Home·The theorem·The proof in one reading·Architecture·Units·Manuscript·Glossary·FAQ·Work items·Data·Method·Provenance·Changes
← III.9 The marginal flow: the running coupling and its reset (notes)Part III·Unit outlineIII.11 Interacting rank, Gaussian comparison and the stress tensor (notes) →

III.10. One full nonlinear renormalization step: bounds on irrelevant terms — detailed notes

Contents

1. What this unit proves 2

2. What it uses 10

3. Strategy 11

4. Steps 12

4.1 The algebraic identity 12

4.2 Packing and geometry 12

4.3 Kernel bounds 13

4.4 Row-by-row lower bounds 14

4.5 Linearization defects on unaffected cells 15

4.6 Assembly of III.10.1 17

4.7 Activities of one step 18

4.8 Derivatives 20

4.9 Start-up fusion and markers 21

4.10 The schedule 21

5. Where the difficulty lies 22

6. Notation 23

7. What remains unexplained 23

The full account of unit III.10: 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: YM/Work/ (items E01–E09, H01–H04, I01–I04), YM/Bid/, YM/Next/FLUCTUATIONS/, YM/AnalyticNext/CONNECTED_ACTIVITY_JETS_E6/, YM/PhysicalNext/RG_NEXT_COMPLETE_REPEATABLE_NONLINEAR_STEP/, YM/PhysicalNext/V9_JOINT_ENDPOINT_LOAD/, YM/Pressure/ (definitions). Size: 49,448 lines in the main-theorem closure.

Conventions. Levels are indexed from coarse to fine, as in III.9.

\(\displaystyle t_c (x) = x^{- 1 / 2} e^{(\log \log x)^2} \text{(coarse field)}, \qquad t = t_c (\beta_F) \text{(fine field)}, \qquad T = \beta_F^{- 49 / 100} \text{(large field proper)},\)

with cluster radius \(B = \lceil (\log x)^2 \rceil\) and path-length cutoff \(R = \lfloor B / 10^5 \rfloor\).

1. What this unit proves

The unit does not construct the effective actions; that is III.9, with III.5–III.8. It proves the estimates on one nonlinear block-spin step that the construction consumes:

III.10.6 collects the start-up fusion and marker estimates. In the energy argument, constants are chosen before \(L\), after \(L\), or after \(W\); this order determines which losses they can absorb. The derivative and Gaussian estimates have their own quantifier orders, stated below.

Data of a step.

\(\displaystyle \Phi_0^{\ast} \Phi_0 {+ \Phi_0'}^{\ast} \Phi_0' + N_0^{\ast} N_0 = 1.\)

The transported kernels cut at length \(R\) need not satisfy this identity; their defects are estimated in §4.5.

Seeds. A store \(A = (A_F, A_I, A_{\flat})\) (Stores) consists of finite sets of fine plaquettes ("costs"), coarse bonds ("aisles") and coarse sites ("flags"). \(\sharp A\) is the sum of their cardinalities, including coincident seeds of different types. \(D_r (A)\) is the coarse \(r\)-neighbourhood of their positions: all parent sites touched by a fine plaquette, both endpoints of a coarse bond, and each flag site. \(A\) is consistent with \((U, k, V)\) at \((x, W)\) (Plot) if the following hold:

Nothing is assumed outside \(D_{5 B} (A)\). The spent seeds \(A^{\text{sp}}\) are the costs with \(|U_p | > T\) and the aisles with \(w_b > T\).

Action gap.

\(\displaystyle \mathcal{A}_F (U) + \frac{a}{2} \sum_b w_b^2 -\mathcal{A}_c (V) .\)

So \(\mathcal{E}_A\) is the action gap of the step (fine action, plus block-spin penalty, minus coarse action), localized near the seeds.

Proposition III.10.1 (energy per seed). Let \(P\) be a request. There is \(L_0\) such that for every block factor \(L \geqslant L_0\) there is \(W_0 (L)\) such that for every \(W \geqslant W_0\) there is \(x_0\) with the following property. Take \(x \geqslant x_0\) and:

Then

\(\displaystyle \kappa \hspace{0.17em} t^2 \hspace{0.17em} \sharp A \hspace{0.27em} \leqslant \hspace{0.27em} \mathcal{E}_A (U, V), \qquad \kappa = \tfrac{1}{2} \min \{\lambda_F / 8, a / 32, \alpha_c / 64\},\)

where \(\lambda_F\) is the linear clip constant of the fine stage and \(\alpha_c\) the quadratic clip constant of the coarse stage.

Status: theorem under the stated admissibility, tree-gauge and consistency hypotheses.

Lean: proved by Fluctuations.grants. This is Integrated.pressure_grants applied to Fluctuations.bare, through Bid.issues.

Proposition III.10.2 (activities of one step). Let \(P\) be a request and \(\mathcal{T}\) target data (Target): constants \(\varepsilon, \delta > 0\), \(e \geqslant 0\), a rate \(r \in (- \tfrac{1}{2}, 0)\), a halo \(h \geqslant 500\) and a local exponential chart of \(G\), with \(4 \varepsilon\) below the support exponent of the fine stage and \(8 \varepsilon\) below both stage widths. Put \(a_{\ast} = e + 2 \varepsilon + \delta\).

There is \(L_0\) such that for every \(L \geqslant L_0\) and every large \(W\) there is \(\mu > 0\) with the following property. For every patch \(D\) of block factor \(L\) satisfying the branch condition there are window constants \(o\) (Target.Order: a decay rate, a spectral gap, a shell width \(K\), a Lipschitz constant \(\gamma\), a drift cap, a plateau \(a_o\) and a trigger level \(\xi\)) and a constant \(C \geqslant 0\) such that the following holds for all large \(x\), every step \(S\) on \(D\) with \(c \geqslant 100 (1 + B)\) and every tube admissible for \(P\) (Works).

\(\displaystyle \sum_{i : o \in X_i} u_i \exp (a_{\ast} N_B (X_i) + 2 \varepsilon d_i) \leqslant \beta_F^r \qquad \text{for every coarse site } o.\)
 Here $N_B(H)$ is the minimum number of radius-$B$ balls whose connected union contains $H$. The letters are the forced activities of $K$, the logarithmic letters, and the warning letters. Their weights are respectively $k_i$, $y_i$ and $E$ times a fixed patch-dependent warning-amplitude bound; warning letters have $d_i=0$;
  1. in every manifest sector, the following bounds hold:

Status: theorem. The target data, fine input bounds and free-kernel compatibility are hypotheses. The input bounds are supplied along the RG trajectory (III.9); the compatibility construction is discussed in §7.

Lean: Target.issue. Target.quotesTogether repackages the output as the bundle Cash (the activity estimates together with witnesses for the letter weights).

Proposition III.10.3 (derivatives of the new activities). Let \(P\) and \(\mathcal{T}\) be as in III.10.2. There is \(L_0\) with the following property. Fix \(L \geqslant L_0\), a patch \(D\) satisfying the branch condition, an order \(i \in \mathbb{N}\), a sector cutoff \(\chi\), \(W, \mu > 0\), window constants \(o\), and chart constants \((A, f)\) satisfying YM.ConnectedActivityE6.Blaze.Eider.Canary. The latter require \(A, f > 0\), \(\log (\exp u) = u\) for \(|u| \leqslant 2 A\), \(\exp (\log (\exp u \hspace{0.17em} g)) = \exp u \hspace{0.17em} g\) for \(|u| \leqslant A\) and \(\|g - 1\| \leqslant f\), and analyticity of \(M \mapsto \log (\exp u \hspace{0.17em} M)\) for \(|u| \leqslant 2 A\), \(\|M - 1\| \leqslant f\). Fix a function \(\mathcal{V}\) of rate \(- 1 / 2\) that is eventually nonnegative (Lemma 4.15).

There are \(C \geqslant 0\), functions \(q, \rho, p\) of rates \(- 1 / 2, 0, 0\), respectively, and \(l \in \mathbb{N}\) such that, for all large \(x\), the following holds for every step on \(D\) satisfying the compatibility condition, every fine representation \(R\), and every bound bundle \(H\) of III.10.2 at \((W, x, \mu)\). One has \(8 a_o t \leqslant A / 2\) and a derivative-bound structure \(\Xi\) (Utopia) with the following properties in every sector at \(x\), of width \(o.K\), cutoff \(\chi\), and with the actual coarse flags.

\(\displaystyle \|D^m V_e (y_0)\| \leqslant c_G \mathcal{V} (x)^m \qquad (0 \leqslant m \leqslant i),\)

where \(c_G\) is the fixed matrix-norm bound in Mode. Transport the fluctuation coordinates compatibly with this family: the physical and Gaussian coordinates represent the same field, their physical displacement has all derivatives through order \(i\) bounded by \(q (x)\), and, near \(y_0\), the resulting fine field agrees with the held fine field on every cold bond and on every hot bond whose physical coordinate has norm at most \(8 a_o t\) (YM.ConnectedActivityE6.Blaze.Silica.Clutch). In a region disjoint from the coarse flags, each letter \(a_j\) is \(C^i\) at \(y_0\) and satisfies

\(\displaystyle \|D^m a_j (y_0)\| \leqslant \text{cost}_j (\rho (x) \text{span}_j)^m \qquad (0 \leqslant m \leqslant i)\)

(Brush). Packet spans are one; other spans include their support and path lengths. The costs use the packet estimate and the nonpacket weights before extracting the catastrophe-probability factor.

The same \(\Xi\) satisfies an integrated estimate (Herd). Let \(\mathfrak{p}\) be a finite molecule with connected support-overlap graph and total support disjoint from the flags. Require that its packet stores are connected with pairwise disjoint radius-\(hB\) neighbourhoods, that its old forced supports are pairwise disjoint, and that every site in an old forced inventory touches a fine-plaquette seed in the union of the packet stores. Along a family \(V (y)\) as above, the prescribed transported integrand is \(C^i\) at \(y_0\) and has every derivative of order at most \(i\) bounded by an integrable function \(g\). With \(d \nu = d \lambda_F \hspace{0.17em} d \gamma\), the product of tree-slice Haar measure and Gaussian fluctuation measure,

\(\displaystyle \int g \hspace{0.17em} d \nu \leqslant |p (x) | \prod_{j \in \mathfrak{p}} \text{fare}_j \hspace{0.17em} (1 + \text{shape} (\mathfrak{p}))^l .\)

The fares include a factor \(\text{rad}\) for each distinct warning address. For one measurable, gauge-invariant source letter of amplitude at most \(w_M\), local with its eligibility tests included, the same estimate replaces its fare by \(w_M\). A forced source must vanish outside the original eligibility set; a regular source occupies an unused regular-letter address. The remaining warning and shape factors are unchanged.

Status: theorem, conditional on the input bound bundle and compatibility. The constants and functions are uniform in the volume. Captain and rocketsC also control iterated left derivatives in spatially weighted directions.

Lean: Utopia.whittle (letter jets), timing (integrable majorants).

Proposition III.10.4 (Gaussian sector of the step). Fix a patch \(D\), a request \(Q\) for which \(D\) satisfies the branch condition, an aperture \(B_0 > 0\), a sector cutoff \(\chi\), and \(k, p \in \mathbb{N}\). There are \(\sigma, d > 0\) and \(\pi \geqslant 1\) (Charges) such that the following holds for all large \(x\).

Take \(0 < \beta \leqslant 2 \beta_F\), a budget point of power \(p\) and threshold \(t_c (x)\), a step on \(D\) with \(c \geqslant 100 (1 + B)\) admissible for \(Q\), a forest, tags, and a store \(A\) whose flags equal the flags of \(V\) on the whole coarse torus. Freeze \(A\), the forest and tags while differentiating. Let \(\Gamma\) be the sector precision matrix, \(\tilde{\Gamma}\) its local approximate inverse, \(m\) the smooth sector mean at aperture \(B_0 t\), and \(h = - g^{- 1}\) times its stationarity residual, with \(g = \beta^{- 1 / 2}\). These are the objects in Value. Then \(\Gamma\) and \(\tilde{\Gamma}\) are positive definite, and \(m\) equals the stationary logarithmic mean for every frozen-link input.

For a coarse site \(o\) and \(0 \leqslant m \leqslant k\), let \(\mathcal{D}^m\) be an iterated left derivative in \(m\) directions satisfying \(|A_e | \leqslant t_c (x) (1 + \text{dist} (e, o))^p\). If \(x_i\) is the coarse location of coordinate \(i\), put

\(\displaystyle G_i = (1 + \text{dist} (x_i, o))^{pk} .\)

The bounds are

\(\displaystyle |\mathcal{D}^m \Gamma_{ij} | \leqslant \pi e^{- \sigma \text{dist} (x_i, x_j)} G_i^m, \qquad |\mathcal{D}^m (\Gamma \tilde{\Gamma} - 1)_{ij} | \leqslant e^{- dB} \pi e^{- \sigma \text{dist} (x_i, x_j)} G_i^m,\)
\(\displaystyle |\mathcal{D}^m m_i | \leqslant \pi tG_i^m, \qquad |\mathcal{D}^m h_i | \leqslant e^{- dB} \pi G_i^m .\)

Thus the spatial exponent at derivative order \(m\) is \(pkm\).

Status: theorem. JointLoad.stage also supplies a covariance cap and a trigger-tail estimate; its scales are given in §7.

Lean: JointLoad.capital.

Proposition III.10.5 (schedule).

Status: theorems. Part (a) is unconditional (pressure_pricing applies Bid.pricing_order to III.10.1). Divisibility is what lets torus tilings fit in III.12–III.13.

Lean: Wait.engine, notation_ticket.

Proposition III.10.6 (start-up fusion H01–H04; markers I01–I04).

(H) Setting. Fix a budget point and:

The fused gas (fused) groups the forced activities and the factors \(e^{\ell_i} - 1\) into connected clusters (pellets). Each pellet contains at least one forced activity, and the forced supports in a pellet are pairwise disjoint.

Hypotheses. Fix a bijective enumeration of the pellets. Let \(\kappa, \gamma \geqslant 0\).

Conclusions.

(I)

\(\displaystyle \mathcal{F}_r = 2 (2 + 4 P_r) + 1 + 6 (\alpha_r + P_r^2) + 4 d_G \left( \log 16 + \tfrac{1}{2} \log \hat{\beta}_r \right),\)

with \(P_r\) the stage's path budget, \(\alpha_r\) its direct clip coefficient and \(d_G = \dim_{\mathbb{R}} M_m (\mathbb{C})\) for the defining matrix representation. This is the hypothesis of YM.Stock.Tabs. Fix \(k \geqslant 1\) and levels \(m \leqslant i\) such that \(kB_i^2 \leqslant a_m / a_i\) (an open stop). For every \(N \geqslant i\) and every period, the centred copies at level \(N\) are the embedded sites from level \(m\); their parents agree at consecutive levels from \(m\) to \(N\); and distinct embedded sites are separated by at least \(kB_r^2\) for every \(i \leqslant r \leqslant N\).

\(\displaystyle w_i = \frac{\hat{\beta}_i^{- r / 2}}{p \hspace{0.17em} \text{columns}_i \prod_{m < i} \text{columns}_{m + 1}}\)

are positive, and for every \(\varepsilon > 0\) eventually \(5 - \log (w_i / \text{order}_i !) \leqslant \varepsilon \hspace{0.17em} \text{scale}_i\).

Status: theorems. In the main-theorem closure:

Lean:

2. What it uses

From other units.

Standard mathematics. The analytic method is Balaban's lattice-gauge renormalization group. Relevant references are Renormalization group approach to lattice gauge field theories, I, CMP 109 (1987), 249–301, II, CMP 116 (1988), 1–22, and Large field renormalization, I, CMP 122 (1989), 175–202, II, CMP 122 (1989), 355–392. The summability estimates use rooted polymer bounds of Kotecký–Preiss, CMP 103 (1986), 491–498, tree-graph estimates, Schur's test, Gaussian exponential moments, and Taylor estimates on compact charts. Exact correspondences with Balaban's lemmas remain open (§7).

Imported constants and assumptions.

3. Strategy

The method combines large-field suppression with Gaussian integration in the small-field region. Seeds record the fine plaquettes, mismatches and coarse flags that exceed their thresholds. The deterministic estimate gives an action cost \(\kappa t^2\) per seed. Since \(\beta \geqslant \beta_F / 2\), its Boltzmann factor is at most \(e^{- (\kappa / 2) e^{2 (\log \log \beta_F)^2}}\). This pays every fixed polylogarithmic entropy loss and is smaller than every inverse power of \(\beta_F\), while \(t = \beta_F^{- 1 / 2 + o (1)}\) remains perturbatively small.

An exact completed-square decomposition separates the fine, interior, coarse and gauge contributions from two kernel defects. Large \(L\) makes the fine prediction small; large \(W\) separates mismatch seeds from the interior prediction. Coarse plaquettes above a fixed threshold lie near spent seeds, whose larger cost \(tT\) pays the affected-cell losses. Away from those seeds, local gauge fixing and the flat tower identities reduce the defects to chart errors and exponentially small kernel tails. Summation over a radius-\(B\) neighbourhood still leaves an error \(o (t^2)\) per seed.

The resulting gas has packet, old forced, logarithmic and warning letters. The energy bound and annulus estimate control packets; Gaussian tails control warnings. Rooted support bounds and tree-graph estimates sum connected molecules. Derivatives are taken with sector data fixed and fluctuation coordinates transported compatibly. The Leibniz rule controls products, and Cauchy–Schwarz retains the square root of the simultaneous-warning probability when integrating their jets.

4. Steps

Unless a lemma gives different hypotheses, the energy estimates in §§4.2–4.6 use the step, tube, forest, fields, tags and consistent store of III.10.1. A sufficiently large threshold in \(x\) may depend on the fixed request, \(L, W\), and any additional parameters already chosen, but never on the period. Row and column bounds refer to sums of operator norms; vector bounds are pointwise in the coordinate index. Distances are native torus max-distances.

4.1 The algebraic identity

Lemma 4.1 (completed squares). Notation:

The rows are:

Their sum over the torus is \(\mathcal{A}_F (U) + \frac{a}{2} \sum_b w_b^2 -\mathcal{A}_c (V)\). The functional \(\mathcal{E}_A\) keeps only the rows located in \(D_{2 B} (A)\).

Proof: adjointness for the three kernels; all cross terms cancel. The cancellation is algebraic and works for any coarse coordinate vector \(v\) if the rows are defined with that vector.

Lean: Tube.locates; rows workF, workI, workC, workN, workL, workD.

4.2 Packing and geometry

Lemma 4.2 (packing; E01). For every \(\delta > 0\) there is \(N\), depending on \(P\) and \(\delta\) but not on \(L\) or \(W\), such that for all \(L\), all \(W > 0\), all large \(x\) and all admissible data with \(A\) consistent:

Proof: the macro box has radius \((j + 7) L\). Counting possible parent sites gives a bounded total multiplicity for each fine plaquette. If no spent seed is within distance \(j + 8\) of a coarse plaquette, consistency bounds every relevant input by \(T\); the predictor-curvature bound then gives \(|V_q | \leqslant C (L, W) T \leqslant \delta\) for large \(x\). Thus every tide has a nearby spent seed, and \(j + 8 \leqslant 4 B / 5\) eventually. A flag is incident to a coarse plaquette above \(t_c\); each plaquette has at most four incident sites. Four-dimensional ball counts give the remaining bounds.

Lean: Bid.E01 (Payments, MacroStore), Work.E01.answer.

Lemma 4.3 (shell comparison; E02). For every gauge group \(G\), block factor \(L\), index \(j\) and kernel choices \(K\), there is \(C > 0\) with the following property. Take a step using \(K\), block factor \(L\) and index \(j\), whose fine side is at least \(100\) times the macro radius, with \(c > 1\); any \(U\) in the tree gauge; any \(V\); and any fine link \(e\) with usable tags. Then

\(\displaystyle \chi (U_e \tilde{V}_e^{- 1}) \leqslant C \hspace{0.17em} \text{port} (\bar{e}, U) + w_{(\bar{e}, \mu)},\)

where \(\tilde{V}\) is the replica of \(V\) on fine links and \(\bar{e}\) the parent block. The last term is present only when the ends of \(e\) lie in different blocks.

Sketch: close \(e\) by the two tree paths, lift the loop to \(\mathbb{Z}^4\), and bound the chord of the product by the sum of chords along a walk whose length is bounded by the macro budget. For a crossing link, the predictor estimate controls the discrepancy between the two block roots; the triangle inequality then adds the coarse mismatch \(w_b\).

Lean: Bid.E02, Work.E02.answer.

4.3 Kernel bounds

Lemma 4.4 (rows and columns; E03). There is \(A > 0\), independent of \(L\) and \(W\), such that:

Take absolute values of the cover coefficients. Adjoint transport has norm one, periodization charges each cover term once per row or column, and the fine residue multiplicity is \(L^4\) (anchored_of_terms).

Lean: Bid.E03 (Allowance), Work.E03.answer.

Lemma 4.5 (path series and predictions; E04, given E03). There is \(A > 0\), independent of \(L\) and \(W\), such that, with \(s = t\) at the fine stage and \(s = t_c\) at the coarse stage:

Proof: sum the absolute path coefficients with their exponential length weights. Periodization gives the column count; the kept test bounds each contributing plaquette by the threshold times the same exponential weight. Thinning bounds the coarse coordinates, and the row estimates of Lemma 4.4 bound their predictions. The comparison \(t_c (x) \leqslant 2 t\) holds eventually at fixed \(L\).

Lean: Bid.E04 (LaneCounts, Maximum), Work.E04.answer.

4.4 Row-by-row lower bounds

Lemma 4.6 (coarse rows; E05). There are \(\delta, B_1 > 0\), independent of \(L\) and \(W\), such that for every coarse plaquette with \(\tau = |V_q |\):

Sketch: in matrix space the curvature is the tangential projection of \(V_q - 1\), so \(\chi^2 - | \text{curv} |^2\) is the squared normal displacement divided by the fixed matrix-to-colour normalization, which is \(O (\chi^4)\) in a tubular neighbourhood of \(G\).

Lean: Bid.E05 (Crowds), Work.E05.answer.

Lemma 4.7 (fine and interior rows; E06, given E03, E04). There is \(L_1\) such that for every \(L \geqslant L_1\) and every \(W \geqslant 2 A / \sqrt{a}\) (with \(A\) from E04), for large \(x\), with \(d = \alpha_F / 128\):

\(\displaystyle \tfrac{\lambda_F}{2} \hspace{0.17em} t \hspace{-0.17em} \hspace{-0.17em} \sum_{p \in A_F, \hspace{0.17em} |U_p | > t} \hspace{-0.17em} \hspace{-0.17em} |U_p | + d \hspace{0.17em} t^2 \hspace{0.17em} \sharp \{p \in A_F : |U_p | \leqslant t\} \leqslant \sum \text{fine rows}, \qquad \tfrac{a}{8} \sum_{b \in A_I} w_b^2 \leqslant \sum \text{interior rows} .\)

Proof:

Lean: Bid.E06 (Seals), Work.E06.answer.

Lemma 4.8 (affected cells; E07, given E01, E03, E04). For every fixed \(\delta > 0\) there is a pre-\(L\) constant \(C_7 > 0\), and for every \(L\) and \(W > 0\) a constant \(Z\), such that for large \(x\):

\(\displaystyle \text{defects} (\text{hit}) + \frac{C_7}{L^2} t \hspace{-0.17em} \hspace{-0.17em} \sum_{p \in A_F, |U_p | > t} \hspace{-0.17em} \hspace{-0.17em} |U_p | + C_7 t \sum_{b \in A_I} w_b + Zt^2 B^4 \sharp A^{\text{sp}} \geqslant \lambda_c t_c \sum_{\text{tides}} |V_q | .\)

Proof:

  1. By qdc, on the hit set \(|V_q | \leqslant r_0 t + H_q\). The macro excess \(H\) is bounded, through the charges of E01, by a pre-\(L\) constant times \(L^{- 2} \sum_{\text{swell}} |U_p | + \sum_{A_I} w_b\).

  2. The defect rows on hit cells are bounded below by Cauchy–Schwarz, using \(|v| \leqslant At\) and the kernel bounds of E03–E04. The inputs \(X\), \(Z\), \(X^c\) split into a uniform part, \(O (t)\) per cell, and the excess.

  3. The uniform part is paid by \(Zt^2 \sharp \text{hit} \leqslant ZNt^2 B^4 \sharp A^{\text{sp}}\); the excess is paid by the two middle terms.

No cancellation of linear terms is used.

Lean: Bid.E07 (Lapse), Work.E07.answer.

Lemma 4.9 (unaffected cells; E08, given E01, E02). For every \(L\) and \(W > 0\) there is \(C\) such that for large \(x\) the following holds. For every coarse plaquette \(q\) of \(D_{2 B} (A)\) outside the hit set there is a gauge transformation \(z\) with:

Sketch: there is no spent seed within \(4 B / 5\) of \(q\), so all tolls and wires near \(q\) are at most \(T\). An axial gauge on a box of the coarse cover, with E02 transferring the bound to fine links, loses a polynomial in \(B\).

Lean: Bid.E08 (Stitches), Work.E08.answer.

4.5 Linearization defects on unaffected cells

Lemma 4.10 (flat identities). Consider the linearized uncut kernels \(\Phi_0, \Phi_0', N_0\) (scalar matrices built from the towers), and the linearized inputs for log-fields \(u\) (fine) and \(v\) (coarse): \(X_0 (u)\) (face curls and path series of \(u\)) and \(Z_0 (u, v) = \sqrt{a_{\text{pen}}} \hspace{0.17em} (v - \text{probe} \hspace{0.17em} u)\), where \(a_{\text{pen}}\) is the penalty. Under the admissibility identities and agreement of the beds with the fine and coarse path data, for every \(u, v\) and every coarse two-layer vector \(w\),

\(\displaystyle \Phi_0^{\ast} X_0 (u) {+ \Phi_0'}^{\ast} Z_0 (u, v) - X_0^c (v) = 0, \qquad w - (\Phi_0^{\ast} \Phi_0 {+ \Phi_0'}^{\ast} \Phi_0' + N_0^{\ast} N_0) w = 0.\)

Proof: the flat tower identities, applied componentwise to the scalar matrices and the attached path data.

Lean: roll_l, roll_d.

Lemma 4.11 (chart estimates on the torus). For a patch \(D\) satisfying the branch condition there are \(C, \tau, r > 0\) with the following property. For every admissible step on \(D\), all tags, all \(R\), and all log-fields \(a\), \(b\) with \(\|a\|, \|b\| \leqslant h \leqslant r\), set \(U = \exp a\) and \(V = \exp b\). Define the defect vectors \(\mathfrak{l}= \Phi^{\ast} X + (\Phi')^{\ast} Z - X^c\) and \(\mathfrak{d} (w) = w - (\Phi^{\ast} \Phi + (\Phi')^{\ast} \Phi' + N^{\ast} N) w\). At every coarse plaquette, componentwise in its two coordinates,

Here \(\omega = \omega (R, h, \tau) = (1 + R) h + e^{- \tau R}\) (widen).

Proof: compare every actual object with its linearization, then subtract Lemma 4.10.

Lean: global_chart_cost.

Lemma 4.12 (interior cost). For the same \(D\) there are \(C, \tau, r > 0\) with the following property. Suppose that in some gauge \(z\) both \(U\) and \(V\) are within \(h \leqslant r\) of the identity on all bonds within distance \(3 R + 4 + \text{reach}\) of a coarse plaquette \(q\), where \(\text{reach}\) is the predictor reach and \(h \geqslant 0\). The step must be on \(D\) and admissible for \(P\); the tags and cut length \(R \in \mathbb{N}\) are arbitrary. Then the bounds of Lemma 4.11 hold at \(q\) for every coarse coordinate vector bounded by a nonnegative constant.

Proof:

  1. Gauge covariance: the defect vectors transform by \(\text{Ad} (z_q)\) (gauge_gaps).

  2. Locality: the defect vectors depend only on nearby bonds (local_gaps).

  3. Extension: a configuration that is \(h\)-close to \(1\) on a ball agrees there with \(\exp\) of a global log-field of norm \(O (h)\) (extend_interior).

  4. Apply Lemma 4.11.

Lean: interior_cost.

Lemma 4.13 (unaffected defects; EBare). Assume E03, E04 and the conclusion of E08 at \((P, L, W)\). Then for every \(\varepsilon > 0\) and all large \(x\), the sum of the defect rows over the unhit coarse plaquettes of \(D_{2 B} (A)\) is at least \(- \varepsilon t^2 \sharp A\).

Proof:

  1. Take \(h = CB^{10} T\) from E08 and \(\omega = \omega (R, h, \tau)\). Lemma 4.12 applies, since \(3 R + 4 + \text{reach} \leqslant 2 B / 5\).

  2. Cauchy–Schwarz with \(|v| \leqslant At\) gives, on every unhit cell, defects at least \(- \ell\), where \(\ell = 2 (At) (Ch \omega) + (At) (CAt \hspace{0.17em} \omega)\).

  3. There are at most \(NB^4 \sharp A\) such cells (E01 with \(\delta = 1\)).

  4. \(NB^4 \ell \leqslant \varepsilon t^2\) for large \(x\), because \(B^n \omega \to 0\) and \(B^n h \omega / t \to 0\) for every \(n\) (radius_stitch_errors). Indeed \(h = \beta_F^{- 0.49 + o (1)}\), \(h^2 / t = \beta_F^{- 0.48 + o (1)}\), and \(e^{- \tau R} \leqslant \beta_F^{- 1}\) because \(R \asymp (\log x)^2\).

  5. If no patch with the request's data satisfies the branch condition, the statement is vacuous.

Lean: Bid.EBare (Belly), Fluctuations.bare.

4.6 Assembly of III.10.1

Lemma 4.14 (assembly; E09). For a fixed request, the packing estimate of Lemma 4.2, fine/interior bounds of Lemma 4.7, coarse bounds of Lemma 4.6 at some \(\delta > 0\), affected-cell bound of Lemma 4.8 at that \(\delta\), and unaffected-cell bound of Lemma 4.13 for every \(L, W > 0\), imply III.10.1.

Choice of constants:

Proof: add the rows.

  1. Swell costs: \(\frac{\lambda_F}{2} t |U_p |\). Half of it pays the \(C_7 L^{- 2}\) term of Lemma 4.8; the remainder is at least \(2 mt^2\), and at least \(m \hspace{0.17em} tT\) when the cost is spent.

  2. Other costs: at least \(\alpha_F t^2 / 128 \geqslant mt^2\).

  3. Aisles: \(\frac{a}{8} w^2 - C_7 tw \geqslant \frac{a}{16} w^2 \geqslant mt^2\); spent aisles carry \(m \hspace{0.17em} tT\).

  4. Tides: paid by Lemma 4.8.

  5. Coarse plaquettes in the intermediate range: \(\frac{\alpha_c}{4} t_c |V_q | \geqslant \frac{\alpha_c}{16} t^2\), which pays \(4 mt^2\) per intermediate plaquette, since \(\sharp A_{\flat} \leqslant 4 (\sharp \text{intermediate} + N \sharp A^{\text{sp}})\).

  6. Small coarse plaquettes: in total at least \(- \frac{m}{4} t^2 \sharp A\).

  7. Gauge rows: \(\geqslant 0\).

  8. Defects: on hit cells they are inside Lemma 4.8; on unhit cells they total at least \(- \frac{m}{4} t^2 \sharp A\).

  9. Spent seeds: their \(m \hspace{0.17em} tT\) covers the flag charge against \(A^{\text{sp}}\) and the \(Z\) term.

The total is at least \(\frac{m}{2} t^2 \sharp A\).

Lean: Bid.E09, Work.E09.answer (with assemble and bank), Bid.issues.

4.7 Activities of one step

The next estimates use the hypotheses of III.10.2. In statements about packets, logarithmic letters or warnings, the tube satisfies the compatibility condition, the input satisfies the stated fine-level bounds, and sectors have the actual coarse flags. Auxiliary window constants are fixed before the threshold in \(x\).

Lemma 4.15 (rate calculus). A function \(w\) has rate \(a\) if for every \(v > a\), \(|w (x) | \leqslant \beta_F^{\hspace{0.17em} v}\) eventually. Products of functions with rates \(a, b\) have rate \(a + b\); their sums have rate \(\max (a, b)\). Eventual absolute domination preserves a rate, and constants have rate \(0\). Thus \(t\) has rate \(- \frac{1}{2}\), and \(B\) and \(\text{bill}\) have rate \(0\).

Lean: Rate.

Lemma 4.16 (from the action gap to packet weights).

((a) and (c) proof read; (b) statement only)

Lemma 4.17 (old and logarithmic letters). Fix \(- 1 / 2 < s < 0\) and any \(u \in \mathbb{R}\). Assume the sector drift and residual estimates, with a residual envelope of rate \(- 1\), as well as the common hypotheses of this subsection. For sufficiently large \(L\) and then all large \(x\), the old forced letters have a rooted bound with total \(\beta_F^u\), and all logarithmic letters have one with total \(\beta_F^s\). The latter include inherited regular activities, local Jacobians and Gaussian determinant corrections. The weights bound both \(| \ell_i |\) and \(|e^{\ell_i} - 1|\). Taking \(u = s\) gives the two totals used in III.10.2.

Proof: transport the incoming forced bounds to coarse supports, and split the logarithmic alphabet into inherited activities, Jacobians and Gaussian corrections. Their envelopes have rate \(- 1 / 2\) or smaller. Choose an intermediate rate strictly between \(- 1 / 2\) and \(s\) to absorb their finite sum, and use \(|e^z - 1| \leqslant 2 |z|\) once the individual logarithmic weights are at most one.

Lean: recommends.

Lemma 4.18 (window, drift, no new seeds). For each patch there are window constants such that, for large \(x\):

Proof: exponential row bounds and coercivity give the precision estimate; the local-inverse error tends to zero and yields a positive covariance. Applying the approximate inverse to the linear term gives the centring drift and a small linear residual. For the no-new-seed assertion, choose \(\xi\) so the Gaussian fluctuation contributes at most one third of the available margin. Linear cancellation from the compatibility identities and the small nonlinear remainder pay the other two thirds. The fine and mismatch prediction bounds then give the strict thresholds \(t / 4\) and \(Wt\).

Lemma 4.19 (catastrophes). For every set \(v\) of warning addresses, the Gaussian measure of the catastrophe set is at most \(E^{\sharp v}\), with \(E\) eventually smaller than every power of \(\beta_F\). The threshold \(\xi t \sqrt{\beta}\) equals \(\xi e^{(\log \log \beta_F)^2} (\beta / \beta_F)^{1 / 2}\) in Gaussian units.

Proof: the covariance has a volume-independent Schur bound. Apply the Gaussian rectangle estimate to simultaneous coordinate excursions and sum the possible witnesses in each radius-\(2 B\) neighbourhood. The resulting envelope can be taken as \(CB^4 e^{- cM^{1 / 2}}\), where \(M = e^{(\log \log \beta_F)^2}\) and \(c > 0\). It is smaller than every inverse power of \(\beta_F\). Repeated warning locations are paid through the fixed multiplicity bound.

Lean: Coat.alarm_bills.

Lemma 4.20 (assembly of III.10.2). The preceding estimates imply III.10.2 with the following parameter choices.

  1. The floor on \(L\) is the sum of the floors from Lemma 4.16(a), the no-new-seed margin, and Lemma 4.17.

  2. \(W\) is taken large for the energy estimate and the no-new-seed margin; \(\mu\) comes from Lemma 4.16(a) and does not depend on the patch.

  3. For each patch:

  4. All conditions holding for large \(x\) are merged into one threshold. The three letter totals have rate \(s < r\), so their sum is eventually at most \(\beta_F^{\hspace{0.17em} r}\).

Lean: Target.issue.

4.8 Derivatives

Lemma 4.21 (letter jets). Under the hypotheses of III.10.3, the letter-jet structure exists before integration. Choose \(- 1 / 2 < s < r\) and set \(\text{rad} = \beta_F^s\). The warning amplitude before probability extraction is a patch constant. The jet scale is a sum \(\rho (x) = B_0 (x) + A_0 (x)\) of two rate-zero envelopes, chosen independently of the input representation and volume.

Proof: combine the packet-jet estimate, the old and logarithmic letter estimates, and the differentiated warning cutoffs, all along compatible transported fluctuation coordinates. Choose the catastrophe envelope of rate \(2 s - 1\); its eventual bound by \(\beta_F^{2 s}\) gives the probability factor \(\text{rad}^2\). The three rooted letter totals have rate \(s < r\).

Lean: Utopia.whittle.

Lemma 4.22 (molecule integrals). The letter bounds of Lemma 4.21 imply the integrated majorants of III.10.3, for the compatible connected molecules specified there.

Proof: the Leibniz rule for Brush adds the letter jet scales. Their sum is bounded by twice the molecule's shape, giving, for rate-zero transport envelopes \(U, J'\) and fixed integers \(k, K \geqslant 0\), a pointwise factor

\(\displaystyle (1 + |U (x) | + 2 (\rho (x) + J' (x)))^i (1 + \text{shape})^i \hspace{0.17em} \text{facescale}^{ki} .\)

The bound vanishes outside the catastrophe set for the molecule's distinct warning addresses. If that set is \(\mathcal{C}\) and its address set is \(v\), Cauchy–Schwarz gives

\(\displaystyle \int \boldsymbol{1}_{\mathcal{C}} \hspace{0.17em} \text{facescale}^{ki} \hspace{0.17em} d \gamma \leqslant \left( \int \text{facescale}^{2 ki} \hspace{0.17em} d \gamma \right)^{1 / 2} \gamma (\mathcal{C})^{1 / 2} \leqslant e (x) (1 + \text{shape})^K \text{rad}^{\sharp v},\)

where \(e\) has rate zero. Integration over the probability measure on the tree slice adds no factor. Multiplying by the pointwise coefficient gives a rate-zero prefactor \(p\) and shape power \(l = K + i\). For a marked source, remove its ordinary fare and insert its amplitude bound.

Lean: timing.

4.9 Start-up fusion and markers

Lemma 4.23 (H01–H04). Under the respective hypotheses in III.10.6(H), the four fusion conclusions hold.

(proof read for H01, H02 and H04; H03 forced-family proof read, symmetry moves statement only)

Lean: H01.gathered, H02.answer, H03.fused_forcedFamily, H04.answer.

Lemma 4.24 (I01–I04). Under the respective hypotheses in III.10.6(I), the four marker conclusions hold.

Lean: I01.answer, I02.answer, I03.answer, I04.answer.

4.10 The schedule

Lemma 4.25 (pricing order). Under the hypotheses of III.10.5(a), the following choices produce its floors and raised entry couplings.

  1. For each pair of adjacent stages, III.10.1 gives a floor \(b_{ij}\); set \(k_i = \max (b_{ii}, b_{i, i - 1}, \text{sheet floor}_i)\).

  2. For \(L_i \geqslant k_i\), choose \(W \geqslant m\) in the admissible range and a threshold \(x_{ij}\).

  3. Raise \(\beta^{\ast}_i\) to \(\max (\beta^{\ast}_i, x_{ii}, x_{i, i - 1})\).

  4. The conclusion persists under further raising (recheck).

Lean: Bid.pricing_order.

Lemma 4.26 (divisible schedule). Under the hypotheses of III.10.5(b), the following construction gives a schedule satisfying all entry, plateau and divisibility requirements.

  1. Entry couplings come from the bar (using \(G\) simple), leads from the wait data, and a monotone stage-index function \(J\) with steps at most \(1\) and \(J \to \infty\) by the finite-request scheduling lemma of III.2.

  2. The scales are \(s_i\); the initial length \(M\) is the minimal length plus the largest bill over the first stages; the periods are \(4 (M + p + 1)\).

The checks:

Lean: Wait.engine.

5. Where the difficulty lies

The energy estimate must preserve \(L\)-independent coefficients on the large fine inputs and \(W\)-independent coefficients on the mismatches. The fine prediction is \(O (t / L^2)\), so increasing \(L\) absorbs its cross terms. The interior prediction is \(O (t)\), so increasing \(W\) absorbs the mismatch cross terms. Losses depending on both \(L, W\) are charged to spent seeds and absorbed only by increasing \(x\).

On unaffected cells, the transported, truncated kernels have no exact completeness identity. The flat identity cancels the linear part, while local gauge fixing, chart estimates and exponential tails control the remaining defects. This is the unaffected-defect estimate Bid.EBare. The constant \(\kappa\) is independent of \(L, W\); the coupling threshold is not.

The activity and derivative estimates additionally require compatible quadratic forms, controlled fluctuation transport and probability bounds for simultaneous warnings. Their product estimates retain one gain for each distinct warning address before the connected-molecule sum. The schedule is numerical; start-up fusion and marker estimates serve separate parts of the construction.

6. Notation

Written proof Lean (glossary) Meaning
\(x = \hat{\beta}_m\), \(\beta_F = x + b_G \log L\) betaF coarse and fine reference couplings
\(t_c (x)\), \(t = t_c (\beta_F)\) temp, fiery small-field thresholds; fiery is at the fine coupling
\(T = \beta_F^{- 49 / 100}\) largeT spent-seed threshold
\(B\), \(R\) radiusB, till cluster radius \(\lceil (\log x)^2 \rceil\), path cutoff \(\lfloor B / 10^5 \rfloor\)
\(\lvert U_p \rvert\), \(w_b\) toll, wire plaquette size; mismatch \(\chi (V_b Q_b^{- 1})\)
request \(P\); admissible Teller, Teller.Funds data of one call of the step estimator
\(A = (A_F, A_I, A_{\flat})\), consistent Stores, Stores.Plot seeds; two-sided thresholds on \(D_{5 B}\)
\(A^{\text{sp}}\), tides, hit spend, tides, hit seeds above \(T\); coarse plaquettes above \(\delta\); cells near \(A^{\text{sp}}\)
\(\mathcal{E}_A\) Stores.assessed localized action gap
\(v\), \(\Phi, \Phi', N\) vision, Φ, Φ', N coarse coordinates; transported kernels cut at \(R\)
\(\Phi_0, \Phi_0', N_0\) flatF, flatI, flatN flat, uncut scalar tower operators
\(\omega (R, h, \tau)\) widen \((1 + R) h + e^{- \tau R}\)
rate \(a\) Rate \(\leqslant \beta_F^{a + o (1)}\)
target data \(\mathcal{T}\) Target \(\varepsilon, \delta, e\), rate \(r\), halo, chart
letter weights, KP condition Sales, Buy rooted Kotecký–Preiss and tree-graph tests
new-gas bounds Cash, Utopia, Herd, Captain III.10.2 bundle; letter jets; molecule integrals
\(N_B (H)\) coverNumber minimum connected ball-cover size
\(\rho (x)\), \(\text{rad}\) Utopia.active, Utopia.rad jet scale of rate zero; square root of a warning-probability bound
Gaussian sector Value, Charges precision, approximate inverse, mean, residual

7. What remains unexplained

  1. Conversion to the norms of III.9. The rooted optional-letter bound and the packet/tree test are explicit here. Their complete conversion to the outgoing \(E\)- and \(K\)-norms of II.2 and III.5, including the reset, remains to be written out. The relevant per-step assemblies are the activity-bound bundle, its jet version, and the molecule-integral bounds.

  2. Free-kernel compatibility along the trajectory. The condition requires fine and coarse beds satisfying the diagonal and flat tower identities, invariance under lattice symmetries, attachment to the chosen path data, and positive invertibility bounds. YM.PhysicalNext.Repeat.Pane.ovation constructs it on the auxiliary torus from the trajectory's beds, assuming their lattice-symmetry property at every period. The proof of that symmetry property for the final selected trajectory, and the precise III.3 attribution of all tower inputs, have not been checked here.

  3. Depth of the analytic proof checks. The assembly proofs of III.10.2–III.10.4 and their defining estimates have been read. The underlying stationary coercivity, local-inverse, packet-jet and individual letter-jet estimates have not all been read through; their detailed analytic proofs remain to be checked. The proof of H03's forced-family structure was read; its full lift of lattice-symmetry permutations was not checked through. These limitations do not alter the conditional statements above.

  4. Additional reset estimates. meritBar and rimfire give summable jet bounds proportional to the input difference, under matched-input and derivative-control hypotheses. erudite gives positivity of a residual density and a \(\beta_F^{- 23 / 20}\) jet bound on the off-diagonal remainder under matching, finite-bound and small-chart hypotheses. Their complete hypotheses and role in III.9's reset are not included in this unit's numbered outputs.

  5. Gaussian trigger scales. In JointLoad.stage, the scales are \(b (x) = 1 + B (x)\) and \(p (x) = e^{(\log \log \beta_F)^2}\); the trigger estimate absorbs fixed powers of \(b\) into \(e^{- cp^2}\) per distinct endpoint. Its full use in the preconstruction has not been traced here.

  6. Comparison with Balaban. The cited papers describe the corresponding small-field, cluster and large-field operations. A lemma-by-lemma identification, including threshold and normalization choices, remains open. The bibliographic dates and volumes above do not establish such an identification.

← III.9 The marginal flow: the running coupling and its reset (notes)Part III·Unit outlineIII.11 Interacting rank, Gaussian comparison and the stress tensor (notes) →