The Setting
Let be a positive integer dimension, let and , and for any finite set equip with the norm (denoted ). From families of directions in , one builds a vertex set of rounded coordinate cells on and equips with the norm (denoted ). Given a map whose squared Euclidean distances approximate a reference chart distance on and a map satisfying , the question is how to bound the expected distance between values of a chart cell function for at two independent uniform points of , and how to combine the explicit error terms into a single constant and weight in .
Mollifier and Metric Constants
For a real number and a natural number , the inverse power is the real number ; in particular, , , and .✦ The smooth bump function is a bump function centered at with inner radius and outer radius , equal to on and supported in .✦ The standard mollifier is the normalization of to unit Lebesgue integral, so that is smooth, nonnegative, supported in , and satisfies .✦ The mollifier derivative integral is the real number ✦, and at smoothing width the mollifier Lipschitz constant is the real number .✦
The profile function is the function defined by ✦, and the profile constant is the real number .✦ In terms of and , the source metric constant is the real number ✦ and the contraction constant is the real number .✦ For any natural number , the contraction weight is the real number .✦
For natural numbers and and real parameters , the macro cost is the real number given by (1): (1) For a natural number , the macro cost error is the real number .
Directions and Frame Charts
Identify with its Euclidean inner product and norm , indexing coordinates by . For a natural number , a tuple , and vectors , the empirical covariance of along is the real number ✦, and the empirical sine transform of at is the vector in .✦ The bad projection set for a real threshold and a vector consists of all tuples such that for some . Given a natural number , a subset , and a family in , the number of bad samples is the count of indices with .✦
For a natural number , a direction family consists of a family of tuples in satisfying for all and , nonparallelism for all , , and , and covariance bounds for every and unit vector , together with the Fourier bound for every and every with , and the goodness bound for every unit vector .✦ A pivoted direction family in dimension is a direction family satisfying the pivot condition for all , , and .✦ Associated with a direction family are the scaled directions ✦, while the frame lower bound constant is the real number .✦
Given a continuous linear map between real inner-product spaces with adjoint and a constant satisfying for every , the Gram operator is a bijective continuous linear operator✦, and the left pseudoinverse is the continuous linear map .✦ Equipping with the norm (denoted , with ), for a direction family in dimension with scaled directions and a chart index , the frame chart operator is the continuous linear map with components ✦, and the chart pseudoinverse is the left pseudoinverse of .✦
Rounded Charts and Cell Functions
For a real radius , the cube of radius in is the open cube , so that and .✦ Let be the uniform probability measure on , defined as Lebesgue measure restricted to scaled by ✦; the cube uniform measure on is the -fold product of across the coordinates of .✦
Given a natural number , a family , and a threshold scale , a point is a regular chart angle if for all and ✦, and the integer chart label of is the tuple with entries .✦ For natural numbers and and families in , the rounded chart vertex set is the set of all pairs with and for which some regular chart angle satisfies ✦; for a direction family with scaled directions , the source vertex set is the vertex set .✦ The rounded chart coordinate map is the map sending each vertex to .✦
When , set with the Euclidean norm. The continuous coordinate split is the continuous linear isomorphism sending to , with inverse prepending to the coordinates of for each .✦ For with , the non-pivot linear functional is the continuous linear functional ✦, and the hyperplane slope functional is the continuous linear functional .✦ Given a finite index set and a family in , the chart pivot coordinate is the real number ✦ and the hyperplane slope family is the family .✦
For a continuous linear functional , the slice linear form is defined by .✦ Given a finite index set , a family of continuous linear functionals on , and pivot scalars , the pivoted slice functional at is the continuous linear functional .✦ For a finite set , real offsets , and a function assigning real numbers to subsets of , the slice cell function is the function sending to .✦
On any real normed space , given a finite index set , continuous linear functionals on , a scale , and a subset , a point is an -regular cell point if for all and ✦, and its cell integer label is the function from to with entries ✦; the cell label vertex set consists of all functions equal to for some -regular point .✦ Given pivot scalars and a finite set , the cell sign pattern of is the subset ✦, and for any set , the cell pattern extension is the map assigning to each map a function from subsets of to .✦ For a subset , if some -regular point has cell sign pattern , then equals (which is independent of the choice of ), and otherwise .
In , the split cube domain is the preimage of under ✦, and at threshold scale the chart label range is the integer interval with .✦ Given a finite index set , a family in , scale , a finite set , a subset , a set , and a map , the cell pattern values map is the function ✦, and for each , the Euclidean chart function is the composition (with offsets and ).✦ For natural numbers and , families in with , a threshold scale , and a chart index , the cell-to-chart vertex map is the map sending each label to , and for any map and index , the concrete cell function is the function .✦
The Claim
Theorem 1. Let be an integer satisfying and , and let be a pivoted direction family in dimension with scaled directions and source vertex set . Suppose satisfies the macroscopic approximation bound for all and in , as well as the local step bound whenever and . Let be a finite set and be a map satisfying for all . Then for every chart index , the average difference of the concrete cell function under satisfies (2): (2)✦ Furthermore, for every integer satisfying , the constants satisfy the numerical inequality (3): (3)✦
In the paper, the result reads: for the metric space and chart maps from Proposition 4.2, there is an absolute constant such that every map (for every finite ) satisfying obeys , where and are independent and uniform, and here , , , , and stand for the paper’s symbols while stands for . Rather than using the chart metric and the composition on chart , (2) holds for any chart via a map satisfying the macroscopic and local chart bounds and the concrete cell function of , accompanied by the numerical inequality (3).
What Is Checked
The integral bound (2) is stated over an integer satisfying and , a pivoted direction family in dimension , a map obeying the macroscopic approximation bound against and the local unit-step bound for vertices and in , a finite set with a map whose increments are dominated by , and any chart index , while the numerical bound (3) is stated over any integer satisfying . For any positive integer and any pivoted direction family in dimension , the frame lower bound constant is strictly positive✦, each frame chart operator satisfies the energy lower bound for all , and every scaled direction coordinate satisfies for all , , and .✦