PDNE: Performance-Dependent Network Evolution

• PDNE is defined as a closed-loop framework where network topology is iteratively optimized based on task-specific performance signals.
• It integrates diverse methods—from wireless topology control to spectral rewiring and reservoir computing—to balance structural dynamics with functional outcomes.
• Empirical studies demonstrate that performance-driven evolution can enhance connectivity, reduce path lengths, and yield compact, efficient network architectures.

Searching arXiv for the cited PDNE-related papers and metadata. arXiv search query: (Kwon et al., 2017) OR (Chidambaran et al., 2018) OR (Yadav et al., 2024) OR (Yadav, 13 Mar 2026) OR (Li et al., 6 Aug 2025) OR (Karalus et al., 2012) OR (Brede, 2011) OR (Assunção et al., 2019) OR (Furfaro et al., 2022) OR (Whetsell et al., 16 Jun 2026) Performance-Dependent Network Evolution (PDNE) denotes a class of adaptive processes in which network structure changes as a function of measured performance. Across the arXiv literature, the term covers several distinct but structurally related formulations: distributed topology control in wireless ad hoc networks (Kwon et al., 2017), spectrum-targeted rewiring of graphs for prescribed diffusion dynamics (Karalus et al., 2012), coevolutionary growth–rewiring models driven by pathlength optimization (Brede, 2011), performance-conditioned reservoir growth and pruning (Yadav et al., 2024, Yadav, 13 Mar 2026), multi-objective neuroevolution of artificial neural network topologies (Chidambaran et al., 2018), adaptive training-resource evolution for deep neural architectures (Assunção et al., 2019), and longitudinal co-evolution between collaboration networks and national scientific performance (Whetsell et al., 16 Jun 2026). In all of these formulations, topology is not treated as fixed background structure; it is the object of optimization or co-adaptation under feedback from throughput, prediction error, task success, spectral distance, average path length, or citation-based performance.

1. Conceptual definition and scope

PDNE is most precisely understood as a closed-loop coupling between structural modification and performance evaluation. A network, or a networked system, is repeatedly evaluated with respect to a task-specific criterion; structural changes are then accepted, rejected, or biased according to whether they improve that criterion. The performance signal may be local or global, myopic or foresighted, scalar or multi-objective, and directly measured on task execution or inferred from a dynamical operator’s spectrum.

The literature exhibits several distinct instantiations of this general idea. In wireless networking, each intermediate node adapts its transmission range to maximize a long-term utility balancing throughput or goodput gain against transmission power cost (Kwon et al., 2017). In spectral graph evolution, rewiring is accepted only if the integrated density of states of the graph Laplacian moves closer to a prescribed target, thereby shaping diffusion dynamics (Karalus et al., 2012). In coevolutionary graph growth, random node addition is interleaved with rewiring steps that reduce shortest-path distances, with the relative timescale of optimization controlling whether the resulting topology remains exponential, becomes heavy-tailed, or condenses into super-hubs (Brede, 2011). In reservoir computing, nodes are added or deleted only when prediction error decreases or does not worsen, yielding compact task-optimized reservoirs (Yadav et al., 2024, Yadav, 13 Mar 2026). In neuroevolution, candidate artificial neural networks are selected and diversified according to mission performance and, in one formulation, an explicit experience-gain objective (Chidambaran et al., 2018).

A central implication of this body of work is that PDNE is not a single algorithm. Rather, it is a design principle: topology evolves under explicit performance pressure. The resulting performance dependence may act through reinforcement-style utilities (Kwon et al., 2017), elitist non-dominated sorting (Chidambaran et al., 2018), deterministic hill-climbing acceptance (Karalus et al., 2012, Yadav et al., 2024), grammar-based mutation with adaptive training budgets (Assunção et al., 2019), or stochastic actor-oriented co-evolution of ties and actor behavior (Whetsell et al., 16 Jun 2026).

2. Core mechanisms and mathematical formulations

At the algorithmic level, PDNE systems typically share four ingredients: a structural state space, an evaluation metric, a set of admissible topology-changing operators, and an acceptance or selection rule.

In distributed wireless PDNE, the state of each node is the expected number of effective neighbors, the action is the change in transmission range, and the utility is

$U(s, a, s') = u + \omega R(s, s') - (1 - \omega)a,$

with immediate throughput gain

$R(s, s') = \gamma(s') - \gamma(s),$

where $\gamma(s)$ is a concave increasing function of throughput versus effective neighbors (Kwon et al., 2017). The local policy is obtained by value iteration on a per-node Markov decision process:

$$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$

The Bellman operator is a $\rho$-contraction, so value iteration converges to the optimal value function under the assumptions stated in the paper (Kwon et al., 2017).

In the spectral formulation, performance is defined as the squared $L^2$ distance between the current logarithmic integrated density of states and a target spectrum:

$$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$

with target behavior encoded through

$I(\lambda)\propto \lambda^{d_s/2},$

which implies

$$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$

for diffusion dynamics (Karalus et al., 2012). Rewiring is accepted only if $\Delta$ decreases.

In pathlength-driven graph evolution, the objective is the average shortest-path length

$R(s, s') = \gamma(s') - \gamma(s),$0

while local rewiring acceptance is based on the node-level proxy

$R(s, s') = \gamma(s') - \gamma(s),$1

A candidate rewiring is accepted if it reduces $R(s, s') = \gamma(s') - \gamma(s),$2; the number of optimization steps per growth step, denoted $R(s, s') = \gamma(s') - \gamma(s),$3, controls the relative strength of optimization and assembly (Brede, 2011).

Reservoir-based PDNE formulations use prediction error as the direct structural criterion. In the task-agnostic minimal-reservoir framework, the network is grown from a two-node seed and modified through node addition and deletion, with changes retained only if they improve the error metric, typically normalized mean squared error:

$R(s, s') = \gamma(s') - \gamma(s),$4

The reservoir update is

$R(s, s') = \gamma(s') - \gamma(s),$5

and the readout is trained by ridge regression (Yadav et al., 2024).

In the Wilson–Cowan reservoir study, prediction performance is quantified channel-wise by

$R(s, s') = \gamma(s') - \gamma(s),$6

and node additions are accepted only if test NMSE strictly decreases across all channels, while deletions are accepted if test NMSE does not increase in any channel (Yadav, 13 Mar 2026). This yields an explicit grow–prune cycle with stopping threshold $R(s, s') = \gamma(s') - \gamma(s),$7 (Yadav, 13 Mar 2026).

Multi-objective neuroevolution introduces a second kind of performance dependence. In MENTOR, the performance objective is

$R(s, s') = \gamma(s') - \gamma(s),$8

with

$R(s, s') = \gamma(s') - \gamma(s),$9

while the experience-gain objective is defined from a minimum spanning tree in experience space:

$\gamma(s)$0

Selection is then driven by elitist non-dominated sorting over $\gamma(s)$1 rather than by a single scalar objective (Chidambaran et al., 2018).

A distinct but related variant appears in longitudinal collaboration networks. There, PDNE is formalized through coupled stochastic actor-oriented models with network evaluation

$\gamma(s)$2

and behavior evaluation

$\gamma(s)$3

so that tie formation depends on performance and performance depends on network position (Whetsell et al., 16 Jun 2026).

3. Major research lines

The existing literature does not define PDNE around a single application domain. Instead, the term organizes several research programs in which structure and function are co-determined.

Domain	Structural variable	Performance signal
Wireless ad hoc networks	Transmission range and topology	Throughput or goodput gain versus power cost
Spectral graph evolution	Edge rewiring	Distance to target spectrum
Reservoir computing	Node growth and pruning	NMSE or task-specific prediction error
Neuroevolution	ANN topology and weights	Mission performance and experience-gain
Deep architecture evolution	Training-budget attribute and architecture	Validation or test accuracy under adaptive training
Collaboration networks	Backbone ties between countries	FWCI and network position
Topological compression	Degree-preserving rewiring	Average shortest path length

In wireless ad hoc networking, PDNE is coupled to network coding. Random linear network coding over $\gamma(s)$4 yields packet anonymity, and the paper’s proposition states that repeated mixing makes packet information and terminal fields asymptotically identical across multi-hop RLNC paths (Kwon et al., 2017). This decouples one-hop forwarding value from the full global dependency structure, allowing each intermediate node to solve a local MDP instead of a global combinatorial control problem. The resulting topology adapts to mobility, link failures, and density changes while remaining distributed (Kwon et al., 2017).

In spectral graph evolution, the objective is not a graph statistic such as degree variance or algebraic connectivity, but the full spectrum of the dynamics’ time-evolution operator (Karalus et al., 2012). The paper demonstrates the emergence of sub-diffusive behavior by targeting spectral dimensions $\gamma(s)$5 and $\gamma(s)$6, below the normal-diffusion value $\gamma(s)$7 for a two-dimensional lattice. This is a notably strong formulation of PDNE because performance is specified at the level of the entire dynamical operator rather than a few summary observables (Karalus et al., 2012).

In coevolutionary growth models, the key insight is that random assembly and pathlength minimization act on distinct timescales (Brede, 2011). Random attachment alone yields exponential degree distributions; optimization alone pushes the system toward star-like structures. Intermediate timescales generate power-law tails, hierarchical clustering, and nontrivial degree mixing (Brede, 2011). This suggests that PDNE can be viewed as a balance between exploration by growth and exploitation by performance optimization.

In reservoir computing, PDNE is used to discover minimal yet effective recurrent substrates. One formulation starts from a two-node seed and evolves task-specific minimal networks that satisfy a target accuracy threshold on benchmarks including Sin–Cos mappings, NARMA-5/10/15, Lorenz chaotic trajectory generation, and Van der Pol limit-cycle generation (Yadav et al., 2024). Another formulation evolves compact reservoirs for Wilson–Cowan population dynamics, starting from $\gamma(s)$8 nodes and accepting only additions that strictly improve channel-wise NMSE and deletions that do not degrade it (Yadav, 13 Mar 2026).

In neuroevolution, the term is tied to topological and parametric evolution of ANNs under measured task performance. MENTOR modifies NEAT through altered selection, speciation, and mutation, and explicitly adds an experience-gain objective to mitigate deception and overfitting to the design-of-experiments scenarios (Chidambaran et al., 2018). By contrast, Fast-DENSER++ makes training time itself evolvable, so candidate deep architectures are evaluated under progressively longer budgets as required, producing fully trained models at the end of evolution without post-evolution fine-tuning (Assunção et al., 2019).

The conceptual breadth of PDNE is further widened by two additional studies. The APN framework proposes a protein-inspired genotype–phenotype mapping in which “silicon DNA” encodes network modules and evolution is conditioned on performance, but the paper is primarily conceptual and does not report empirical experiments (Lao et al., 2024). In global science studies, the co-evolution of international research collaboration and national scientific performance is modeled empirically as reciprocal selection and influence, with performance proxied by Elsevier’s fractional FWCI and collaboration ties derived from Web of Science backbone networks (Whetsell et al., 16 Jun 2026).

4. Empirical regularities and reported outcomes

Several PDNE studies report measurable structural or performance gains relative to non-adaptive baselines, but the nature of these gains depends on the domain.

In wireless networking, simulations with two sources and two terminals, PPP-distributed intermediates of density $\gamma(s)$9, state space size $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$0, action space size $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$1, $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$2, and typical discount factor $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$3 showed that the proposed strategy outperformed myopic and static baselines in goodput and connectivity (Kwon et al., 2017). In the Wi-Fi Direct evaluation over a $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$4 area with node density $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$5, the proposed method achieved goodput $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$6 Mbps, successful connectivity $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$7, and energy efficiency $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$8 Mbps/dBm, compared with Myopic at $$V_{\tau+1}(s) = \max_{a \in A} \sum_{s' \in S} P(s' \mid s, a)\big[U(s,a,s') + \rho V_\tau(s')\big].$$9 Mbps and $\rho$0, Traskov at $\rho$1 Mbps and $\rho$2, and TCLE at $\rho$3 Mbps and $\rho$4 (Kwon et al., 2017).

In spectral diffusion shaping, the mean distance-to-target $\rho$5 decreases rapidly during evolution, and the resulting return probability follows the prescribed sub-diffusive slopes. The evolved networks exhibit $\rho$6 for $\rho$7 and $\rho$8 for $\rho$9 (Karalus et al., 2012). At the same time, the evolved topologies become more degree-heterogeneous and assortative, while clustering remains broadly distributed. This indicates that very similar spectral performance can be realized by structurally diverse networks (Karalus et al., 2012).

In growth–optimization coevolution, several regime transitions are reported as the optimization timescale $L^2$0 changes (Brede, 2011). Under global optimization, around $L^2$1 the degree distribution follows $L^2$2 with $L^2$3, while degree-dependent clustering and nearest-neighbor degree exhibit critical sign changes around the same region. Specifically, the paper reports $L^2$4 for $L^2$5 at $L^2$6, compared with $L^2$7 at $L^2$8 and $L^2$9 at $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$0; similarly, $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$1 shows $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$2 at $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$3, compared with $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$4 at $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$5 and $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$6 at $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$7 (Brede, 2011). For local optimization with $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$8, heavy-tailed regimes appear already for $$\Delta(\tilde I,\tilde I^{\mathrm{target}})=\int_{\ln \tilde \lambda_{\min}}^\infty \left|\tilde I(\ln \tilde \lambda)-\tilde I^{\mathrm{target}}(\ln \tilde \lambda)\right|^2\, d(\ln \tilde \lambda),$$9, but with $I(\lambda)\propto \lambda^{d_s/2},$0 over much of that range (Brede, 2011).

Reservoir PDNE yields compact structures with striking task dependence. The minimal-network study reports average sizes of approximately $I(\lambda)\propto \lambda^{d_s/2},$1 for Sin–Cos-1, $I(\lambda)\propto \lambda^{d_s/2},$2 for Sin–Cos-2, $I(\lambda)\propto \lambda^{d_s/2},$3 for Van der Pol limit-cycle generation, $I(\lambda)\propto \lambda^{d_s/2},$4 for Lorenz chaotic trajectory generation, $I(\lambda)\propto \lambda^{d_s/2},$5 for NARMA-5, $I(\lambda)\propto \lambda^{d_s/2},$6 for NARMA-10, and $I(\lambda)\propto \lambda^{d_s/2},$7 for NARMA-15 (Yadav et al., 2024). These evolved reservoirs obey the scaling

$I(\lambda)\propto \lambda^{d_s/2},$8

with $I(\lambda)\propto \lambda^{d_s/2},$9 and $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$0, implying an exponent $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$1 in $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$2 (Yadav et al., 2024). The same study reports a strong asymmetry between input and readout-node fractions in evolved networks: for PDNE, $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$3, $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$4, common nodes $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$5, and unique nodes $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$6 (Yadav et al., 2024).

In Wilson–Cowan modeling, PDNE evolves reservoirs that accurately predict $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$7 and $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$8 across unseen stimulus amplitudes and generalize zero-shot to stimuli with varying pulse number, position, and amplitude (Yadav, 13 Mar 2026). Across $$P_0(t)=\frac{1}{N}\sum_{\nu=1}^{N} e^{-\lambda_\nu t}\propto t^{-d_s/2}$$9 repetitions, the final network size is reported as $\Delta$0 nodes, with approximately $\Delta$1 E-specific, $\Delta$2 I-specific, and $\Delta$3 Shared nodes (Yadav, 13 Mar 2026). Population-level connectivity recovers the correct Wilson–Cowan signs for three of four interaction types: E$\Delta$4E is $\Delta$5, E$\Delta$6I is $\Delta$7, I$\Delta$8E is $\Delta$9, while I$R(s, s') = \gamma(s') - \gamma(s),$00I is $R(s, s') = \gamma(s') - \gamma(s),$01 even though the target model has $R(s, s') = \gamma(s') - \gamma(s),$02 (Yadav, 13 Mar 2026).

The neuroevolution study reports that dual-stage evolution, consisting of an initial multi-objective stage on $R(s, s') = \gamma(s') - \gamma(s),$03 followed by single-objective performance refinement, yields superior generalization in unseen scenarios relative to single-objective evolution (Chidambaran et al., 2018). In novel V-REP layouts, the UGV single-stage method failed all $R(s, s') = \gamma(s') - \gamma(s),$04 missions, whereas the dual-stage method succeeded in $R(s, s') = \gamma(s') - \gamma(s),$05 (Chidambaran et al., 2018). For the Swarm-Robot case, both methods reached destinations in multiple scenarios, but dual-stage reached faster and single-stage showed fewer collisions (Chidambaran et al., 2018).

Fast-DENSER++ reports statistically significant gains over Fast-DENSER when models are evaluated as fully trained outputs. Across $R(s, s') = \gamma(s') - \gamma(s),$06 runs on CIFAR-10 with $R(s, s') = \gamma(s') - \gamma(s),$07 generations and $R(s, s') = \gamma(s') - \gamma(s),$08-ES with $R(s, s') = \gamma(s') - \gamma(s),$09, validation accuracy is $R(s, s') = \gamma(s') - \gamma(s),$10 for Fast-DENSER++ versus $R(s, s') = \gamma(s') - \gamma(s),$11 for Fast-DENSER, and test accuracy under evolutionary evaluation is $R(s, s') = \gamma(s') - \gamma(s),$12 versus $R(s, s') = \gamma(s') - \gamma(s),$13, with Mann–Whitney $R(s, s') = \gamma(s') - \gamma(s),$14 and $R(s, s') = \gamma(s') - \gamma(s),$15, respectively (Assunção et al., 2019). The method is slower than Fast-DENSER, at $R(s, s') = \gamma(s') - \gamma(s),$16 hours per generation versus $R(s, s') = \gamma(s') - \gamma(s),$17, but remains far faster than DENSER at $R(s, s') = \gamma(s') - \gamma(s),$18 hours per generation (Assunção et al., 2019).

In collaboration-network co-evolution, the full stochastic actor-oriented model supports reciprocal performance dependence. In the network equation, the FWCI main effect is $R(s, s') = \gamma(s') - \gamma(s),$19 with $R(s, s') = \gamma(s') - \gamma(s),$20, and the FWCI$R(s, s') = \gamma(s') - \gamma(s),$21distance interaction is $R(s, s') = \gamma(s') - \gamma(s),$22 with $R(s, s') = \gamma(s') - \gamma(s),$23, both with $R(s, s') = \gamma(s') - \gamma(s),$24 (Whetsell et al., 16 Jun 2026). In the behavior equation, degree has $R(s, s') = \gamma(s') - \gamma(s),$25 and average alter performance has $R(s, s') = \gamma(s') - \gamma(s),$26, again with $R(s, s') = \gamma(s') - \gamma(s),$27 (Whetsell et al., 16 Jun 2026). This indicates that higher-performing countries are more likely to form backbone ties and that centrality and collaborator performance increase subsequent national FWCI.

5. Structural principles, assumptions, and recurrent themes

Several recurrent design principles appear across otherwise disparate PDNE formulations.

A first principle is acceptance by improvement. In the spectral diffusion framework, rewiring is accepted only if the distance $R(s, s') = \gamma(s') - \gamma(s),$28 to the target DOS decreases (Karalus et al., 2012). In minimal-reservoir PDNE, additions and deletions are accepted only if they improve task error (Yadav et al., 2024). In the Wilson–Cowan reservoir study, additions must strictly decrease test NMSE across all output channels, while deletions must not increase it (Yadav, 13 Mar 2026). In topological compression, rewiring is directed by bounds on average-distance change so that removals minimize expected damage and additions maximize expected benefit (Li et al., 6 Aug 2025). This common mechanism suggests that many PDNE methods are hill-climbing or monotone-improvement processes, even when the specific state spaces differ.

A second principle is performance-conditioned structural sparsification or compactification. The Wilson–Cowan study starts from a minimal seed reservoir and reaches compact sparse networks with density decreasing during growth and then stabilizing (Yadav, 13 Mar 2026). The minimal-reservoir study reports unexpectedly sparse topologies obeying a node-density scaling law (Yadav et al., 2024). Topological compression explicitly rewires connected graphs to reduce average shortest-path length while preserving $R(s, s') = \gamma(s') - \gamma(s),$29, $R(s, s') = \gamma(s') - \gamma(s),$30, connectivity, and the degree distribution over the whole process (Li et al., 6 Aug 2025). In Fast-DENSER++, the compactness pressure is indirect: time-budget growth is granted to promising individuals, while non-time mutations reset the training budget to the default, thereby avoiding over-allocation to unproven offspring (Assunção et al., 2019).

A third principle is decoupling or abstraction of complex dependencies. Wireless PDNE relies on packet anonymity induced by RLNC so that a node can model the network as a local node–environment interaction (Kwon et al., 2017). Spectrum-based PDNE replaces direct dynamical simulation with whole-spectrum matching (Karalus et al., 2012). Collaboration-network PDNE uses SAOM ministeps to decompose co-evolution into actor-level tie and behavior updates (Whetsell et al., 16 Jun 2026). These are different technical strategies, but each reduces a high-dimensional global problem into tractable local or summary evaluations.

A fourth principle is timescale sensitivity. In graph growth with pathlength optimization, the balance between assembly and optimization is determined by $R(s, s') = \gamma(s') - \gamma(s),$31 (Brede, 2011). In wireless PDNE, the foresight parameter $R(s, s') = \gamma(s') - \gamma(s),$32 controls the trade-off between convergence speed and long-term planning (Kwon et al., 2017). In Fast-DENSER++, the training-time mutation and fairness retraining mechanism make evaluation depth itself performance-dependent (Assunção et al., 2019). This suggests that PDNE is often not only about what is optimized, but also when optimization is allowed to act relative to growth, learning, or environmental change.

These frameworks also rest on strong assumptions. Wireless PDNE assumes homogeneous PPP node placement, i.i.d. link failures with rate $R(s, s') = \gamma(s') - \gamma(s),$33, and a coarse state representation given by the number of effective neighbors (Kwon et al., 2017). Spectral PDNE assumes linear dynamics and an operator whose eigenvalue distribution is the relevant performance descriptor (Karalus et al., 2012). The minimal-reservoir study does not report weight-initialization details, leakage values, or formal convergence guarantees (Yadav et al., 2024). The Wilson–Cowan application uses $R(s, s') = \gamma(s') - \gamma(s),$34, strict decrease-only addition, and no explicit complexity penalty beyond pruning acceptance (Yadav, 13 Mar 2026). The collaboration study discretizes FWCI into $R(s, s') = \gamma(s') - \gamma(s),$35 ordinal categories and binarizes weighted co-authorship networks through a disparity filter, changing the interpretation from all collaborations to backbone ties (Whetsell et al., 16 Jun 2026).

6. Relation to adjacent paradigms, limitations, and future directions

PDNE overlaps with, but is not equivalent to, several adjacent paradigms.

It is not identical to preferential attachment. In the growth–optimization model, new nodes attach uniformly at random, and heavy tails arise from subsequent performance-driven rewiring rather than degree-proportional attachment (Brede, 2011). In the fitness-driven deactivation model, which can also be read as a performance-conditioned network evolution process, incoming links go uniformly to active nodes, while deactivation probability is inversely proportional to nodal fitness,

$R(s, s') = \gamma(s') - \gamma(s),$36

so degree heterogeneity emerges through fitness-controlled persistence rather than direct preferential attachment (Xu et al., 2010). The model yields structured exponential networks under homogeneous fitness distributions and structured scale-free networks under heterogeneous fitness distributions, while recovering the clustering scalings $R(s, s') = \gamma(s') - \gamma(s),$37 and $R(s, s') = \gamma(s') - \gamma(s),$38 (Xu et al., 2010).

It is also not identical to standard NEAT-style neuroevolution. MENTOR retains direct genome encoding, innovation numbers, and structural mutation, but departs from original NEAT by using NSGA-II-style elitist non-dominated sorting, tournament selection within species, and experience-gain as a second objective (Chidambaran et al., 2018). Likewise, the hybrid feed-forward ANN functionalized by evolution combines backpropagation for weights with Pseudo-Darwinian structural mutation across generations, rather than evolving weights and topology jointly (Furfaro et al., 2022). That study reports structural convergence on MNIST, with best evolved accuracy $R(s, s') = \gamma(s') - \gamma(s),$39 versus $R(s, s') = \gamma(s') - \gamma(s),$40 for controls after $R(s, s') = \gamma(s') - \gamma(s),$41 epochs, as well as early solutions in cart-pole relative to non-evolving baselines (Furfaro et al., 2022).

A common misconception would be to treat PDNE as necessarily multi-objective. Some formulations are explicitly multi-objective, as in MENTOR’s optimization over performance and experience-gain (Chidambaran et al., 2018). Others are strictly single-objective or threshold-based, such as NMSE minimization in Wilson–Cowan reservoir evolution (Yadav, 13 Mar 2026), pathlength minimization in graph optimization (Brede, 2011), or average-distance reduction in topological compression (Li et al., 6 Aug 2025). Conversely, another misconception would be to assume that PDNE always requires an explicit complexity penalty. The Wilson–Cowan reservoir study states that acceptance and stopping rely on NMSE rather than an explicit multi-term objective combining error and complexity; compactness is enforced by grow–prune acceptance tests (Yadav, 13 Mar 2026).

Limitations are likewise domain-specific. Spectral matching can leave large structural degeneracy, with many topologies achieving similar dynamics (Karalus et al., 2012). Topological compression decreases clustering coefficient $R(s, s') = \gamma(s') - \gamma(s),$42 as it reduces average distance $R(s, s') = \gamma(s') - \gamma(s),$43, which may erode modularity or other desirable mesoscopic organization (Li et al., 6 Aug 2025). Strict greedy acceptance can slow convergence or trap search near thresholds, as noted for performance-evolved Wilson–Cowan reservoirs (Yadav, 13 Mar 2026). Collaboration-network PDNE may amplify stratification because performance-dependent selection, preferential attachment, and transitivity jointly favor already central high-performing countries (Whetsell et al., 16 Jun 2026).

The future directions proposed across the literature are notably convergent. Wireless PDNE suggests SINR-aware rewards, POMDP formulations, and multi-objective utilities including latency, fairness, and reliability (Kwon et al., 2017). Reservoir-based PDNE suggests extension to multi-population, spatially structured, and chaotic neuronal models, as well as incorporation of biophysical constraints such as Dale’s principle (Yadav, 13 Mar 2026). Neuroevolution work proposes deeper ANN topologies, broader benchmark suites such as OpenAI Gym, and intra-generational learning (Chidambaran et al., 2018). Topological compression suggests multi-objective formulations that jointly control average distance, clustering, modularity, robustness, and spectral constraints (Li et al., 6 Aug 2025). The APN program proposes biologically constrained genotype–phenotype mappings and ecological multi-objective evolution, but remains prospective rather than experimentally established (Lao et al., 2024).

Taken together, these works indicate that PDNE is best regarded as a general research framework for structure–function co-adaptation. Its central claim is not that one particular topological update rule is universal, but that network architecture can be treated as a dynamical variable whose evolution is directly conditioned on measured system performance. Across communication, learning, neuroscience, complex systems, and scientometrics, that claim has produced a common methodological pattern: define performance, restrict admissible structural mutations, couple evaluation to acceptance or selection, and study the resulting emergent organization (Kwon et al., 2017, Karalus et al., 2012, Brede, 2011, Yadav et al., 2024, Yadav, 13 Mar 2026, Whetsell et al., 16 Jun 2026).