The Setting
Signs that keep every prefix sum of vectors in short can be read off a Gaussian process, which must then stay in prescribed balls at many indices and times at once. Lemma 5.2 of the paper on the Euclidean Steinitz-Bergström theorem, One energy interval at one spectral scale, bounds the probability that it does.
Here the process is a stationary Gaussian chain in , read through real matrices of operator norm at most , each with a size . How likely is it that every stays below a fixed multiple of , when the matrices vary slowly along a parameter in and the steps lie in a window set by a spectral scale , and how fast does that probability decay over a long horizon? The library also answers the second question for the matrices that a filter builds at one scale.
The Gaussian Chain
Fix coefficients and a horizon , and let the noise , with , follow the standard Gaussian measure on ; every probability below is under this measure. The stationary Gaussian autoregressive chain in is given in each coordinate by (1) The autoregressive row of is the family of weights of the noises in , and the autoregressive evaluation at step is the matrix sending to .
A real matrix is a contraction when for every , and the squared Hilbert-Schmidt norm of a real matrix is the sum of the squares of its entries, (2)✦
The Filter at One Scale
A filter step on real matrices is a triple of a real matrix , a vector and a column ; it acts by , with the th unit vector of .✦ A filter step is good when is invertible and (3) A finite sequence of filter steps has fresh columns for a real matrix when every column of into which a step of writes is zero✦, and the run of on is the matrix obtained by applying the steps of to one after another, first to last.✦ For weights on the columns, the state of is the matrix .✦
A move on real matrices is a pair of such matrices, its contraction and its gain ; it carries a matrix to a matrix at a cost .✦ The move of a filter step at the weights is ✦, and the energy cost of moves from is (4)✦
The lift of a real matrix along is the matrix .✦
The cutoff is the continuous piecewise linear function on with (5)✦ and the size weight is the function for the exponent .✦ The constants are the gain constant ✦, the increment constant ✦, the moment constant ✦, the epoch factor ✦ and the survival radius (6)✦
The filter at scale of , , the weights and a family of prefixes of , indexed by a set , consists of real matrices acting on , sizes and a horizon . The matrices , the sizes and the horizon are built from the run of each on and its state at the weights , the weights rescaled by the scale , through the lift, the cutoff, the size weight, the energy cost of the moves of and the epoch factor.
The Claim
Theorem 1. Let , let be a natural number and let be the chain (1). In (a) and (b), let be a finite family of real contractions with sizes , and let and be real numbers with and for every .
(a) If steps and parameters satisfy and for all , then (7)✦
(b) If , and heights satisfy whenever , then (8)✦
(c) Let be a real matrix with orthonormal rows, , let be a finite sequence of good filter steps with fresh columns for whose steps write into pairwise distinct columns, and let be prefixes of . Let be positive weights on the columns with for every column , and let . Then the matrices , sizes and horizon of the filter at scale satisfy (9)✦
In the paper, the result reads: with the paper’s process at time and its energy cost at scale up to index , there is an absolute constant such that, if the values , , lie in an interval of length at most , the paper’s fixed bound on that length, then for every the probability that for every and every is at least . Part (a) is this statement in discrete time, the steps standing for the times and the parameters for the energies .
What Is Checked
The claims (7) and (8) are stated over the stationary Gaussian autoregressive chain (1), through its autoregressive rows and evaluations under the standard Gaussian measure, and over contractions and the Hilbert-Schmidt norm (2); the claim (9) is stated over the same chain and over good filter steps (3) with fresh columns, their runs and states, the moves of the steps and their energy cost (4), the lift, the cutoff (5), the size weight, the epoch factor and the survival radius (6).