---
title: 'Jumping Connections: Bypassing Locality in Systems'
url: https://www.emergentmind.com/topics/jumping-connections
type: topic
---

# Jumping Connections: Bypassing Locality in Systems

Searching arXiv for recent papers on "jumping connections" to ground the article.
{"query":"all:\"jumping connections\" OR ti:\"jumping connections\"","max_results":10,"sort_by":"submittedDate","sort_order":"descending"}
Searching arXiv for "jumping connections".
“Jumping connections” is not a single standardized construction. In the literatures surveyed here, the term and closely related notions denote mechanisms that introduce nonlocal transfer, discontinuous update, or stratified change into systems whose default behavior is local or layer-by-layer. This includes skip pathways in graph neural networks, fixed-distance jumps in discrete-time quantum walks, bundle-type changes on moduli spaces of parabolic connections, jump maps in hybrid consensus and wireless-sensor routing, and pointer-jumping compositions in communication complexity [2507.05533], [1106.2036], [1611.00971], [2110.14277], [0802.2843], [1004.2322].

## 1. Taxonomy of the term

A concise way to organize the subject is by the mathematical object on which the “jump” acts. In some settings it is an explicit operator on states; in others it is a structural shortcut between layers; in still others it is a change of stratum or filtration index.

| Domain | Meaning of “jumping connection” or cognate notion |
|---|---|
| Graph neural networks | A skip-connection from an intermediate hidden layer directly to the output |
| Quantum walks | A fixed-distance permutation that supplements the nearest-neighbor step |
| Moduli of parabolic connections | A jumping locus where the underlying bundle type changes |
| Hybrid consensus | A time-driven jump map interleaved with continuous flow |
| Wireless sensor routing | A jumping transmission mode activated under faults, congestion, or void regions |
| Communication complexity | Pointer-jumping through layered function composition |

A plausible unifying description is that jumping connections relax strict locality. In the GCN setting, they bypass depth-induced attenuation; in the quantum-walk setting, they bypass nearest-neighbor transport; in moduli theory, they connect generic and exceptional bundle types; in hybrid systems, they interleave continuous and discrete propagators; and in pointer jumping, they encode traversal through compositions of maps rather than through explicit sequential revelation.

## 2. Skip pathways in graph neural networks

In the GCN literature represented here, a jumping connection is a skip-connection from the first hidden layer directly to the final output of a two-hidden-layer network. With node features $X\in\mathbb R^{d\times N}$, normalized adjacency $A\in\mathbb R^{N\times N}$, hidden weights $W\in\mathbb R^{m\times d}$ and $U\in\mathbb R^{m\times m}$, output weight $C\in\mathbb R^{k\times m}$, and entry-wise ReLU $\sigma$, the network is
$$
\mathrm{out}(X,A;W,U)=C\,\sigma(WXA)+C\,\sigma(U\,\sigma(WXA)\,A).
$$
The corresponding idealized model replaces $A$ by a sparse effective adjacency $A^*$ that retains only the “essential” edges for propagation [2507.05533].

The central theoretical result is a joint learning-dynamics and generalization analysis for GCNs with both jumping connections and layer-wise graph sparsification. The analysis states that the learned model’s generalization accuracy closely approximates the highest achievable accuracy within a broad class of target functions dependent on $A^*$. The same analysis also identifies an asymmetry between layers: in a two-hidden-layer GCN, generalization is more affected by deviations of the first-layer sparsified matrix from $A^*$ than by deviations in the second layer. The paper formulates this asymmetry quantitatively through layer-specific conditions such as
$$
\|A^{1t}-A^*\|_1 \lesssim \epsilon_0/C_s(\mathcal F),\qquad
\|A^{2t}-A^*\|_1 \lesssim \epsilon_0/(\alpha\,C_s(\mathcal F)\,C_s(\mathcal G)).
$$
Because $\alpha\,C_s(\mathcal G)<1$, the second layer can be more aggressively sparsified [2507.05533].

The mechanism is explicit in the node-wise output representation. Writing $U=VC$ and denoting the first-layer skip output by
$$
\mathrm{out}^1(X,A;W)\coloneqq C\,\sigma(WXA),
$$
the $n$-th node output is
$$
\mathrm{out}_n(X,A;W,V)=\mathrm{out}_n^1(X,A;W)+C\,\sigma\!\bigl(V\,\mathrm{out}^1(X,A;W)\,a_n\bigr),
$$
where $a_n$ is the $n$-th column of $A$. The first term reaches the output without passing through the second hidden layer. This is why first-layer adjacency errors propagate directly, whereas second-layer errors are filtered through the composite term. The paper’s practical guideline is therefore conservative sparsification in the first few hops and more aggressive sparsification in deeper hops, and it states that the same principle extends to deep GCNs with many skip-connections, including Jumping Knowledge networks [2507.05533].

## 3. Fixed-distance jumps in discrete-time quantum walks

In one-dimensional discrete-time quantum walks, jumps are introduced by enlarging the usual walk dynamics on
$$
H=H_S\otimes H_C,\qquad
H_S=\mathrm{span}\{|n\rangle:n\in\mathbb Z\},\qquad
H_C=\mathrm{span}\{|L\rangle,|R\rangle\}.
$$
The clean walk uses a Hadamard coin
$$
C=I_S\otimes M,\qquad
M=\frac{1}{\sqrt 2}\begin{bmatrix}1&1\\1&-1\end{bmatrix},
$$
followed by the nearest-neighbor shift
$$
S=\sum_{n\in\mathbb Z}\bigl(|n+1,R\rangle\langle n,L|+|n-1,L\rangle\langle n,R|\bigr),
$$
so that one clean step is $U^{QW}=SC$. The jump extension adds, at each timestep, a permutation $S^{\mathrm{jump}}_\omega(t)$ built from disjoint transpositions of pairs of channels $\{n,n+j\!\!\mod N\}$, where $j$ is a fixed jump distance. One step of the disordered walk is
$$
U^{\mathrm{jump}}_\omega(t)=S^{\mathrm{jump}}_\omega(t)\,(SC).
$$
Two disorder models are distinguished: static disorder, where one permutation is frozen for all times, and dynamic disorder, where the permutation is sampled independently at each timestep [1106.2036].

The asymptotic distributions differ sharply between these regimes. Dynamic disorder yields a Gaussian-like position distribution with linearly growing variance, $\sigma^2(t)\sim v(p,j)t$. Numerically, for fixed $j$, the coefficient satisfies $v(p,j)\propto pj$ to leading order, so standard-deviation curves versus $p$ are essentially straight lines. Static disorder instead produces localization and parity-sensitive non-Gaussian profiles. For odd jump size $j$, the long-time diagonal probability approaches a stationary profile well-fitted by a discrete Laplace envelope modulated by secondary peaks of period $j$. For even $j$, the long-time profile develops a three-peak structure: one sharp, localized central peak at $n=0$ and two symmetrically displaced broad peaks around $n\approx \pm j/2$, with small oscillations of period $j$ superimposed [1106.2036].

A particularly notable numerical result is the universal scaling law for the long-time variance in the static odd-$j$ case. If
$$
\mathrm{Var}_{p,j}=\sum_n n^2 P_{p,j}(n),
$$
and one rescales
$$
x=p\,j^\alpha,\qquad y=j^{-\beta}\mathrm{Var}_{p,j},
$$
with empirically determined exponents $\alpha\simeq 1.04$ and $\beta\simeq 1.67$, data for various $j$ collapse onto a single U-shaped master curve $y=F(x)$, minimal around $x\approx 2$. A plausible implication is that the localization width is governed not by $p$ and $j$ separately, but by a scaled average jump parameter [1106.2036].

## 4. Jumping loci in moduli theory and jumping indices in irregular Hodge theory

In algebraic geometry, the language of “jumping” refers not to transport across edges but to changes in geometric type. For rank-$2$ $\mathfrak{sl}_2$-parabolic connections on $\mathbb P^1$ with $\det E\simeq \mathcal O_{\mathbb P^1}(-1)$, Komyo and Saito stratify the moduli space $M$ by bundle type
$$
E\simeq \mathcal O(k)\oplus \mathcal O(-k-1),\qquad
0\le k\le \lfloor (n-3)/2\rfloor.
$$
The open stratum $M^0$ is the non-jumping locus with $E\simeq \mathcal O(-1)\oplus \mathcal O(-1)$. For $k=1$, one has $E\simeq \mathcal O(1)\oplus \mathcal O(-2)$, and the locus $M^1\subset M$ is the jumping locus; its preimage in the cyclic moduli $\widehat M$ is an exceptional divisor. Apparent singularities $q_i$ and dual parameters $p_i$ furnish coordinates on the non-jumping locus and lead to an injective morphism
$$
M^0\hookrightarrow \mathrm{Hilb}^{\,n-3}(K_n'),
$$
where $K_n'$ is obtained from a blown-up Hirzebruch surface. For $n=5$, $\widehat M\to M$ is the blow-up along $M^1$, and the map by apparent singularities and dual parameters embeds $\widehat M$ into a further blow-up $\widetilde{\mathrm{Hilb}}^2(K_5')$ [1611.00971].

The geometric content of the jump is explicit in the degeneration process. On a chart with local parameters $(q_1,p_1;q_2,p_2;\lambda)$, one performs an elementary modification of the non-jumping family; as $q_2\to q_1$, the underlying bundle changes from $\mathcal O(-1)\oplus \mathcal O(-1)$ to $\mathcal O(1)\oplus \mathcal O(-2)$. The two apparent points coalesce, their duals blow up, and finite symmetric combinations such as
$$
s=\frac{1/p_2'-1/p_1'}{q_2'-q_1'}
$$
survive as local coordinates on the final blow-up. Here “jumping” denotes passage to a different bundle-type stratum rather than a dynamical hop [1611.00971].

A related but distinct use appears in irregular Hodge theory. For an irreducible holonomic connection $\mathcal E$ on $\mathbb P^1$ with possibly irregular singularities, Sabbah’s canonical decreasing irregular Hodge filtration
$$
\cdots \supset F_{\mathrm{irr}}^\beta H \supset F_{\mathrm{irr}}^{\beta'}H\supset\cdots
$$
has jumps at finitely many real numbers $\{\beta_1<\cdots<\beta_r\}$, called the jumping indices. Jakob and Reiter compute these indices and their multiplicities for four rigid $G_2$-connections $E_1,E_2,E_3,E_4$. For $E_1$, $E_2$, and $E_4$, all with a single irregular slope $\alpha=\tfrac12$, the nonzero irregular Hodge numbers are
$$
d_{1/2}^0=2,\qquad d_{1/2}^1=3,\qquad d_{1/2}^2=2.
$$
For $E_3$, the jumps depend on the relation between parameters $a$ and $b$ and can involve $\{2b,1,2(1-b)+1\}$, yielding several distinct multiplicity tables. In this context, a jumping connection is indexed by discontinuities of a filtration rather than by graph-theoretic or dynamical shortcuts [2007.11248].

## 5. Jump maps in hybrid and communication networks

In multi-agent consensus with time-driven jumps, the system state is hybrid: $x\in\mathbb R^N$ evolves continuously and a clock $s$ resets at jump times. With flow graph $\mathcal G_f$ and jump graph $\mathcal G_j$, Laplacians $L_{\rm f}$ and $L_{\rm j}$, and coupling gain $\alpha$, the decentralized dynamics are
$$
\dot x=-L_{\rm f}x,\qquad \dot s=1 \quad \text{on flows},
$$
and
$$
x^+=(I-\alpha L_{\rm j})x,\qquad s^+=0 \quad \text{on jumps}.
$$
The relevant invariant object is the hybrid multi-consensus subspace
$$
\mathcal K=\mathrm{Im}\,P(\pi_{\mathcal H}^\star),
$$
where $\pi_{\mathcal H}^\star$ is the coarsest partition of the vertex set that is simultaneously an almost equitable partition for both $\mathcal G_f$ and $\mathcal G_j$. Under the bounds
$$
\alpha<\min_{\lambda\in\sigma(L_{\rm j})\setminus\{0\}}\frac{2\,\Re\lambda}{|\lambda|^2},
\qquad
\alpha<\min_{\lambda\in\sigma(L_{\rm j})\setminus\{0\}}\frac{1}{\Re\lambda},
$$
the subspace $\mathcal K$ is globally asymptotically stable [2110.14277].

In wireless sensor networks, the jump is operational rather than algebraic. The DMRF protocol maintains a Forwarding Candidate Set (FCS) of one-hop neighbors and switches from ordinary hop-by-hop forwarding to jumping transmission mode when a failure, congestion, or void region is detected, or when the remaining-time factor
$$
\lambda_i=\mathrm{TminDelay}_i/L_i
$$
falls below a threshold $\theta_{\rm jump}$. For jumps from node $i$ to node $j$, the success ratio and jumping probability are
$$
\mathrm{Suc}_{ij}=S_{ij}/T_{ij},\qquad
P_{ij}=\mathrm{Suc}_{ij}\Big/\sum_{k=1}^m \mathrm{Suc}_{ik},
$$
with $m=|FCS_i|$. Failures and acknowledgments update $T_{ij}$ and $S_{ij}$, and the resulting feedback mechanism shifts probability mass toward reliable neighbors [1004.2322].

The published simulation setup uses 100 nodes in a $20\,\mathrm m\times 20\,\mathrm m$ area, communication radius $30\,\mathrm m$, MAC/PHY $802.11$/CSMA-CA at $200\,\mathrm{Kb/s}$, and deadlines equal to $\mathrm{TminDelay}\times 1.2$. Reported delivery ratios for DMRF are $98\%$, $92\%$, $85\%$, and $80\%$ at faulty-node ratios $0.1$, $0.2$, $0.3$, and $0.4$, respectively. Under increasing void-circle radius, average delay remains $31\,\mathrm{ms}$, $32\,\mathrm{ms}$, and $35\,\mathrm{ms}$ for radii $4\,\mathrm m$, $6\,\mathrm m$, and $8\,\mathrm m$, while competing hop-by-hop schemes increase sharply [1004.2322].

## 6. Pointer jumping and the complexity-theoretic meaning of jumps

In communication complexity, “jumping” refers to evaluation along a layered composition of pointers. In the one-way number-on-the-forehead model with players $\mathrm{plr}_1,\dots,\mathrm{plr}_k$, the $k$-layer pointer-jumping problem has one root in layer $L_0$, $n$ vertices in each intermediate layer, and either $2$ output vertices for the Boolean problem $\mathrm{mpj}_k$ or $n$ output vertices for the non-Boolean version. The Boolean recursion is
$$
\mathrm{mpj}_k(i,f_2,\dots,f_{k-1},x)
=
x\bigl(f_{k-1}\circ\cdots\circ f_2(i)\bigr),
$$
and deterministic one-way NOF communication complexity is denoted $C(\cdot)$ [0802.2843].

The principal upper bound is sublinear for every fixed $k\ge 3$:
$$
C(\mathrm{mpj}_k)
=
O\!\Bigl(n\bigl(k\,\log\log n/\log n\bigr)^{\frac{k-2}{k-1}}\Bigr).
$$
In particular,
$$
C(\mathrm{mpj}_3)=O\bigl(n\sqrt{\log\log n/\log n}\bigr)=o(n).
$$
This rules out an $\Omega(n)$ lower bound for Boolean pointer jumping in the full one-way NOF model for constant $k\ge 3$ [0802.2843].

The same paper isolates a restricted class of protocols in which every player is collapsing: player $\mathrm{plr}_j$ sees the layers behind her individually but the layers ahead only through their composition. Under this restriction, any deterministic one-way NOF protocol for $\mathrm{mpj}_k$ must satisfy
$$
\mathrm{cost}(P)\ge n-\tfrac12\log n-2=n-O(\log n).
$$
The lower bound is proved using “crossing pairs,” a combinatorial device replacing earlier information-theoretic styles of argument. At the same time, the non-Boolean pointer-jumping problem $\widehat{\mathrm{pj}}_k$ still admits a collapsing-player upper bound of
$$
O\bigl(n\,\log^{(k-1)}n\bigr)
$$
when all but one of the layer functions are permutations [0802.2843].

A common misconception would be to treat this usage as analogous to skip-connections in neural networks. The formal resemblance is limited: both involve bypassing naive layerwise progression, but the object of study here is one-way communication under partial information, not representational flow in a learnable architecture.

## 7. Structural themes and scope

Across these literatures, three motifs recur. First, jumping connections alter locality. Fixed-distance permutations in quantum walks supplement nearest-neighbor motion; skip-connections in GCNs bypass depth; DMRF jumps bypass impaired one-hop neighborhoods; pointer jumping composes hidden layers into a single effective map. Second, jumps often induce different asymptotic regimes: Gaussian-like spreading versus localization in quantum walks, layer-sensitive sparsification constraints in GCNs, or constant clusters versus hybrid consensus arcs in multi-agent systems [1106.2036], [2507.05533], [2110.14277]. Third, in geometry and Hodge theory the jump is stratificational rather than dynamical, marking discontinuous changes of bundle type or filtration grade [1611.00971], [2007.11248].

This suggests that “jumping connections” functions best as a family resemblance term rather than a universal definition. Its precise content depends on whether the ambient structure is a Hilbert-space evolution, a neural message-passing architecture, a moduli stack, a hybrid dynamical system, a routing protocol, or a communication game. What remains stable across these cases is the introduction of a mechanism that bypasses an underlying local rule while preserving enough structure to admit rigorous analysis.

Source: https://www.emergentmind.com/topics/jumping-connections