The Setting
Vectors in and signs give the signed walk of prefix sums , for . Here the signs are read off one point of the coefficient space , the Euclidean space of real vectors indexed by a finite set . Each vector owns a coordinate , and its sign is the difference between and a prediction , where is a linear functional on that vanishes on the standard basis vectors with .
The point is sought in a convex body on which every prediction stays in the slab . It is then fed to a linear filter started from a real matrix with rows and columns indexed by , whose state after vectors changes once per vector, and the question is what is.
Clipping and Good Bodies
The clip at is the function given by (1)✦
Let be a Euclidean space, the complex space of square-integrable functions for Lebesgue measure, and the unitary Fourier transform on it. For a measurable real function on with , the Fourier multiplier with multiplier is the bounded operator on , with the multiplication by .✦ The heat operator of a measurable function on is the Fourier multiplier .✦ The compressed heat operator of a measurable set and such a function is the operator on , where is the multiplication by the indicator of .✦
For nonnegative weights , the Gaussian symbol of is the function on given by (2)✦ A measurable set satisfies the diagonal heat bound with constant if, for all nonnegative weights , the operator norm on obeys (3)✦ A good convex body in is a nonempty, open, bounded, convex subset of that is symmetric, with for every , and satisfies the diagonal heat bound with constant .✦
Filter Steps
A filter step is a triple of a real matrix , a vector and a column index ✦; it acts on a real matrix with rows and columns indexed by by (4)✦ where the column insertion is the matrix with in column and zeros elsewhere. The run of steps on applies them in order, and .✦ The matrix acts on as the linear map into ✦, and a row indexed by acts as the linear functional ✦.
The mass of a vector is .✦ The normalizer and the symmetric contraction of a vector are the real number and the matrix given by (5) The raw predictor row of a matrix against is the row indexed by (6)✦
The Claim
Theorem 1. (a) Let be a good convex body in , let be distinct coordinates in , and let be linear functionals on with whenever and with for every and every . Then there are signs and a point with for every outside and (7)✦
(b) Let be a filter step whose matrix is the symmetric contraction of its vector, let be a real matrix with rows and columns indexed by , and let and a real number satisfy . Then (8)✦
(c) Let be filter steps , each with the symmetric contraction of its vector as matrix, let be a real matrix with rows and columns indexed by , with run , and let and real numbers satisfy for every . Then, for every , (9)✦
In the paper, the result reads: the clipped shifts of the predictions are inactive, , and then for every .
What Is Checked
The item defines the clip (1), and its claim is stated over good convex bodies, with the diagonal heat bound (3) built from the Gaussian symbol (2), the Fourier multiplier, the heat operator and the compressed heat operator, and over filter steps (4) with their column insertion and run, the mass, the normalizer and the symmetric contraction (5), and the raw predictor row (6).
The library also proves that the Gaussian symbol is nonnegative✦, and that multiplication by a measurable function bounded by almost everywhere maps into itself with norm at most .