Wavelet-Packet Content for Positive Operators

Myung-Sin Song, James Tian · 2026

Read Wavelet-Packet Content for Positive Operators by Myung-Sin Song, James Tian (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14174.txt.

Introduction

Myung-Sin Song and James Tian attach a notion of content to wavelet packet trees in 2026. Wavelet packets split a signal into a tree of finer frequency bands. The team gives a simple way to attach content to the tree's nodes, given a positive operator. At a fixed packet depth, packet projections split the operator.

The split yields positive pieces of the original operator. This decomposition induces a boundary measure on the packet path space. Vector-dependent densities show how energy distributes across the tree. Such maps help analysts see where an operator's strength concentrates.

Catalog record

CREATORMyung-Sin Song, James Tian
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-14174

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Wavelet-Packet Content for Positive Operators Myung-Sin Song, James Tian Abstract . We give a simple way to attach “content” to the nodes of a wavelet packet tree when a positive operator is given. At a fixed packet depth, the packet projections split the operator into positive pieces, and this decomposition induces a boundary measure on the packet path space, together with vector-dependent densities that show how energy is distributed across the tree. We then study a sequential extraction procedure and two depth-fixed greedy rules for choosing packet blocks, one based on trace weights and one based on Hilbert-Schmidt weights. The main results are explicit geometric decay estimates for the remainder under these greedy removals. In the Hilbert-Schmidt case we also isolate a coherence quantity that measures how close the operator is to being block-diagonal in the packet partition. We close with a concrete patch-based denoising procedure for images, where packet blocks are selected by these content weights computed from an empirical second-moment operator; the construction ensures that both the approximants and the remainders stay positive at every step. 2020 Mathematics Subject Classification Primary 42C40; Secondary 47B10, 47A63, 41A65, 94A08 1. Introduction Wavelet packets organize information along a tree. At a fixed depth one has a finite packet partition of the ambient space into mutually orthogonal subspaces, and this is the basic idea behind many best-basis and subband-selection procedures in harmonic analysis and signal processing. In this paper we use the packet tree as a tool to study positive operators, with a simple principle: all approximations should remain inside the positive cone at every step. Fix a wavelet packet system \left\{W_{w}:w\in\mathcal{T}\right\} and the associated packet projections \left\{P_{w}\right\}_{w\in\mathcal{T}} satisfying the usual tree relations. Given a positive operator R\in B\left(H\right)_{+} , the most naive way to attach information to a node w is by compression P_{w}RP_{w} , or at the scalar level by weights like \text{tr}\left(P_{w}R\right) and \left\langle x,P_{w}Rx\right\rangle . Our starting point is that, for an operator-level theory which survives iteration, the natural object is instead the packet content operator C_{w}\left(R\right)=R^{1/2}P_{w}R^{1/2}\in B\left(H\right)_{+}. This choice has two immediate advantages. First, it is positivity-preserving by construction. Second, the packet content pieces telescope along the tree, and at each fixed depth n they form a positive decomposition R=\sum_{\left|w\right|=n}C_{w}\left(R\right), so that every partial sum is a positive approximation and the remaining tail is also positive. This is the structure that sits behind all of the extraction schemes studied here. A recurrent theme is that the interest of the content operators is not in a single trace or a single quadratic form. In particular, in simple diagonal models (such as the Shannon multiplier example, Section 2.2 ) the numbers \text{tr}\left(C_{w}\left(R\right)\right) coincide with the naive compression weights \text{tr}\left(P_{w}R\right) . This is the “scalar trap”: at the level of one statistic, the square-root conjugation may look invisible. What is genuinely new is what happens when one uses the packet content operators iteratively along extraction paths, so that operator-valued interactions between different packet blocks and different stages become visible. To organize these scalar shadows of the operator-valued decomposition, we attach to each positive trace-class R a canonical boundary measure on the packet path space. Theorem 3.1 shows that there is a unique finite Borel measure \mu_{R} on \partial\mathcal{T} whose cylinder masses are given by \mu_{R}\left(\left[w\right]\right)=\text{tr}\left(C_{w}\left(R\right)\right), and that for each x\in H there is a nonnegative density e_{x}\in L^{1}\left(\partial\mathcal{T},\mu_{R}\right) representing the quadratic data \left\langle x,C_{w}\left(R\right)x\right\rangle as integrals over cylinders. We use this view as a convenient tool for greedy rules: at a fixed depth, one can read the content weights as a finite mass distribution over the depth- n cylinders, and extraction becomes a deterministic rule that removes mass from the heaviest available cylinders. The first quantitative statement comes once we freeze a depth n and only allow packets at that level. In Section 4 we study a trace-class greedy extraction scheme which, at each step, removes the depth- n content block of maximal trace. Theorem 4.1 gives a geometric decay estimate for the trace of the remainder, with a rate depending only on the number of packets at depth n . Thus, at fixed depth, repeatedly removing the heaviest content block forces the remainder to drop at a controlled rate. In Section 5 we turn to a Hilbert-Schmidt (HS) analogue. This is where genuinely noncommutative behavior appears. This approach is sensitive to how packet blocks interact, not merely to their individual traces. We introduce a depth- n coherence parameter \Gamma_{n}\left(A\right) which measures how far a positive HS operator is from being block-diagonal with respect to the depth- n packet partition. The resulting HS-greedy scheme admits a decay estimate (Theorem 5.3 ), and the coherence parameter provides a natural way to quantify when the HS rule improves over the trace only picture. Finally, in Section 6 we include a simple application to patch-based image denoising which keeps the wavelet packet structure explicit. One forms an empirical second-moment operator from noisy patches, computes depth- n packet block weights, and denoises by projecting onto the selected packet blocks. The main reason this fits naturally with the rest of the paper is that the underlying approximation mechanism is the same: the packet content decomposition ensures that both the approximants and the remainders stay positive at every step, while the boundary measure language from Theorem 3.1 gives an interpretation of the block selection rule. A few remarks on motivation and related literature. Wavelet packets, best-basis selection, and packet-based compression have a well-developed mathematical literature; see for example the early packet and optimal-decomposition viewpoint [ Tor92 , CMQW93 , Wic94 , Tut95 ] , later structural work on packet libraries [ Sal11 ] , and recent developments where packet ideas are pushed toward graph-based and imaging-oriented constructions [ CLS21 , BS23 , SS22 , CN18 ] . For concrete diagonal and near-diagonal regimes, Shannon-type packet models remain a useful testing ground [ CW03 , YS03 ] , and multilevel-systems (which is also close in spirit to the present tree formulation) is emphasized in [ JS23 ] . Connections to Cuntz-algebra and generalized Walsh constructions, which often serve as algebraic models for wavelet and packet bases, can be found in [ DPS14 , DPS19 , DP13 ] . Related, though coming from a different direction, are recent works on operator-generated frame-like expansions (operator orbit frames), which also emphasize structured expansions driven by an underlying operator [ BW24 ] From the operator-theoretic side, the requirement that approximants and remainders remain positive is a familiar theme in matrix and operator inequalities, and it is closely tied to Schur complements and related shorting operations [ Bha07 , MMP07 , And71 , AT75 , BM88 , Gol86 , ACS06 , Jan05 ] . The same circle of ideas is intertwined with classical results on operator ranges and two-projection geometry [ FW71 , NRRR79 , Met97 , AD69 , BS10 ] . One way to read the present construction is that the packet tree provides a structured family of compressions, while the square-root conjugation packages those compressions into positive pieces that are stable under truncation and iteration, which is exactly what one wants if the extraction rule is to be analyzed quantitatively at the operator level. The greedy rules in Sections 4

Read the complete text