Unmasked Self-Attention Mechanism

• Unmasked self-attention is a technique that removes traditional masking to allow every token to attend to all others, capturing richer contextual relationships.
• It improves interpretability by exposing the complete inter-token interactions during both training and inference phases of transformer architectures.
• Its implementation benefits diverse applications in NLP and computer vision, leading to enhanced model performance and deeper insights into attention dynamics.

A binary tree ordering is any systematic process or rule by which the nodes, subtrees, or entire structures within a class of binary trees are arranged according to a specific criterion. The concept encompasses multiple formalizations: (1) traverse-based orderings of the set of nodes (e.g., preorder, postorder, inorder); (2) partial orders or posets induced by tree structure or operations (e.g., the ancestor-descendant relation, rotation posets); (3) enumerative bijections between binary trees and other discrete sets (e.g., natural numbers, permutations); (4) succinct encodings that preserve subtree or traversal order and permit order-number queries. Binary tree ordering is critical for efficient algorithmic processing, succinct data representation, and the analysis of structural and probabilistic properties relevant in combinatorics, theoretical computer science, and computational phylogeny.

1. Traversal-Based Orderings

Standard binary tree traversals induce explicit orderings of the nodes:

• Preorder: Root-first traversal records the sequence $(r,\ \text{Preorder}(L),\ \text{Preorder}(R))$, where $L$, $R$ are left/right subtrees. Every preorder traversal of a BST with $n$ nodes is a permutation of $n$ avoiding pattern $(2,3,1)$; the set of preorder traversals of all $n$-node BSTs coincides with the $(2,3,1)$-avoiding permutations (Levy et al., 2019).
• Postorder: Leaves are visited before root: $(\text{Postorder}(L),\ \text{Postorder}(R),\ r)$. These orderings correspond to $(3,1,2)$-avoiding permutations (Levy et al., 2019).
• Inorder: Left subtree, root, right subtree; orders the keys in increasing order for BSTs and underpins range-minimum query (RMQ) applications (Davoodi et al., 2014).

Traversal orderings are foundational in runtime algorithms (e.g., splaying, sorting) and data encodings (e.g., Zaks’ sequence for BP representation).

2. Partial Orders and Sorting in Tree Posets

The structure of a binary tree naturally induces a partial order (poset) via the transitive, antisymmetric parent-of relation. Sorting in this tree-poset—not merely linearizations—is a central combinatorial and algorithmic problem:

• Tree-poset model: Given tree $L$0 with $L$1 nodes, $L$2 induces the partial order; $L$3 is the maximal outdegree, $L$4 the width (antichain size, i.e., number of leaves) (Roychoudhury et al., 2022).
• Sorting complexity:
  • Randomized: $L$5 time/queries, optimal for bounded-degree trees such as binary trees ($L$6).
  • Deterministic: $L$7; for complete binary trees ($L$8), this is $L$9.
• Algorithmic techniques: Recursive centroid (pseudo-centroid) finding, sampling-based subtree size estimates, chain decomposition for deterministic algorithms (Roychoudhury et al., 2022).
• Lower bound: Any tree-poset sorter must make $R$0 comparisons; this matches the randomized upper bound in strongly bounded-degree regimes.
• DFS-Numbering: For practical use, a final DFS numbering after sorting supports constant-time comparison queries.

These sorting and ordering procedures are crucial for problems such as reconstructing tree structure given only a comparison oracle, comparing hierarchical clusterings, or supporting efficient range queries.

3. Orderings of Tree Classes: Enumerations and Bijections

Binary tree structures—ordered, unordered, ranked—have distinct enumeration and ordering schemes:

Tree Class	Size Parameter	Enumeration Formula	Canonical Ordering
Ordered (plane)	$R$1 leaves	$R$2 (Catalan)	Implicit plane order (left/right)
Ordered, ranked (permutation trees)	$R$3 leaves (or $R$4 internal nodes)	$R$5	Bijective with permutations of $R$6 (rankings)
Unordered (nonplane)	$R$7 leaves	No closed form	No canonical left/right assignment

• Generating functions: Ordinary and exponential GFs encode counts and probabilities of ordered trees, as in $R$8 for plane trees and $R$9 for permutation trees (Wiehe, 2020).

Enumerative orderings are critical in probabilistic models (e.g., Yule process) and enable uniform sampling and indexing.

4. Succinct Encodings, Numberings, and Order-Preserving Representations

Succinct binary tree representations rely on order-preserving encodings:

• Ordinal-binary transformations: Four symmetric bijections $n$0 map an $n$1-node binary tree + dummy root to an ordinal tree with $n$2 nodes, preserving critical orderings (preorder, postorder, inorder, etc.) (Davoodi et al., 2014).
• Bit-complexity: All variants can be stored in $n$3 bits, supporting navigation and ordering queries in $n$4 time.
• Zaks’ sequence: Encodes a binary tree as a length $n$5 BP (balanced-parentheses) string, with direct support for inorder and postorder numbering (Davoodi et al., 2014).
• Order-preserving properties: These representations support constant-time queries to parent, child, next/previous sibling, depth, subtree size, LCA, and any of the (pre-, in-, post-) orderings.

Such encodings are essential for range queries (e.g., RMQ via Cartesian tree), space-efficient text indexing, and supporting algorithmic operations on massive trees.

5. Orderings in Stochastic and Evolutionary Models

The process by which binary trees are generated—especially in evolutionary models—imposes distinct orderings on the space of trees:

• Yule (pure-birth) model: Sequentially splits extant leaves, generating labeled permutation trees (ordered, ranked) with uniform probability $n$6 (Wiehe, 2020).
• Plane shape vs. permutation labeling: The ordered plane shape is counted by Catalan numbers; when combined with rankings, the class matches permutations.
• Grafting construction: Each new leaf is attached by uniformly choosing among $n$7 existing segments; this process yields the uniform law on ordered ranked trees.
• Structural statistics: Node balance, number of “cherries” (two-leaf subtrees), and higher-order caterpillars can be tracked by bivariate or multivariate generating functions, revealing distributional properties specific to ordering (e.g., mean number of cherries $n$8 in Yule trees).

Orderings induced by these stochastic models drive inference in phylogenetics and population genetics.

6. One-to-One Rank/Unrank Bijections and Enumeration

A canonical bijection exists between $n$9 and the set of full binary trees (each internal node has exactly two children, each leaf has none):

• Type and order: For $n$0 with prime factorization $n$1, define $n$2; the $n$3-th element among numbers of given type $n$4 is denoted $n$5.
• Unranking map $n$6 (index $n$7 tree):

$n$8

• Ranking map $n$9 (tree $(2,3,1)$0 index):

$(2,3,1)$1

This pairing yields a one-to-one correspondence between $(2,3,1)$2 and the set of full binary trees (Skliar et al., 2020).

• Enumeration: The number of full binary trees with $(2,3,1)$3 internal nodes is $(2,3,1)$4 (Catalan number).
• Algorithmic process: Factorization and counting in the “type” sequence implement both ranking and unranking. Examples verify the mapping and its inversion.

These bijections enable lexicographic ordering, efficient encoding, and exhaustive generation of all binary trees up to a given size.

7. Pattern-Avoiding Orderings and Dynamic Algorithms

Certain permutation classes—those avoiding small patterns—are intimately linked to binary tree traversals:

• Pattern-avoiding: $(2,3,1)$5 is a preorder if and only if it avoids $(2,3,1)$6; postorder if and only if it avoids $(2,3,1)$7 (Levy et al., 2019).
• Dynamic splaying: Inserting (by splaying) the preorder or postorder of any BST into an empty BST has linear total cost ($(2,3,1)$8) due to the structure imposed by pattern-avoidance. The analysis employs a simple potential function tracking touched and untouched nodes and their contributions to subtree accesses.
• Dynamic Finger property: For arbitrary starting trees, as long as the destination BST is weight-balanced, the total cost to bring it to a specific ordering (either preorder or postorder) remains linear in $(2,3,1)$9, bound via the Dynamic Finger Theorem (Levy et al., 2019).

These results exemplify how orderings with special structure—arising from traversal patterns or partial orders—enable efficient dynamic access and sorting algorithms, providing evidence towards conjectures on dynamic optimality of splay trees.

---

Concluding remarks:

The study of binary tree orderings encompasses a diverse spectrum: strict traversal-induced permutations, partial orders governing ancestor/descendant relationships, global enumerative bijections with other mathematical objects, stochastic orderings from evolutionary processes, and sublinear/constant-time-supporting succinct encodings. Advanced algorithms for sorting tree-posets and encoding trees exploit these structures, yielding both theoretical optimality and practical utility in data indexing, algorithms, and combinatorial enumeration (Roychoudhury et al., 2022, Wiehe, 2020, Davoodi et al., 2014, Levy et al., 2019, Skliar et al., 2020).