The Setting
A report can reveal the old block labels or reveal an independent fallback label on an event determined by the old labels. The comparison is between the retained variance of the total score and the Fourier degree of the report cells.
For a finite set and a real function , write for the uniform mean of , with value zero when is empty.✦ The fiber indicator of a report at a label is the indicator function of the fiber over ✦, and the fiber probability is its uniform mean, as in (1).✦ For a report , a real observable and a label , the restricted moment is the uniform mean of weighted by the indicator of the fiber over .✦ The fiber conditional mean of a real observable under a report , at a label , is the real number obtained by dividing the restricted moment by the fiber probability, with value zero on an empty fiber.✦ For a report and a real observable , the conditional expectation is the function in (1). (1)✦
Let be a finite block alphabet, a label set, and a finite leaf index set with and . The selector and leaf space is ✦; the full block space is , whose last coordinate is the fallback block.✦ For , the old score and full score are the sums in (2). For , the old label tuple applies to each selector and leaf block. (2)✦
Write for the old label tuple space.✦ For an event , the event-switched observation is the tagged report in (3); the tags and distinguish its two kinds of value. (3)✦
The quadratic deficit , for a real weight and , is the weighted mean in (4).✦ For a report and an event , the event weight is the function given by . (4)
A label score determines the switched sum predictor on tagged reports by (5), using the old sum on . (5)✦
Table Reports on the Sign Cube
The sign cube on a finite coordinate set is .✦ For a real function and , the Fourier character and real Fourier-Walsh coefficient are given in (6). For a real function , the degree is its ordinary real multilinear Fourier degree in (6), with degree zero for constants and the zero function. (6)✦
For a cube observation , its conditional cube sum, or score, is the conditional expectation of the coordinate sign sum .✦ For a cube observation , the retained variance is the second moment of its centered conditional score , as in (7). (7)✦
A selector table assigns a leaf to each selector label. For Boolean leaf tests , the unique failing leaf event is defined by (8). (8)✦
For observations , the selector and leaf tuple and the table-switched observation are the maps in (9). (9)
The switching block index set has a selector index and a fallback index in addition to the leaves.✦ The block coordinate identification sends to the blocks , where .✦ The switching block identification groups those blocks as an old tuple and a fallback block.✦ The switching cube identification is the map from to in (10), obtained by composing the block coordinate identification with the switching block identification✦; for a block observation , the table-switched cube observation uses in every slot. (10)✦
The Claim
Theorem 1. Let , , and satisfy and . For every , the prediction error satisfies (11). (11)✦ If also , every real satisfies (12). (12)✦ For arbitrary , and , the conditional sum identity is (13). (13)✦
Let be a nonnegative integer. If every fiber indicator of , and each has degree at most , every report value of the table-switched observation satisfies (14). (14)✦ For a block observation whose fiber indicators all have degree at most , the table-switched cube observation satisfies (15) for every report value . (15)✦ For every block observation , its switched score satisfies (16). (16)✦ If for and , its retained variance satisfies (17). (17)✦
In the paper, the result reads: for arbitrary independent centered variance-one original inputs with sum and an independent centered variance-one input , the tagged report reveals the original tuple off and only on , has cell cost at most , and, with , and , satisfies (18) for every real . (18)
What Is Checked
The switching identities are stated over uniform finite product alphabets with an event on the old label tuple; the degree bounds concern fiber indicators of observations on finite sign cubes, including the block observation obtained by regrouping the product coordinates through .