Convergence of the interacting scale histories · Part 6

Searchable transcription

folding is
4
X
j=0
|ρj| 2k(1 + z)B.power k qj.
(RB4b)
gleam =

The radii and reference cardinalities have rate zero, and has rate20; so does this allowance. The small chart required here is supplied on any healthy neighborhood by axial gauge: its link distances are bounded by a fixed multiple of (2 times the threshold. The interior extension and its inverse chart place it inside radius2 furl. The radius has rate zero and the threshold rate2, whereas decays more slowly; the scheduling thus gives the strict margin eventually.

The reference offers are indexed by the coupling itself: at a curve parameter, the selected offer is the one at). It is therefore the same offer for two presentations with the same coupling, independently of how that point was reached. The fixed-reference finite-volume estimates hold eventually in their approximation depth simultaneously on the compact coupling band; (RB4b) gives the number denoted by gleam in (FCC1). The compatibility needed above comes from the same local cylinder formulas in the different reference views.

Here is also the weighted-tail step behind the retirement argument. At one finite torus all supports have cover number bounded by its site count. Thus multiplying any summable envelope by its support exponential, even the early exponent + 3, costs only a finite constant depending on that torus. Changing its derivative anchor also costs a finite constant:

Searchable transcription

K =
(1 + D)p
k,
(RB5)

where is the maximal torus separation; the direction triangle inequality gives this bound. Consequently the local source estimate supplies one summable envelope in the common anchored listing. If the debris indicator of each fixed address is eventually zero, dominated convergence for series makes its weighted debris sum tend to zero at that parameter. This eventual removal is literal: a nonkept address eventually belongs to the finite retirement roster, whereas a kept address never belongs to the debris.

Searchable transcription

Continuity in the deviation budget extends that small-tail bound to a parameter neighborhood.
More explicitly, spend half a requested allowance on the tail at the center and half on the full source
difference to a neighboring parameter, converting the latter to the common anchored weighted norm
by the finite constant (RB5) and its support-weight counterpart. Restricting the difference to the
debris subfamily cannot increase its positive majorant. Thus the neighborhood works for every
sufficiently large depth.
A finite subcover of the compact band gives a single depth.
Combin-
ing it with the uniform beacon depth proves the simultaneous retirement used in (FCC1). The
canonical fixed-layout curve argument already given then transfers this depth to every compatible
presentation; the construction here supplies its weighted continuity and actual finite beacons.

Here are the reference and continuity inputs with their finite-volume dependence exposed. In this section alone is a fixed reference side, is the fixed budget,order, power, andsupportExponent. The five signed fourth-difference coefficients in (RB1) are denoted throughout by; the formerly used for them in (RB4b) mean the same coefficients.

Searchable transcription

They are unrelated to the parameter λ in the local fusion argument.
The representation constant used here is royalty(G) = √m, where m is the dimension of the
fixed faithful unitary representation. In particular it is at least one. It is distinct from the Fourier
aliasing price, which depends on a strip width and an analytic norm.
The reference-chart radius and margin have the explicit values

log log log (RB6)

The logarithm of the normalized finite Gaussian integral is centered by subtracting its value at the identity. On the reference chart, with no flags, its derivatives through the purchased order satisfyD . The lower bound on its unnormalized integral is strictly positive. These statements are obtained on the actual reference forest and its Gaussian marginal, with the same stationary section and coupling, by the Gaussian differentiation and connected estimates (ASP2a)–(ASP2c), the local integral normalization (PRI2)–(PRI5), and the logarithm series (ASP1)–(ASP2). In particular the bound concerns the centered logarithm, not the integral

itself. The reference side must exceed the prescribed patch and covariance read radius. The path width, path norm, loop allowance, correction parameter and derivative horizon are fixed first; the odd block length is then above their finite lower bounds; finally is above their common delay. Bounds for the two adjacent loop allowances and the correction parameter put both sides in these same estimates. The exact logarithmic density identity on the strict healthy chart has the form

Searchable transcription

f(P) = c −
X
y∈Site(n)
Fy(actual P) −
X
α
Aα(actual P).
(RB7)

Here is the retained finite ordinary reference density and the emitted regular source. Their actual occurrence indices and supports are retained in the sum. The nonretained source consists of the omitted ordinary occurrences and omitted short stationary paths, not of a difference between Gaussian integrals on different tori. Its anchored early allowance is at most once its approximation depth is large. The retained-source part has allowance at most turnFee(). Summing these two bounds gives turnFee( per site. Summation over sites and centering (RB7) gives the second term of (RB4a); the reference logarithm gives its first term after the logarithmic chart chain rule already proved there. This explains every summand in

Searchable transcription

We give the tail argument that supplies the omitted-source allowance used in (RB7). A regular
address is a pair (oα, Xα), its anchor and finite support. For nonnegative numbers aα, its early
bound u means
X
sup
aα exp{(e + 3) covB(Xα)} ≤u.
(RB8)
o
α:oα=o

Searchable transcription

Sum over the V anchors. Since each exponential is at least one, P
α aα ≤V u. This proves absolute
summability of the derivative envelopes, including derivatives in any fixed finite list of directions:
divide those directions by one common scalar until they satisfy the threshold bound, apply (RB8),
then multiply by that scalar to the derivative order. The same scalar works for every address.
Termwise differentiation on the strict healthy open set therefore follows from uniform absolute
convergence of the derivative series through order k.
Changing every anchor to the one in the common listing multiplies its envelope by at most
Kanc = ((1 + D)p)k, as in (RB5). On this fixed torus,

Searchable transcription

covB(X) ≤V,
Ksup = exp{(e + 3)V }

bounds every early support weight. Thus an unweighted summable envelope gives the summable weighted envelope. For the debris predicate

is not kept and does not belong to the depth- finite roster

Searchable transcription

one has 1Pm(α) →0 at every fixed address: ordinary occurrences exhaust their finite shelves and
short paths exhaust their length cutoffs. Dominated convergence for the nonnegative series yields
P
α 1Pm(α)gα →0. This proves a small early bound, rather than merely a small unweighted sum.

Searchable transcription

For clarity, source deviation has the following meaning. Two sources have identical regular and
forced address layouts. There are nonnegative envelopes cα, dα such that on each regular healthy
domain every left derivative through order k of their difference is bounded by cα, and on every
configuration every forced coefficient difference is bounded by dα. Both total weighted sums
X
α
ee covB(Xα)cα,
X
α
ee covB(Yα)dα

are at most the deviation allowance. Regular directions use the anchor-dependent budget (AB1)– (AB4). Forced coefficients use their full configuration domain.

The dependence on the coupling and input activities is continuous in this deviation sense. On a finite set of occurrence addresses, the local integrands and their derivatives vary continuously: the Gaussian covariance and mean, finite Haar factors, cutoffprofiles and stationary coordinates are the actual ones already constructed. Their local integral bounds (ASP2b)–(ASP2c) provide a common integrable majorant. The normalized denominator stays positive in a sufficiently small neighborhood. Differentiation and dominated convergence hence give continuity of each regular derivative supremum on its compact direction and healthy test sets; the full-domain forced bounds use the corresponding full-domain envelopes. For the infinite family, the signed connected expansion gives a common summable majorant on that neighborhood, by (ASP3)–(ASP5). Given 0, first choose a finite address set whose common tail is below4, then use finite continuity to make its sum below2. The two tails contribute at most2. This proves deviation continuity with the actual support weights.

More explicitly, the dependence of the stationary path-only terms on the coupling is affine: if ) denotes one such term, then

Searchable transcription

Aα(β) −Aα(β′) = (β −β′){Aα(1) −Aα(0)}.

The two early envelopes at couplings zero and one give a summable Lipschitz envelope for this difference. The other connected terms are compared by the same marked interpolation before summation. If the input source discrepancy is, their regular and forced absolute envelopes have sums bounded by and for fixed finite, whenever 0. Signed translations cost

Searchable transcription

Z = |Seat(n)| Rim(n) turnFee(B)M.

Reindexing does not change a sum; multiplication by the symmetry normalization and the directionchange allowance gives exactly these factors. Converting a total unweighted envelope to a weighted one costs at most . Thus the outgoing source deviation is bounded by All these constants may depend on this reference torus, budget, fixed path bounds and derivative horizon, but are independent of the approximation depth.

Finally fix a parameter value and an early target 0. Pay2 for its debris tail using (RB8). Choose a parameter neighborhood with full source deviation below); then its restricted debris differences have early allowance at most2. The triangle inequality gives the same tail bound throughout that neighborhood for all sufficiently large depths. A finite subcover of the compact coupling band makes the depth uniform. Applying this to the five fixed reference tori proves the common allowance in (RB7), then (RB4a) and (RB4b), and hence the beacon term in (FCC1). No torus-uniform version of or is used.

Derivation and transfer of the reference chart estimate

The reference chart comparison uses a stationary expansion on fixed reference tori. Constants at this point may depend on the chosen reference torus and finite horizon. The subsequent periodic-copy

and activity arguments identify where that dependence is admissible. A finite reference calibration should therefore not be read as a uniform inverse estimate at every level.

We first derive the protected-logarithm jet input. Fix the stage and derivative order before the delay. Put, and fineMonitor(). Work at centers with; eventually and the stationary radius exceeds. Every) then satisfies, which is the margin needed for the Gaussian logarithm estimates throughout this ball. Let be the actual protected integral,) its truncated Gaussian approximation with fluctuation cutoff, and = Ωexp(coverLog) 0. The Gaussian comparison on the whole closed ball), using the vertical gap

Searchable transcription

0 ≤I(z) −pV (δ, z) ≤pe−gD2/8,
V (δ, z) ≥a = e−1/2.
(PL1)

The positive omitted tail is bounded by the quadratic Gaussian tail estimate; the local comparison includes the chart and Haar factors. Here Ωis field independent, whereas

is a positive function of the center. The inequalities and the following logarithm argument are pointwise in that center. The cutoffs equal one near zero fluctuation, where the density is strictly positive, so 0. Its Gaussian determinant is positive by the vertical gap. Writing gives 0. Hence the logarithmic discrepancy

Searchable transcription

f = log I −log Ω+ coverLog −log V
satisfies
|f| ≤a−1e−gD2/8,
(PL2)

Searchable transcription

by 0 ≤log(1 + q) ≤q. This is only a value estimate.
We record how the actual integral comparison in (PL1) is obtained, so that its positive error and
its domain are explicit. On the fixed forest fiber let Uz be the set represented by the fluctuation chart
U = draw(Vz, μ(z) + δi), ∥i∥≤D, where μ is the stationary graph. The chart radius inequalities
are (1 + trailCost)u + δD < initial and the corresponding color-scaled displacement is below the
group-chart radius. On this set all protecting flags vanish. The exact change of variables, including
its Haar Jacobian and the vertical Gaussian determinant, gives
Z
Uz
Iz(U) dνF (U) = p(z)V (δ, z),
I(z) = p(z)V (δ, z) +
Z
Ucz
Iz(U) dνF (U).
(PL1a)

Here is the nonnegative summed fiber density used in the exact push. This is why the comparison has a one-sided error.

The outside-field energy bound is paid in the same chart. Denote by the number of free bonds, by 0 the single-bond chart normalization, by the common finite size bound, and by respectively the local derivative and determinant expenses. With4, its normalized logarithmic expense is

Searchable transcription

YD = −k log wg(β) −RD + vC loopsPrice + K0 + T0.
(PL1b)

Searchable transcription

The fiber complement estimate is
R
Ucz Iz ≤p(z)eYD. It uses the local link reconstruction to put every
small-toll, small-wire configuration in the safety chart. Otherwise its fine or prediction penalty pays
at least the threshold energy. The parameter inequality

Searchable transcription

β meter(vC)u + RD ≤βppenaltym2
0/2

makes that alternative at least as expensive as the fluctuation tail; is a fixed positive safety threshold, not the shrinking radius. The quadratic lower bound within the chart pays the other alternative. Both comparisons use the same actual field and forest fiber.

The delay makes2, and also the stronger inequality2. Here the extra expense used to locate a clean chart point is

Searchable transcription

E = β{vCCcl∥z∥5 + vCC′(θ′3 + θ′4)} + K0 + T0 + vCAcoarse/2.

The two terms multiplied by and the loop expense are eventually bounded by one: the field radius has rate16, the threshold has rate2, and the remaining size factors have rate zero. The nondecaying expense is affine in the fine monitor,)fineMonitor(), with bounded by the already fixed polynomial size budgets. The normalization and determinant expenses have the same type of bound. The square4)fineMonitor( dominates this fixed finite list, after the stage parameters and positive are fixed. This proves both inequalities. The stronger one, together with the positive lower bound for the total push, forces a clean point inside; the local chord and displacement estimates then keep all chart flags zero on the surrounding radius- ball. Thus (PL1a) applies throughout that ball. The first inequality and (PL1b) give its error , as claimed. The truncated Gaussian lower bound and its jet estimates use the same size and radius choices. Their derivation is the local Gaussian expansion above; no full-field lower bound for an inverse density is used here.

The Gaussian approximation uses the actual functions constructed above. Keep the retained maximum-norm real coordinates, and give the vertical real coordinate array its Euclidean norm, denoting it by. The linear map) that joins these arrays into the joint maximum-norm array has norm at most one. In this notation set

Searchable transcription

φ(x, h) = F(L(x, h)),
ψ(x, h) = loop(U(x, h)) −
X
e free
log JG(tint(h)e).
(GL0)

Here is the main-action-plus-prediction score (AS1), is its actual fine configuration, loop is the retained quadratic-loop contribution, and is the local Haar density of the chart in the earlier primitive integral. The minus sign makes contain the positive Haar factor. Let) be the vertical Euclidean version of the unique graph (US7), and let) be its Hessian in these vertical coordinates. The equality of counting and Euclidean color squared norms proves

Searchable transcription

Dhφ(x, ρ(x)) = 0,
D2
hφ(x, ρ(x))[z, z] = zTM(x)z,
M(x) ≥(g/2)I.
(GL0a)

Thus the inverse covariance bound used below is; it is the retained half-gap of the actual stationary Hessian.

We explain the common derivative hypotheses for these functions. The joined points used below satisfy

where is the stationary Lipschitz price, enlarged to include the retained identity coordinate. The radius inequalities before (PL1a) put this whole set inside the purchased stationary domain and put every color logarithm inside half the Haar-chart radius. The graph jets are (US8). The score jets follow from (AS6) after the norm-at-most-one map; converting coordinate multilinear bounds to operator bounds costs at most the number of joint coordinates to the derivative order. The loop jets follow by summing the absolutely weighted path jets, multiplied by the fine-site count. A Haar summand is the fixed analytic log composed with a color projection of norm at most 1+dim g, so its order- price is at most its purchased chart price times (2+dim g; summing costs the free-bond count.

The centered determinant jets required below are controlled on this same domain. Differentiate log det) and. Repeating gives a finite sum of

traces of products of inverse factors and derivatives of, indexed by ordered derivative partitions. Each trace is bounded by the vertical dimension times the product of the operator norms. The inverse factor has bound 2, and derivatives of are obtained from two score derivatives and the graph chain rule, all at the already purchased higher orders. This gives a finite polynomial price in those jet bounds, the coordinate counts and 2. The value of the centered determinant logarithm follows by integrating its first derivative from zero. These operations give the common price for the graph, classical score, quantum contribution and centered half-log-determinant used below. Its dependence on the auxiliary size bounds is polynomial at fixed derivative order; those sizes have the assigned polynomial logarithmic bounds. Consequently has rate zero. This establishes the hypotheses of the local Gaussian calculation on the entire displaced set, using the actual score, chart and graph.

Define the Taylor-integral expression

Searchable transcription

A(b, x, z) =
Z 1
0
(1 −t)D2
hφ(x, ρ(x) + tbz)[z, z], dt.
(GL1)

Searchable transcription

Stationarity and twice integrating a line derivative give φ(x, ρ(x) + bz) −φ(x, ρ(x)) = b2A(b, x, z),
including b = 0 by the displayed integral. In particular A(0, x, z) = 1
2zTM(x)z. Put

Searchable transcription

G(b, x, z) = 1
2 log det M(x) −1
2 log det M(0) + A(0, 0, z) −A(b, x, z)
+ ψ(x, ρ(x)) −ψ(x, ρ(x) + bz).
(GL2)

Then the same truncated approximation in (PL1) is

Searchable transcription

V (b, x) =
Z
∥z∥≤D
eG(b,x,z)dγM(0)−1(z).
(GL3)

At = 0, multiplying its fixed Gaussian density by the exponential in (GL2) gives exactly the density of. Hence (GL4)

This identity retains the moving determinant and covariance; replacing either by its value at zero would lose the derivative argument.

Searchable transcription

Here is a uniform estimate for the mass in (GL4). If a centered Gaussian covariance Q satisfies
Q ≤cI, its one-coordinate exponential moment gives P(|Zi| ≥v) ≤2e−sv+cs2/2 for s ≥0: apply
Markov’s inequality to e±sZi and add. If the number of coordinates is at most P ≥1, and Pv ≤D,
the complement of the radius-D ball is contained in the union of those coordinate events. Indeed,
√
Pv ≤Pv ≤D. Thus its probability is at most 2Pe−sv+cs2/2.
if every |Zi| < v, then ∥Z∥≤
Take P = 1 + Q0(t), where Q0 is the fixed polynomial size budget, and write H = fineMonitor(t).
√
√
Eventually P ≤
H. For D = dH, choose v = d
H and s = v/c. The error is at most

Searchable transcription

2P exp(−d2H/(2c)).
(GL5)

The monitor is chosen to dominate these polynomial size budgets and logarithmic coupling powers, so (GL5) is smaller than for every fixed exponent, after increasing the delay. This estimate is uniform for the moving covariances, since the inverse-Hessian bound gives one common For completeness the derivative price in (GL3) is finite without any derivative of a moving integration set. At order, put

Searchable transcription

Us = 2(1 + C + D),
As = D2Bs(Us, C),
Fs = 2C + 2As + 2Bs(Us, C).

Here bounds the purchased section, score, quantum and centered determinant jets, with two extra score and section orders for (GL1). The affine displacement map in (GL1) has price evaluating its second score derivative on two copies of gives the factor. Integration against , followed by the sum in (GL2), bounds all jets of by. Since) = 0, its first derivative also gives max(). The shrinking chart makes this at most one. Consequently on the fixed cutoffball. Its base Gaussian mass is eventually at least one half by (GL5), so. The exponential composition rule bounds its jets by ); the fixed Gaussian measure is a probability, so integration preserves this bound. The logarithm on the compact interval [] has a finite fixed jet price, giving the joint ()-jet bound). All these prices have rate zero.

Finally (GL2) is unchanged by simultaneously replacing () with (). Symmetry of the cutoffball and base Gaussian makes), and therefore its logarithm is even in . Each of its-derivatives has zero first-derivative at zero. Integrating its second-derivative bounds the difference from its = 0 value by, through order, using the purchased joint order = 0, (GL4)–(GL5) and log) for 11 give an arbitrarily small polynomial value error on the surrounding ball. Apply the interpolation (PL4) below with a rough joint bound and a sufficiently small chosen value exponent. This bounds the-jets of log. Combining the even difference and this zero-parameter estimate gives the asserted Gaussian logarithm jet price at, as well as its lower bound. This proves the Gaussian input used in (PL1) and (RBG2).

The separate rough derivative estimate for the discrepancy on that closed ball is, for a requested order sup (PL3)

For this rough estimate one can avoid estimating an inverse integral on a whole chart. We give the relative derivative bound used in the construction. Write the exact push as an integral over a fixed forest fiber of a finite sum of nonnegative functions

Searchable transcription

I(x) =
Z X
α
aα(U)e−βNα(x,U) dνF (U),
aα(U) ≥0.
(PL3a)

The coefficients include the fine density and all factors independent of the retained coordinate The only-dependence here is the prediction penalty, with the fiber configuration fixed. In particular this argument does not require a nearby point to retain a selected protecting flag. Let bound the number of coarse bonds. The link product and trace estimates in (AS3) give the penalty jet price colorScale)

The predictor at fixed is a constant group element of matrix norm colorScale; hence this price is uniform in and the finite choice For clarity write the composition polynomial as

Searchable transcription

Bn(a, b) = n!b(2 + a)n.

It bounds the finite partition sum for composition when the inner and outer jets have respective prices at scale one. Subtracting the value of at the center costs a factor two in its jet bound. If bounds the exponential jets at zero, each integrand in (PL3a) therefore has jets through order bounded by its own value times

Searchable transcription

Kn = Bn(2|β|AN, En).

Integrating preserves this relative bound. Differentiation under the integral follows from the same estimate on a neighborhood together with the bounded fine-density envelope and the finite Haar fiber measure. Consequently) for. At each center where (PL1) gives 0, normalize by this single positive value: Its jets at are bounded by, and bounds the logarithm jets at one, the chain rule gives jets of loglog) = log) bounded by). The size of has canceled before estimating. This is the normalization that supplies rough logarithm derivatives uniformly over the surrounding ball.

The other terms of have explicit prices as well. If) is the stationary chart-jet price, set

Searchable transcription

Tn(t) = Bn(2 + 2C(t), C(t)),
On(β, t) = βTn(t) + Tn(t) + C(t).

Composition of the classical action, quantum action and Gaussian determinant on the stationary graph bounds the centered cover logarithm by 2. Let) be the rate-zero Gaussian-log coefficient and put; since1, the Gaussian logarithm jet price is at most Combining the centered logarithms and then restoring the value bound (PL2), which is eventually at most one, gives the rough price

Searchable transcription

M(t) = Bn(Kn(2y, r(t)), Ln) + 2On(2y, t) + 2Q(t) + 1,
n = j + 1.
(PL3b)

The actual coupling lies below 2, so positivity and monotonicity of these polynomials justify that substitution. The bond-count bound), chart coefficient) and) have rate zero. Thus has rate zero, has rate, and) has rate. The cover term has rate one and the Gaussian term rate zero. Their sum has rate at most + 1, and hence is eventually at most. This proves (PL3), including the zeroth derivative by (PL2). It uses actual integral derivatives and the surrounding ball, rather than differentiating the small value error.

Here is the interpolation and its order of choices. If on a ball of radius and , integration along unit directions gives

Searchable transcription

∥Di+1h(0)∥≤2vi/s + Cs.
(PL4)

For multilinear derivatives one tests unit vectors before applying this line estimate. Suppose rough derivatives through order + 1 are bounded by on a ball of radius 0. For any target exponent, a sufficiently small value bound on that ball implies every jet through order at its center is bounded by. At order zero take. For the step from order to + 1, choose

Searchable transcription

c = a + max(0, d −b) + 1,
b′ = b −1 −c,

then use the order- assertion with target and ball exponent to determine Eventually , so the smaller balls around points within of the center lie in the original ball. The induction gives there. Formula (PL4) gives 2 , whose two exponents are strictly below; their sum is eventually at most. Lower jets already satisfy the stronger bound.

Apply this recursion with. It fixes first; then choose the tail accuracy and delay so that (PL2) is at most. Thus the discrepancy jets are bounded by Gaussian differentiation bounds the separate local Gaussian logarithm jets by , with of rate zero. Keeping that term in leaves the discrepancy in (RBG1). Since

Searchable transcription

max
i≤n ui∥DiY (v)∥≤y−2.
(PL5)

This includes = 0. The critical extension and actual chart composition use the already purchased higher jets and finite composition prices. Those prices are fixed before the tail and reference delays.