows have room \(\lfloor B / 4 \rfloor \to \infty\) as \(x \to \infty\).

Lean: Aff.bedrock, Aff.install_promises.

Lemma 4.36 (auxiliary Gaussian estimates). Fix the step-estimator and patch of III.5.14. The auxiliary Gaussian (Mint) uses the completed precision \(G = D^T D + I_{\text{inactive}}\), the affine constant \(h\) (including the clipped frozen coordinates), source \(D^T h\), and selected mean \(\mu = - \text{within} (G)^T D^T h\).

For the field estimates below, assume a funded step with this patch, coarse side \(\geqslant 100 (1 + B)\), and the flag equality of III.5.14. Write \(H_{n, p} = B^p t^n\). A reserve \(C_{\text{res}} > 0\) supplies uniform bounds on the upper and lower row sums and the coarse predicted fields. A compatible pair of quadratic forms means upper and lower forms attached to the tube, satisfying its diagonal and renormalization identities, the two quadratic bounds and orientation covariance. The shift and plateau estimates assume such a pair and the window bounds of III.5.13(b). Each large-\(x\) threshold is uniform in the remaining volume and field data.

5. Where the difficulty lies

The tree-graph argument is standard. The estimates must also retain arbitrary label sets, repeated occurrences, site pins and covering weights at the growing scale \(B = \lceil (\log x)^2 \rceil\). The entropy factor \(4 (2 B + 1)^4\) is sufficient because it grows only polynomially in \(\log x\).

The gain scale is \(P (\beta_F)^2 = e^{2 (\log \log \beta_F)^2}\). It absorbs \(\text{bill}^d P (\beta_F)^z\) for fixed \(d\) and \(z < 2\), but grows more slowly than every positive power of \(\beta_F\). The deletion route gives support-indexed logarithm coefficients and a common field-independent envelope; the graph route also gives occurrence tails and multiset bounds. Both give deletion ratios uniform in volume, although the full logarithm is extensive.

The sector estimate needs global seed records to preserve the local point-charge prerequisites under lattice orientations. Its remainder \(\ell_A\) has no volume-uniform bound here. For the Gaussian estimates, the substantive inputs are the gauge-fixed Poincaré inequality and control of the transported flat rows away from the flagged region. Finite range grows with \(B\); uniform exponential row bounds provide the useful control.

6. Notation

Written proof Lean Meaning
\(\text{SB} (f ; C)\) SumBudget \(f \geqslant 0\) summable, \(\sum f \leqslant C\)
\(\text{Hit}_B (w ; a, \gamma)\) Ac.Hit site-sum condition
\(\text{cov}_B\), \(\text{ball}_B\) coverNumber, ballCost connected covering number; \((2 B + 1)^4\)
\(\text{Ready} (w ; a, \gamma, \delta, q)\) Trees.Ready, degreePrice smallness of the cluster expansion
\(R \subseteq\) overlap Ursell.Tests, blocked incompatibility implies overlap; the overlap relation
\(G_R (z)\), \(\phi^T_R (z)\) Ursell.pot, Ursell.coefficient incompatibility graph of a tuple; Ursell coefficient
\(\Xi (w)\); cluster term; log series Gas.total, Gas.term, Gas.logSum gas, its cluster terms and their sum
\(g_n\), \(\ell_m\) Coef.gasRow, Coef.logRow coefficients of \(\Xi\) and of the log series
\(w^Y\), \(p\) Gas.cut, Gas.missing weights off \(Y\); clusters meeting \(Y\)
multisets, \(\prod w^n / n!\) Multiple.Reps, manyW labels with multiplicities
packet \(A\), number of seeds, \(D_r (A)\) Stores, shares, dome finite set of typed seeds; radius-\(r\) neighbourhood of their locations
connected packet, \(K_{\text{hit}}\), \(K_{\text{step}}\), \(\tau_L\) Packets.connected, hitCost, stepCost, typesAt animal-counting constants
\(\rho\), packet tail ratioCost, fit, totalCharge per-seed ratio; tail of packet sums
hypotheses of III.5.7(b) Mixed.Costs.Buy, Mixed.Accounts numerical list; weight bounds
\(B\), \(\beta_F\), \(t\), \(P\) radiusB, betaF, fiery, Vac.Numbers.P \(\lceil (\log x)^2 \rceil\); \(x + b \log L\); \(\beta_F^{- 1 / 2} P (\beta_F)\); \(e^{(\log \log s)^2}\)
\(M\), gain \(\beta \mu t^2\) hatWin, Grant.credit guaranteed and actual gain per seed
\(Z (B)\), \(\lambda (A)\) fumes, locker, pieces disjoint-family gas; Möbius pieces of \(- \log Z\)
\(f \ast g\), \(\hat{f}\), \(\|f\|_w\), \(\lambda (f)\) glue, shed, crop, lamp union convolution, cumulative sum, weighted norm, log series
hypotheses of III.5.9 Rate, Deletion, Receipt, Sufficient, Fund, Tags weight, deletion identity, site conditions, coefficient properties
allocation \(D\); point charge Allocation, PointCharge partition of unity over packets; \(\mathcal{E}_A \geqslant \kappa t^2\) per seed
\(\mathcal{E}_A\); \(c_A\); \(\ell_A\); \(\pi_{\beta} e^{- \beta \kappa t^2 k}\) assessed, cutEntry, leftOver, earn sector energy; sector output; remainder; gain
weighted Schur classes; \(M \geqslant r\) Green.Rows, Green.Both, Green.Bottom row/column sums at rate \(\eta\); lower bound
windows, \(\text{within}\), \(E\), \(\text{cov}\), \(\Pi_U\) Win.Windows, within, error, cov, planePrice localized inverse and covariance
uniform matrix data; window constants Win.Promise, Turns, Turns.Takes (mass, bottom, rate, counting); bounds of III.5.13(b)
\(D\), \(\text{gram}\) Aff.Dmat, Aff.gram transported flat row matrix; completed precision matrix
transported flat derivative Aff.derivative row operator used to define \(D\)
auxiliary Gaussian assumptions Aff.Reserve, Aff.Alongside, Aff.cert reserve bounds; compatible quadratic forms; volume and flag tests
\(H_{n, p}\); comparison field \(v\) Rates.allowance, Aff.unit \(B^p t^n\); active stationary extension minus replica extension, transported back by the local gauge

7. What remains unexplained

  1. Imported analytic inputs. The step-estimator’s energy grant in III.10 and the Poincaré inequality in III.6 are used as stated. Their proofs, the full per-bond spectral comparison and the origin of the run’s committed-scale energy hypothesis have not been audited here.

  2. Later applications. The dependency database identifies consumers in the units listed above. Their verification of the site-sum, per-packet and multiset hypotheses remains outside this unit’s audit. In particular, the bounds here do not construct renormalization outputs or bound the sector remainders uniformly in volume.

  3. Unread supporting proofs. The breadth-first construction, the distinct-label forest count, the spanning-walk construction for animals and the integration of allocations restricted to sectors were used through their statements. In the auxiliary Gaussian chain, the statements and the displayed estimates of Lemma 4.36 were checked; its supporting local chart and quadratic-form comparison lemmas were not audited throughout. These remain proof inputs, rather than additional conclusions of the present sketches.

  4. Flag locality. In III.5.9(b), locality on \(P\) applies to the raw coefficients. Their masked versions also inspect flags on \(P\); those flags can depend on plaquettes extending outside \(P\). A written proof must retain this flag dependence or enlarge the physical support before claiming ordinary link locality for the masked coefficients.

  5. Verification scope. This is a source correspondence check. No Lean build, axiom computation or rendered-page check was run.

← III.4 Gaussian integration: change of Gaussian sector, determinants, window law (notes)Part III·Unit outlineIII.6 The stationary branch: background field of the effective action (notes) →