---
title: 'Arrow-Chasing: Techniques and Applications'
url: https://www.emergentmind.com/topics/arrow-chasing
type: topic
---

# Arrow-Chasing: Techniques and Applications

to=arxiv_search 总代理联系 code omitted  北京赛车投注json omitted
Arrow-chasing is a polysemous technical term used in several research areas to denote either a property of structures constrained by arrows, a procedure that follows arrows through a system, or a proof technique that reduces global behavior to local arrow-based rules. In contemporary arXiv literature, the term and closely related notions appear in Ramsey theory through graph arrowing and Folkman numbers, in combinatorial game theory through legal edge orientations on trees, in communication complexity through pointer-chasing, in enumerative combinatorics through the arrow product and arrow patterns, in visual proofs on Pascal’s triangle, in welded-knot theory through Arrow calculus, in homological algebra through diagram-chasing in double complexes, and in autonomous pursuit planning where a chaser repeatedly follows a moving target [1207.3750] [2110.08738] [2508.19158] [1903.09454] [2607.04094] [2508.16388] [1703.04658] [1108.0958] [1904.03421].

## 1. Ramsey arrowing and Folkman-type graph constructions

In graph theory, arrow-chasing is most literally formalized by Ramsey arrowing notation. For a graph \(G\), the relation
\[
G \to (a_1,\dots,a_k)^e
\]
means that every edge \(k\)-coloring of \(G\) contains a monochromatic \(K_{a_i}\) in some color \(i\), and
\[
G \to (a_1,\dots,a_k)^v
\]
is defined analogously for vertex colorings. The edge Folkman number
\[
F_e(a_1,\dots,a_k;p)
\]
is the minimum order of a graph that arrows \((a_1,\dots,a_k)^e\) while containing no \(K_p\); the vertex version \(F_v(a_1,\dots,a_k;H)\) is defined similarly for \(H\)-free graphs [1207.3750] [2605.16542].

A classical instance is
\[
F_e(3,3;4),
\]
the smallest order of a \(K_4\)-free graph \(G\) such that every red/blue edge-coloring of \(G\) yields a monochromatic triangle. The formulation is equivalent to the Erdős–Hajnal question asking whether there exists a \(K_4\)-free graph that is not the union of two triangle-free graphs: if the edge set can be partitioned into two triangle-free spanning subgraphs, then the corresponding 2-coloring avoids monochromatic triangles, and conversely a triangle-free 2-coloring gives such a partition [1207.3750].

The paper "Use of MAX-CUT for Ramsey Arrowing of Triangles" reduces this triangle-arrowing problem to a cut problem on an auxiliary graph \(H_G\). Its vertex set is
\[
V(H_G)=E(G),
\]
and two vertices \(e_1,e_2\in V(H_G)\) are adjacent exactly when there exists an edge \(e_3\) such that \(\{e_1,e_2,e_3\}\) forms a triangle in \(G\). If \(t_\triangle(G)\) denotes the number of triangles of \(G\), then
\[
E(H_G)=3t_\triangle(G).
\]
The central theorem, due to Dudek and Rödl and used in the paper, is
\[
G\to(3,3)\quad\text{if and only if}\quad MC(H_G)<2t_\triangle(G),
\]
where \(MC(H_G)\) is the MAX-CUT value of \(H_G\) [1207.3750].

This reduction turns an arrowing statement into an optimization certificate. A 2-edge-coloring of \(G\) is a bipartition of \(V(H_G)\); each bichromatic triangle of \(G\) contributes exactly \(2\) crossing edges in the induced \(K_3\) of \(H_G\), whereas a monochromatic triangle contributes \(0\). Hence a cut of size \(2t_\triangle(G)\) corresponds exactly to a coloring with no monochromatic triangle, while any strict upper bound below \(2t_\triangle(G)\) certifies \(G\to(3,3)\) [1207.3750].

The paper first uses the spectral MAX-CUT bound
\[
MC(H_G)\le \frac{|E(H_G)|}{2}-\frac{\lambda_{\min}|V(H_G)|}{4},
\]
where \(\lambda_{\min}\) is the minimum adjacency eigenvalue of \(H_G\), and then strengthens it using the Goemans–Williamson semidefinite relaxation
\[
\text{Maximize}\quad \frac{1}{2}\sum_{i<j}w_{i,j}(1-\mathbf v_i\cdot \mathbf v_j)
\]
subject to \(\|\mathbf v_i\|=1\). The SDP optimum is used only as an upper bound on \(MC(H_G)\), not for randomized rounding, because certification of arrowing requires an upper certificate rather than a large cut [1207.3750].

Quantitatively, this program improves the best upper bound on \(F_e(3,3;4)\) from \(941\) to \(860\), and then to \(786\). The 860-vertex graph comes from an induced subgraph of \(G(941,5)\); the 786-vertex graph \(G_{786}\) is obtained from \(L(785,53)\) by adding one new vertex joined to a specified set of 60 vertices. For \(G_{786}\),
\[
|V(G_{786})|=786,\qquad |E(G_{786})|=61290,\qquad t_\triangle(G_{786})=428881,
\]
so \(2t_\triangle(G_{786})=857762\), while SDP computations give an upper bound \(MC(H_{G_{786}})\le 857753\), and an additional computation by Rinaldi using `SpeeDP` improves this to
\[
857742 \le MC(H_{G_{786}})\le 857750.
\]
This yields
\[
F_e(3,3;4)\le 786
\]
while the paper recalls the lower bound \(19\le F_e(3,3;4)\) [1207.3750].

Recent work on small Folkman graphs broadens this arrow-chasing landscape for parameters \(a_i\in\{2,3\}\) and \(k\le 4\). The paper "On Small Folkman Graphs Arrowing \(K_2\) or \(K_3\)" gives new exact values and bounds for both vertex and edge Folkman numbers under forbidden subgraphs such as \(J_n\), \(C_4\), and \(W_5\). Among the new exact values are
\[
F_v(2,2,3;J_5)=10,\qquad F_v(3,3;J_5)=11,\qquad F_v(2,3,3;J_6)=12,\qquad F_v(3,3,3;J_6)=15,
\]
and
\[
F_e(3,3;J_6)=11.
\]
The same paper proves the existence of \(F_e(3,3;W_5)\) by combining a \(C_4\)-free graph \(G\) with \(G\to(3,3)^v\) and the observation that adding one universal vertex converts vertex-arrowing \((3,3)^v\) into edge-arrowing \((3,3)^e\), while preserving \(W_5\)-freeness when \(G\) is \(C_4\)-free [2605.16542].

These results show that in Ramsey theory, arrow-chasing is not merely notation. It is a structural principle: an arrowing statement can be certified by optimization, symmetry-restricted search, extension arguments based on independent sets, SAT unsatisfiability, or by transferring vertex-arrowing to edge-arrowing through universal-vertex constructions [1207.3750] [2605.16542].

## 2. Games of arrows and impartial play on trees

In combinatorial game theory, "The Game of Arrows on 3-Legged Spider Graphs" defines the Game of Arrows as the tree specialization of Francis Su’s Game of Cycles. A move consists of orienting one previously unmarked edge of a graph, subject to the condition that the resulting partial orientation creates no internal sinks or sources; on trees the cycle-cell victory condition is irrelevant, so the game becomes a finite impartial normal-play game in which the last legal move wins [2110.08738].

For a graph \(G\), the set of possible directed edge-marks is
\[
A(G)=\{(v,w)\mid \{v,w\}\in E(G)\}.
\]
A decoration \(X\subseteq A(G)\) is a partial orientation with at most one direction chosen on each edge. A state is a decoration with no internal sinks or sources. If \(\mathcal F_G(X)\) denotes the set of legal followers of \(X\), then the Sprague–Grundy value is defined by
\[
g(X)=\mathrm{mex}\, g(\mathcal F(X)).
\]
For disjoint unions,
\[
g(X\sqcup Y)=g(X)\oplus g(Y).
\]
These are the computational foundations of the paper’s arrow-chasing analysis [2110.08738].

A central methodological innovation is state isomorphism. A state isomorphism from a state \(X\) of \(G\) to a state \(Y\) of \(H\) is a bijection on directed marks of the unmarked subgraphs that commutes with flip and induces a bijection between all descendants of \(X\) and all descendants of \(Y\). This is stronger than matching immediate followers. The local verification mechanism is given by Theorem 4.8: a bijection commuting with flip is a state isomorphism if and only if the induced maps are locally state-preventing at every vertex [2110.08738].

The paper’s main result concerns 3-legged spider graphs \(S(a,b,c)\). After introducing trimming, which removes leaf obstructions and reduces each positive leg length by \(1\), the technical theorem is:
\[
g(S[a,b,c])=0 \qquad \text{for even positive } a,b,c.
\]
Consequently, for the original Game of Arrows, if \(a,b,c\) are odd and nonzero, then the initial state on \(S(a,b,c)\) is a \(P\)-position, so player two has a winning strategy [2110.08738].

The proof is built by decomposing spider states into path-like subgames. The atomic path states have explicit Grundy values:
\[
g(F_n)=\#_1 n,\qquad g(C_n)=\#_1 n\oplus 1,\qquad g(T_n)=n,\qquad g(R_n)=\#_1 n,
\]
where \(\#_1 n=n\bmod 2\). More elaborate spider states are then reduced to disjoint unions of these pieces. A representative decomposition is
\[
S[a(m),b,c]\cong T_{a-m}\sqcup S[\widetilde{m-1},b,c].
\]
The final parity argument shows that every follower of an empty even-legged spider has odd Grundy value, so the empty spider itself has Grundy value \(0\) [2110.08738].

Here arrow-chasing means chasing legal arrow placements through the game tree, but the paper’s deeper contribution is the replacement of global symmetry arguments by a local equivalence calculus on partially oriented states. This suggests a broader interpretation: in impartial graph-orientation games, arrow-chasing is the process of resolving a complex orientation state into smaller canonical arrow-configurations whose Grundy values are computable [2110.08738].

## 3. Pointer chasing and communication complexity

In communication complexity, arrow-chasing appears as pointer chasing: Alice and Bob hold alternating parts of a functional directed graph, and the task is to determine the endpoint after \(k\) alternations. In the formulation of "Pointer Chasing with Unlimited Interaction", Alice holds \(N_A:[n]\to[n]\), Bob holds \(N_B:[n]\to[n]\), and
\[
p_0=1,\qquad p_1=N_A(p_0),\qquad p_2=N_B(p_1),\qquad p_3=N_A(p_2),\ \ldots
\]
More generally,
\[
p_t=
\begin{cases}
N_A(p_{t-1}) & t \text{ odd},\\
N_B(p_{t-1}) & t \text{ even}.
\end{cases}
\]
The \(k\)-step problem can be written as
\[
PC_f^{n,k}(N_A,N_B)=f(\mathrm{walk}_A^k(1)),
\]
where \(f:[n]\to\{0,1\}\) is fixed and nontrivial [2508.19158].

There is a trivial \(k\)-round protocol: send the current pointer value at each step, for total communication \(O(k\log n)\). The paper’s main result is that this protocol remains nearly tight even when the number of rounds is unrestricted. For constant error,
\[
R_\epsilon(PC_f^{n,k})=\Omega\!\big(k\log(n/k)\big),
\]
and for zero error, under \(4k\le n\),
\[
R_0(PC_f^{n,k})=\Omega\!\big(k\log\log k\big).
\]
The first bound shows that if \(k=O(n^\delta)\) for constant \(0<\delta<1\), then the trivial \(O(k\log n)\) protocol is optimal up to constants [2508.19158].

The proofs are reduction-based. The bounded-error lower bound embeds the communication problem
\[
OR_{r,m}(x,y)=\bigvee_{i=1}^m (y_i)_{x_i}
\]
into a pointer-chasing instance by arranging \(m\) gadgets so that the alternating walk checks one queried bit per gadget. The zero-error lower bound reduces from \(Cycle_n\), using the fact that whether a \(2k'\)-step alternating walk returns to the starting point distinguishes Hamiltonicity in a union of perfect matchings when \(k'\) is prime [2508.19158].

A second 2025 paper gives a simple deterministic lower bound using a density-invariant round-elimination argument. For the standard \(k\)-pointer-chasing function on arrays \(a,b\in[n]^n\), with Alice seeing \(b\), Bob seeing \(a\), deterministic \(k\)-round communication is at least \(n/8\). More strongly, the theorem is formulated for dense restricted domains \(A,B\subseteq[n]^n\) with partially fixed coordinates and an “alive” starting pointer condition
\[
\Pr_{a\in A}[a[0]=v] < \frac{2}{n}\qquad\forall v\in[n].
\]
The key invariant is the unfixed density
\[
\frac{|A|}{n^{\,n-|F_A|}}
\]
and likewise for \(B\); fixing a heavy coordinate doubles this density. Repeating the process yields a clean round-elimination contradiction [2507.08919].

The communication-complexity literature therefore uses arrow-chasing in its most literal algorithmic sense: one must follow a sequence of dependent arrows through distributed information. The surprising point is that unrestricted interaction does not substantially compress this dependence [2508.19158] [2507.08919].

## 4. Symbolic, visual, and enumerative arrow methods

A different family of meanings arises when arrows encode local combinatorial structure rather than a path to be followed.

In "Symbolic method and directed graph enumeration", the arrow product is introduced for families \(\mathcal A\) and \(\mathcal B\) of labeled digraphs. The arrow product consists of pairs \((a,b)\), with \(a\in\mathcal A\), \(b\in\mathcal B\), relabeled to have disjoint labels, together with an arbitrary set of directed edges from vertices of \(a\) to vertices of \(b\). The key generating-function rule is:
\[
\mathbf C(z,w)=\mathbf A(z,w)\mathbf B(z,w),
\]
where \(\mathbf A,\mathbf B,\mathbf C\) are graphical generating functions. This makes the arrow product a symbolic device for assembling digraphs from strongly connected components and acyclic inter-component wiring [1903.09454].

The same paper derives compact formulas for classical digraph families. For directed acyclic graphs,
\[
\mathbf{DAG}(z,w,u)=\frac{\mathbf{Set}((u-1)z,w)}{\mathbf{Set}(-z,w)},
\]
where \(u\) marks sources. For strongly connected digraphs,
\[
\mathrm{SCC}(z,w)= -\log\!\left(G(z,w)\odot \frac{1}{G(z,w)}\right).
\]
More generally, if \(A(z,w)\) is the EGF of a family of allowed strongly connected components, then the GGF of digraphs whose SCCs all lie in \(A\) is
\[
\mathbf D_A(z,w)=\frac{1}{\mathbf{Set}(z,w)\odot e^{-A(z,w)}}.
\]
Here arrow-chasing is not path-following but one-way dependency assembly: the arrow product formalizes how arrows may run from one component block to another [1903.09454].

In permutation theory, "When arrow patterns meet classical patterns" studies arrow patterns, introduced to connect one-line notation and cycle notation. An arrow pattern \(\alpha=(\nu;H)\) consists of a word \(\nu\) and arrows \(b_i\to c_i\); a permutation \(\sigma\) contains \(\alpha\) when a pattern occurrence in \(\sigma\) is compatible with specified map relations in \(\hat\sigma=\Phi^{-1}(\sigma)\), where \(\Phi\) erases parentheses in standard cycle notation. The paper completely enumerates the six classes
\[
S_n(\pi,\alpha),\qquad \pi\in S_3,
\]
for
\[
\alpha=(12;1\to 3),
\]
obtaining
\[
a_n(123,\alpha)=2^n-n,
\]
\[
a_n(132,\alpha)=a_n(312,\alpha)=C_n,
\]
and
\[
a_n(213,\alpha)=a_n(231,\alpha)=a_n(321,\alpha)=F_{2n-1}.
\]
For the distinct arrow pattern
\[
\beta=(12;1\to 2),
\]
the paper proves
\[
a_n(321,\beta)=M_n
\]
by two independent bijective arguments, one through non-nesting involutions and one through Dyck paths [2607.04094].

In "Arrow-chasing in Pascal's triangle -- Visual proofs for summation formulas involving binomial coefficients", arrow-chasing becomes an explicitly visual proof technique based only on Pascal’s rule
\[
\binom{n}{k}=\binom{n-1}{k-1}+\binom{n-1}{k},
\qquad
\binom{n}{0}=\binom{n}{n}=1,
\]
with \(\binom{n}{k}=0\) outside \(0\le k\le n\). Two complementary modes are emphasized. Downward arrow-chasing uses Pascal’s rule directly, while upward arrow-chasing uses the weight rule: for a weighted row sum
\[
\sum_{k=0}^n a_k\binom{n}{k},
\]
moving one row upward preserves the total but replaces the weights by
\[
a_k+a_{k+1}.
\]
This yields visual proofs of identities such as
\[
\sum_{k=0}^n \binom{n}{k}=2^n,
\qquad
\sum_{k=m}^n \binom{k}{m}=\binom{n+1}{m+1},
\]
\[
\sum_{k=0}^n k\binom{n}{k}=n\,2^{n-1},
\qquad
\sum_{k=0}^n (-1)^k\binom{n}{k}=0,
\]
\[
\sum_{k=0}^n \binom{n}{k}^2=\binom{2n}{n},
\qquad
(a+b)^n=\sum_{k=0}^n \binom{n}{k}a^{n-k}b^k.
\]
The paper also develops a subtractive variant based on
\[
\binom{n}{k}=\binom{n+1}{k+1}-\binom{n}{k+1}
\]
for certain alternating identities [2508.16388].

These works collectively show that in symbolic and enumerative combinatorics, arrow-chasing denotes local rule propagation: arrows organize how weights, dependencies, or structural constraints move through a combinatorial object [1903.09454] [2607.04094] [2508.16388].

## 5. Topological and homological arrow calculi

Arrow-chasing also names local rewriting systems in topology and homological algebra.

In "Arrow calculus for welded and classical links", Arrow calculus encodes a welded or classical link diagram by a crossingless base diagram together with immersed oriented arrows or higher-order \(w\)-trees. A \(w\)-tree is a connected uni-trivalent tree immersed in the plane of the diagram, with oriented edges, virtual crossings only, and optional twist decorations. A \(w_k\)-tree is a \(w\)-tree with \(k\) tails. An Arrow presentation is a pair \((V,A)\), where \(V\) has no classical crossings and \(A\) is a set of \(w\)-arrows such that surgery on \(V\) along \(A\) yields the original diagram [1703.04658].

The calculus is generated by six basic Arrow moves: virtual isotopy, head/tail reversal, tails exchange, isolated arrow, inverse, and slide. The key completeness theorem states that two Arrow presentations represent equivalent diagrams if and only if they are related by Arrow moves. The theory extends from arrows to \(w\)-tree presentations through the expansion move, which recursively resolves a \(w\)-tree into \(w\)-arrows. The \(w\)-tree version includes higher-order analogues of inverse, slide, head traversal, heads exchange, head–tail exchange, antisymmetry, and the fork move [1703.04658].

An algebraic shadow accompanies the diagrammatics. At a trivalent vertex, incoming labels \(a,b\) induce the commutator label
\[
[a,b]:=a\bar b \bar a b.
\]
Exchange lemmas appear as identities such as
\[
AB = B[\bar B,\bar A]A,
\]
while the fork relation is
\[
[[A,A]]=1.
\]
These formulas make explicit that moving heads and tails past one another introduces higher-degree commutator corrections. In this sense, arrow-chasing is literally the local manipulation of arrow data through a diagram, controlled by exact rewrite rules and graded error terms [1703.04658].

In homological algebra, "On diagram-chasing in double complexes" gives a different but related local calculus. For an object \(A\) in a double complex, besides the horizontal and vertical homology objects, Bergman defines two additional subquotients:
\[
{}^A=(\Ker e\cap \Ker f)/\Im p,
\qquad
A_=\Ker q/(\Im c+\Im d),
\]
where \(p\) and \(q\) are the diagonal composites into and out of \(A\). These are called the receptor and donor. There are canonical intramural maps among \({}^A\), the horizontal and vertical homology at \(A\), and \(A_\), and every arrow \(A\to B\) induces an extramural map
\[
A_ \longrightarrow {}^B.
\]
The main local statement, the Salamander Lemma, gives a 6-term exact sequence attached to each arrow; for a horizontal arrow this has the form
\[
C_ \longrightarrow {}^A \longrightarrow H_h(A) \longrightarrow A_ \longrightarrow {}^B \longrightarrow H_h(B) \longrightarrow {}^D,
\]
with the vertical analogue for vertical arrows [1108.0958].

Classical “long-distance” diagram chases are then reconstructed as concatenations of these local salamander sequences. The Snake Lemma, the Sharp \(3\times 3\) Lemma, and long exact homology sequences become composites of donor–receptor identifications and homology maps. This makes explicit a viewpoint that is often tacit: diagram-chasing is a local transport process through named subquotients, not a mysterious jump between distant corners of a diagram [1108.0958].

A plausible implication is that both papers—despite living in different fields—treat arrow-chasing as a local-to-global technology. In one case arrows are geometric surgery data; in the other they are maps in an abelian category. In both cases global invariants are recovered by systematic motion through a network of local arrow-attached objects [1703.04658] [1108.0958].

## 6. Pursuit, robotics, and systems interfaces

In applied settings, arrow-chasing often returns to literal pursuit.

"Investigating the classical problem of pursuit, in two modes" studies pursuit on a straight line and curved pursuit. In the rotational two-particle model, a fox and a rabbit move with equal speed \(v\), and the rabbit’s velocity makes a constant angle \(\alpha\) with the line joining rabbit and fox. The paper derives coupled equations for the velocity components,
\[
\frac{dv_x}{dt}=\frac{v\sin\alpha}{L-v(1-\cos\alpha)t}\,v_y,
\qquad
\frac{dv_y}{dt}=-\frac{v\sin\alpha}{L-v(1-\cos\alpha)t}\,v_x,
\]
and identifies a logarithmic-spiral law in the symmetric \(N\)-particle cyclic pursuit setting:
\[
r(\theta)=r_0 e^{-\theta/\cot(\alpha/2)}.
\]
For a regular \(N\)-gon of side length \(l\), the radius evolves as
\[
r(t)=\frac{l}{2\sin(\alpha/2)}-v\sin\frac{\alpha}{2}\,t.
\]
The same spiral relation is proposed for moths circling a lamp under a constant-angle navigation rule [2309.02471].

In autonomous robotics, "Online Trajectory Generation of a MAV for Chasing a Moving Target in 3D Dense Environments" treats target chasing as a real-time planning problem for a multirotor equipped with a vision sensor. The environment is known through an occupancy map, and the target path is previewed over a short horizon
\[
x_p(\tau),\qquad t\le \tau\le t+H.
\]
The planner’s central visibility quantity is
\[
\psi(x;x_p)=\min_{y\in L(x,x_p)} \phi(y),
\]
where \(\phi\) is the Euclidean distance field of obstacles and \(L(x,x_p)\) is the line segment from chaser to target. Thus \(\psi>0\) means the target is visible, and larger \(\psi\) means more line-of-sight clearance [1904.03421].

The planning architecture is cascaded. A graph-search preplanner first computes visibility-aware waypoints \(\sigma_{[t_0,t_N]}\) and safe chasing corridors by minimizing a sum of interval distance, visibility cost, and tracking-distance error. The transitional visibility term is
\[
c_v(x_{n-1},x_n)=
\Bigg(
\sqrt{
\left(\int_{L(x_{n-1},x_n)} \psi(y;x_{p,n-1})\,dy\right)
\left(\int_{L(x_{n-1},x_n)} \psi(y;x_{p,n})\,dy\right)
}
\Bigg)^{-1}.
\]
A second stage then solves a convex QP for a piecewise-polynomial minimum-jerk trajectory inside the chasing corridors, minimizing
\[
\sum_{n=1}^{N} \left( \int_{t_{n-1}}^{t_n}\|\dddot{x}_c(\tau)\|^2\,d\tau + \lambda \|x_c(t_n)-x_n\|^2 \right)
\]
subject to initial conditions, \(C^2\) continuity, and corridor membership [1904.03421].

The method is evaluated in the “mini garden” and “complex city” environments. With horizon \(H=5\) s, average total runtime is \(0.11\) s in the mini garden and \(0.08\) s in the complex city, corresponding to about \(10\) Hz replanning. Increasing the visibility weight \(w_v\) improves average visibility and reduces occlusion duration, at the cost of longer travel distance [1904.03421].

A different systems use of “Arrow” appears in data processing rather than pursuit. "Zero-Cost, Arrow-Enabled Data Interface for Apache Spark" introduces Arrow-Spark, a Spark `FileFormat` integration that uses the Apache Arrow Dataset API as a data layer. The system routes data through Arrow IPC messages and JNI into Spark’s `ColumnarBatch` abstraction. The paper’s empirical claim is that consuming data through Apache Arrow is “zero-cost” in the sense that performance is on par with or better than native Spark for the tested workloads. In particular, for CSV ingestion the implementation is reported to be approximately \(4.5\times\) faster than Spark, while the batch-size study identifies \(8192\) rows—corresponding to \(256\) KB for rows of four 64-bit integers—as the best-performing size on the evaluated hardware [2106.13020].

This systems usage is semantically distinct from Ramsey arrowing or pointer chasing, but it still fits the broad encyclopedia pattern: arrow-chasing names technical workflows in which arrows or arrow-like interfaces mediate the transport of information, visibility, or structure across a constrained architecture [1904.03421] [2106.13020].

Source: https://www.emergentmind.com/topics/arrow-chasing