The Setting
Clipping two trace experiments on the same event controls the mass discarded from both laws and the likelihood score of the resulting pair. The finite source experiments here consist of paired side reports and recursively selected leaf laws.
A finite probability law on a finite set is a nonnegative mass function with .✦ For a real function on , an event and a finite probability law on , expectation, event probability and total variation distance are given in (1). (1) For a finite reference law on and a nonnegative density on with , the density law is the probability law with mass . Conditioning a law on an event of positive probability gives the law in (2). For an event and a density on , the common clipping event is the subset of given in (2). (2) The squared even moment norm of a real function under , with half-order , is the probability-weighted quantity in (3). (3)✦ A clipped oracle for laws , with parameters , is a pair of finite laws and a real score satisfying (4); denote the set of these triples by . (4)✦
Paired Reports and Source Windows
The product law and the coordinate product law are laws with the masses in (5). For a probability kernel , the joint law has the mass there; a mixture of laws with mixing law has mass . (5) A family of paired mechanisms on a finite side space consists of laws and weights with at every outcome.✦ For a paired mechanism with laws on and an orientation , the oriented pair law on is or .✦ The hidden pair mass , flagged pair mass , and paired report law are defined by (6); denotes an unreported orientation, with the side pair still returned. (6) The report space for mechanisms is ✦, and its joint report law is .✦ For a reported flag , the orientation posterior is the probability law on that is uniform for and is the point mass at otherwise.✦ For a report vector , the joint orientation posterior is .✦
For leaf laws on a finite set and a Boolean labelling of a finite set of leaves indexed by , the selected leaf law is . A rule from orientation vectors to leaf labellings defines the canonical source oracle , the joint law in (7). (7)✦
Let be natural window and local trial counts. All window slots and nesting indices are zero-based; write . The scheduled global trial counts , window leaf sets , trial sets , and collection sets are the recursive objects in (8). (8) The source leaf and trial counts are ✦ and ✦. A window orientation assignment at depth is a choice of local bits, global bits and one depth- assignment for each child; at depth zero the assignment is unique.✦ These assignments are identified with Boolean functions on ✦, and enumeration gives vectors in .✦ The recursive source leaf orientation propagates the integer template label to each child, using the local assignments and incoming shift for the node anchor and the global assignments for the child’s scheduled shift. With motif parameter , local centers , incoming shift zero and root label , denote its leaf vector by ; at depth zero its label is .✦
The Recursive Budgets
A collection has trial count : the two root collections have counts and , and a descendant collection keeps its count in the child window.✦ For paired mechanisms with common hidden weight , the hidden metadata tail and fitting tail are the quantities in (9). (9) For a natural depth and positive real parameters , the source sparsity count , centered moment budget and clipping cost are defined by (10), with . (10) The hidden binomial cap and hidden smoothing bound are the functions in (11). (11) For a natural branching bound , a finite ordered role shape at depth zero is unique; at depth it is an ordered list of at most depth- shapes, each assigned one of three roles. Let be the number of such shapes.✦ The interior label budget and binary tree budget are defined in (12), where . (12) For , the forced contribution budget is given by (13). (13)✦ For a smoothing bound and a natural degree cutoff , the sharp degree budget and sharp energy budget are the quantities in (14). (14)
For block parameter and depth , the warmup block count is .✦ The dimension of zero-based block , scale slack , scale exponent , block gain and total score exponent are defined in (15). (15) For problem size , the scale exponent product , backwards target scale , half moment order , and gap floor are the quantities in (16). Here is the ceiling truncated below at zero, and the even score exponent is . (16)
The Claim
Theorem 1. Let be a finite law and let be nonnegative densities of mean one under . Fix .
Suppose , , and . The common clipping loss satisfies (17). (17)✦ Suppose an event satisfies on , and put , . For , assume . The conditioned score satisfies (18). (18)✦
For the source construction, let all be positive and let the paired mechanisms have coordinates and common hidden weight , with . Let be leaf laws with , and , where . Assume is positive and even for ; for , assume is even, , and . Suppose also that , , for , and , with . There is a clipped source oracle as in (19), for every natural cutoff . (19)✦
For the level budget, let , , and let the real gap parameter satisfy . Fix a natural index and put . Suppose each of , and is at most , and . The level clipping budgets satisfy (20). (20)✦
In the paper, the result reads: for fixed global parameters, a level with score exponent , and error scale , preliminary laws with common side metadata and conditional nonnegative unit-integral densities relative to a common conditional reference, whose centered norms are at most for and whose difference has squared norm at most uniformly on a metadata event of probability at least , can both be conditioned on the joint event consisting of that metadata event and , changing each law by in total variation and yielding laws with uniformly over levels.
What Is Checked
The claims concern finite density experiments and canonical source oracles formed from paired mechanisms, recursive window assignments and selected product leaf laws. For finite laws and any real weight , the paired report masses sum to one. If and at every outcome, those masses are nonnegative.