Tadpole Graphs: Structure & Applications

• Tadpole graphs are defined by a cycle (head) with an attached path (tail), creating a distinct asymmetric structure with a single junction.
• They appear in diverse studies, including competitive online exploration with optimal competitive ratios, Dirichlet eigenvalue minimization, and quantum graph models.
• Their simple yet nontrivial topology enables precise analysis and sharp bounds across algorithmic, spectral, and nonlinear PDE frameworks.

A tadpole graph is a graph formed by a cycle with an attached path, equivalently a wedge-sum of a cycle and a path at a single vertex. In the combinatorial notation used in recent work, a tadpole graph $T_{k,m}$ consists of a cycle $C$ of $k\ge 3$ vertices together with a path $P$ of $m\ge 1$ vertices whose endpoint is identified with a distinguished cycle vertex; the same object is also described as a dragon or kite (Brandt et al., 2019). In current arXiv literature, tadpole graphs appear in at least three technically distinct settings: competitive online exploration on weighted graphs, extremal spectral problems for Dirichlet eigenvalues, and nonlinear PDEs on metric graphs (Brandt et al., 2019, He et al., 29 Mar 2026, Duboscq et al., 7 May 2025).

1. Structural definitions and standard models

In its finite combinatorial form, a tadpole graph $T_{k,m}$ is specified by a cycle $C_k$ and a path $P_m$ with one endpoint of $P_m$ identified with one vertex of $C_k$. In the notation of Brandt et al., if the stem has $C$0 vertices then $C$1 stem edges are labeled $C$2, while the $C$3 cycle edges are labeled $C$4, so that $C$5 (Brandt et al., 2019). In the Faber–Krahn literature, the special graph $C$6 consists of a $C$7-cycle head and a path tail attached at the neck vertex; one convenient labeling is

$C$8

with cycle edges $C$9 and tail $k\ge 3$0 (He et al., 29 Mar 2026).

The finite tadpole has one vertex of degree $k\ge 3$1 at the junction or neck, one pendant vertex of degree $k\ge 3$2 at the tail end, and all remaining vertices of degree $k\ge 3$3 (Akker et al., 2024, He et al., 21 Mar 2026). This asymmetric degree pattern is the source of several of its extremal properties. In exploration problems it creates a single branching uncertainty at the junction; in Dirichlet eigenvalue problems it creates a single bottleneck between a short cycle and a long path; in metric-graph PDE models it becomes a loop joined to a half-line at one vertex (Duboscq et al., 7 May 2025).

A distinct metric-graph version is the tadpole graph $k\ge 3$4, which depends on a length parameter $k\ge 3$5 and is the union of a loop edge $k\ge 3$6 of total length $k\ge 3$7, identified with $k\ge 3$8, and a half-line edge $k\ge 3$9, identified with $P$0, meeting at a single vertex $P$1 (Duboscq et al., 7 May 2025). This continuum model is not a finite combinatorial graph, but it preserves the same head-tail geometry.

2. Single-agent online exploration on weighted tadpoles

In the online graph-exploration model, a searcher starts at a designated start vertex $P$2, which may lie on the stem or on the cycle. Upon first visiting a vertex $P$3, the algorithm learns the unique ID of $P$4, all neighbors of $P$5, and the weights of the incident edges. Each traversal of an edge $P$6 incurs cost $P$7, and the objective is to return to $P$8 after visiting every vertex at least once while minimizing total cost (Brandt et al., 2019). The performance measure is the competitive ratio

$P$9

where $m\ge 1$0 is the minimum-cost closed walk visiting all vertices (Brandt et al., 2019).

For tadpole graphs, Brandt et al. show that no online algorithm can achieve competitive ratio below $m\ge 1$1, extending a lower-bound construction of Miyazaki et al. to preserve the tadpole structure throughout the adversarial revelation process (Brandt et al., 2019). The adversary grows three unknown rays from a hidden junction and keeps the searcher uncertain which two rays form the cycle; with suitable parameters,

$m\ge 1$2

which can be made greater than $m\ge 1$3 by choosing $m\ge 1$4 large (Brandt et al., 2019).

The matching upper bound is attained by a greedy nearest-unvisited strategy. The algorithm repeatedly computes, in the explored subgraph, the shortest-path distance from the current vertex to every known but unvisited vertex, selects a minimizer, traverses a shortest path to it, and finally returns to $m\ge 1$5 by a shortest path (Brandt et al., 2019). The result is optimal in the competitive sense: on any weighted tadpole graph $m\ge 1$6, this greedy algorithm has competitive ratio at most $m\ge 1$7, and the lower bound shows that this is best possible (Brandt et al., 2019).

The amortized proof depends on a charging scheme in which each step is charged to the cheapest incident edge leading to an unvisited neighbor, while edges on certain return paths are charged individually. Each edge weight is charged at most twice in nearest-neighbor phases, with only $m\ge 1$8 extra path-charges (Brandt et al., 2019). A crucial structural fact is that an optimal tour on a tadpole has exactly one of two forms:

• Shape 1: every cycle edge is traversed once and every stem edge twice.
• Shape 2: one very heavy cycle edge $m\ge 1$9 is omitted and all other edges are traversed twice.

In Shape 2, one has $T_{k,m}$0 the sum of the other cycle edges; in both shapes the analysis yields $T_{k,m}$1 (Brandt et al., 2019). This establishes the head-tail asymmetry of the tadpole as one of the few nontrivial restricted graph classes for which the optimal online competitive ratio is exactly characterized.

3. Multi-agent online exploration: time and energy regimes

The multi-agent version places $T_{k,m}$2 agents at a common start node $T_{k,m}$3 in an initially unknown tadpole graph. Whenever an agent visits a new node, its incident edges and edge weights are revealed to all agents, so the explored part is globally shared (Akker et al., 2024). Two cost models are used. In the time model, the cost is the wall-clock time until the last agent returns to $T_{k,m}$4; in the energy model, the cost is the maximum total distance traversed by any single agent (Akker et al., 2024).

For tadpole graphs, the competitive picture depends strongly on the number of agents. The lower bounds are $T_{k,m}$5 in the time model for any $T_{k,m}$6, and $T_{k,m}$7 in the energy model for $T_{k,m}$8 (Akker et al., 2024). The constructive upper bounds are as follows.

Agents	Time model	Energy model
$T_{k,m}$9	$C_k$0-competitive	$C_k$1-competitive
$C_k$2	$C_k$3	$C_k$4
$C_k$5	$C_k$6	$C_k$7

For $C_k$8, the paper gives a simple randomized AMP-style strategy that is $C_k$9-competitive in both models. At the junction, two outgoing subpaths are chosen uniformly at random and then explored by the cycle AMP rule (Akker et al., 2024). For $P_m$0, the algorithm “AMP-Tadpole-3” sends three agents into the three visible directions but moves only the agent minimizing $P_m$1, thereby balancing traversed distances. This achieves energy-model ratio $P_m$2 and time-model ratio at most $P_m$3 (Akker et al., 2024). For $P_m$4, the “Split-4” strategy partitions the team at each degree-$P_m$5 junction into $P_m$6 subteams, one per outgoing edge, applies AMP inside each team, and achieves the time lower bound of $P_m$7 while preserving optimal energy usage (Akker et al., 2024).

These results show that the combinatorial asymmetry of a tadpole graph does not preclude exact online optimality in stronger distributed models. A plausible implication is that the single junction is sufficiently structured to permit near-perfect load balancing once enough agents are available, especially under global communication.

4. Dirichlet eigenvalues and Faber–Krahn extremality

A second major line of work studies tadpole graphs as minimizers of first Dirichlet eigenvalues. For the normalized combinatorial $P_m$8-Laplacian on a finite connected graph with boundary given by pendant vertices, the operator is

$P_m$9

with Dirichlet condition $P_m$0 and Rayleigh quotient

$P_m$1

The first Dirichlet eigenvalue is the minimum of this quotient over nonzero $P_m$2 satisfying the boundary condition (He et al., 29 Mar 2026).

He and Yu prove that if $P_m$3 is any connected graph whose boundary consists of exactly $P_m$4 edges and $P_m$5, then

$P_m$6

with equality if and only if $P_m$7 is isomorphic to $P_m$8 (He et al., 29 Mar 2026). Their proof uses a maximizing path for a positive first eigenfunction, a degree-sum estimate, and a surgery that copies the eigenfunction onto the tail of $P_m$9 while flattening values on the triangular head. The comparison yields $C_k$0 and $C_k$1, hence the tadpole has no larger Rayleigh quotient (He et al., 29 Mar 2026).

An analogous sharp result holds for the unnormalized combinatorial Laplacian

$C_k$2

whose first Dirichlet eigenvalue satisfies

$C_k$3

He and Yu show that among all connected graphs with boundary on $C_k$4 vertices, and separately among all connected graphs with boundary and $C_k$5 edges, the unique minimizer is again $C_k$6 (He et al., 21 Mar 2026).

Several comparison lemmas explain why the triangular head is extremal. In both spectral papers, larger cycle lengths are disfavored: $C_k$7 in the normalized $C_k$8-Laplacian setting, and $C_k$9 in the combinatorial Laplacian setting (He et al., 29 Mar 2026, He et al., 21 Mar 2026). Paths are also larger: $C$00 and $C$01 (He et al., 29 Mar 2026, He et al., 21 Mar 2026). This makes $C$02 the discrete Faber–Krahn minimizer within the classes considered.

For the concrete case $C$03 in the combinatorial Laplacian model, the symmetry-invariant eigenfunction on $C$04 leads to the cubic

$C$05

whose unique root in $C$06 gives $C$07; by comparison, $C$08 and $C$09 (He et al., 21 Mar 2026). The numerical gap illustrates the general principle that a triangle head plus a long tail yields the smallest first Dirichlet eigenvalue among the admissible competitors.

5. Metric tadpoles and nonlinear Schrödinger ground states

In the metric-graph setting, the tadpole $C$10 is the union of a loop edge of total length $C$11 and a half-line meeting at a vertex $C$12. A function on $C$13 is a pair $C$14 with $C$15 and $C$16, subject to continuity

$C$17

The Hamiltonian with a delta interaction at the vertex is

$C$18

with domain determined by continuity and the derivative jump condition

$C$19

where $C$20 is repulsive (Duboscq et al., 7 May 2025).

The stationary cubic NLS on this graph is

$C$21

and its action functional is

$C$22

where

$C$23

Because $C$24 is unbounded below on $C$25, minimization is restricted to the Nehari manifold

$C$26

and the action ground-state level is

$C$27

(Duboscq et al., 7 May 2025).

The main existence theorem states that for fixed $C$28 and $C$29, ground states exist in two asymptotic regimes. If

$C$30

then there exists $C$31 such that $C$32 for all $C$33; if

$C$34

then there exists $C$35 such that $C$36 for all $C$37, where the line soliton action is $C$38 (Duboscq et al., 7 May 2025). Existence follows from variational arguments and profile decomposition: once $C$39, escape of mass along the tail is ruled out (Duboscq et al., 7 May 2025).

The paper also classifies standing-wave profiles on the loop by phase-plane methods. The loop component satisfies

$C$40

with first integral $C$41, and solutions fall into dnoidal, sech, and cnoidal families according to whether $C$42, $C$43, or $C$44 (Duboscq et al., 7 May 2025). On the tail one always has a translated sech profile,

$C$45

with $C$46 determined by the Robin boundary condition (Duboscq et al., 7 May 2025). Numerical experiments for $C$47, $C$48, and loop lengths $C$49 show convergence in at most about $C$50 steps and a transition of mass from the tail-dominated regime to the loop-dominated regime as $C$51 increases (Duboscq et al., 7 May 2025).

6. Terminology, scope, and related uses

In graph theory, “tadpole graph” refers to the cycle-plus-path structure described above. The term is also used in physics in the phrase “tadpole diagrams,” but this denotes a different object: a one-particle-reducible diagram in constant electromagnetic fields that is constructed algebraically from one-particle-irreducible constant-field diagrams by differentiation with respect to the field strength tensor (Karbstein, 2017). This usage is unrelated to combinatorial tadpole graphs except for the visual metaphor.

Within mathematics, two distinct notions should likewise be separated. A finite tadpole graph such as $C$52 or $C$53 is a discrete combinatorial object used in online algorithms and spectral graph theory (Brandt et al., 2019, He et al., 21 Mar 2026). A metric tadpole such as $C$54 is a quantum graph consisting of a loop and a half-line endowed with differential operators and vertex conditions (Duboscq et al., 7 May 2025). The shared geometry is the same head-tail topology, but the analytical frameworks are different: shortest-path cost and competitive ratio in the former, Dirichlet spectra and nonlinear bound states in the latter.

Taken together, the recent literature presents tadpole graphs as a compact test case where a simple non-tree, non-symmetric topology admits exact or sharp results across several domains. In online exploration, the optimal competitive ratio is exactly $C$55 for a single agent and can drop to $C$56 in the energy model with three agents (Brandt et al., 2019, Akker et al., 2024). In Dirichlet eigenvalue minimization, the triangle-head tadpole $C$57 is the unique extremizer in the graph classes considered (He et al., 29 Mar 2026, He et al., 21 Mar 2026). In nonlinear analysis on metric graphs, the same topology supports asymptotically proven action ground states under a repulsive delta vertex condition (Duboscq et al., 7 May 2025). This convergence of results across algorithmics, spectral graph theory, and PDEs helps explain why the tadpole has become a recurrent model geometry in current arXiv research.