Relaxed Routing Approaches

• Relaxed Routing is a family of techniques that intentionally weakens exact routing constraints to enable tractable, analyzable, and adaptive network solutions.
• It applies linear, convex, and mixed-integer relaxations to transform computationally hard routing problems into efficiently solvable formulations with subsequent rounding or certification.
• The approach balances trade-offs such as throughput versus fidelity and computational speed versus optimality, making it essential in domains like quantum networks, vehicle routing, and dataflow control.

Relaxed Routing denotes a family of routing formulations in which a hard routing restriction is deliberately weakened in order to preserve tractability, analyzability, robustness, or operational responsiveness. In the recent literature, the term refers to linear relaxation of integer routing-and-purification decisions in space–ground integrated quantum networks (Hu et al., 17 Jan 2025), a stress-driven cost adaptation mechanism for network control (Ren et al., 21 Sep 2025), relaxed continuity and relaxed elementarity in vehicle-routing branch-and-price (Bhat et al., 28 Feb 2026), and a soft terminal-constraint formulation of network transport via relaxed Schrödinger bridges (Chen et al., 2018). Other works use the same phrase for relaxing equal-cost shortest-path assumptions, rigid graph dependence, or strict timing/branch semantics in communication networks, trajectory routing, and cyber-physical dataflow (Tam et al., 2011).

1. Semantic scope and recurring formulations

The cited literature uses “Relaxed Routing” in several technically distinct senses. In all of them, the central operation is to replace an exact routing constraint by a softer mechanism that still admits formal analysis or implementation.

Domain	Relaxed object	Representative mechanism
Quantum networking	Integer routing variables	LP relaxation with rounding
Network control	Fixed shortest-path costs	Leaky-integrator cost landscape
Vehicle routing	Continuity or elementarity	Relaxed continuity; ng/LA routes
Backbone TE	ECMP or dynamic core rerouting	Few precomputed paths or trees
CPS/dataflow	Fixed branch execution under timing	Controlled splitters/joiners with mode variables
Stochastic transport	Exact terminal marginal	KL-penalized terminal distribution

In space–ground integrated quantum networks, relaxed routing means converting an integer program over per-request scheduling, entanglement consumption, purification usage, and repeater swapping counts into a linear program by relaxing all integer variables to reals, then rounding while preserving flow, capacity, and end-to-end noise constraints (Hu et al., 17 Jan 2025). In adaptive network control, the same term denotes the System Relaxation Algorithm, where routing costs evolve through a damped integral of structural stress rather than remaining fixed shortest-path weights (Ren et al., 21 Sep 2025).

In vehicle-routing and column-generation settings, “relaxed” usually modifies the pricing problem rather than the master problem. Relaxed continuity decouples exact inter-leg trajectory consistency to obtain lower bounds during pricing, whereas ng-route and Local Area route relaxations weaken elementary-route constraints by permitting controlled revisits (Bhat et al., 28 Feb 2026). In robust stochastic routing, the relaxation applies to the endpoint condition: the terminal marginal need only be close to the target distribution, measured by a Kullback–Leibler penalty, instead of matching it exactly (Chen et al., 2018).

2. Linear, convex, and mixed-integer relaxations

In "Dynamic Routing in Space-Ground Integrated Quantum Networks" (Hu et al., 17 Jan 2025), the routing problem is formulated as an integer program over three communication modes: free-space-only, ground-only, and hybrid paths. The objective is

$\max \sum_{k\in \mathbb{K}} Y_k$

subject to flow conservation, link and repeater capacities, and an end-to-end noise bound,

$$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$

Here, fidelity is linearized into additive edge noise by $\mu_e=\log(1/\gamma_e)$, time-varying free-space edges use $\mu_e=\log(1/f_e(t))$, and purification on free-space edges is suppressed through $p_e=0$. Relaxed routing consists of replacing the integer variables $Y_k, x^k_e, \phi^k_e, \alpha^k_r$ by real-valued variables, solving the resulting LP in polynomial time, and then rounding in a constraint-preserving way. The relaxation is used because exact integer optimization is computationally prohibitive on operational timescales when entanglement generation, swapping, and satellite-ground visibility vary over time. Its principal benefit is timely scheduling of additional purification while balancing throughput against fidelity.

In "Coflow Scheduling in Data Centers: Routing and Bandwidth Allocation" (Shi et al., 2018), relaxed routing is a convex relaxation of an unsplittable coflow-routing problem. The original MINLP uses $x^i_{uv}\in\{-1,0,1\}$ and end-to-end rates $b_i$ to minimize the Coflow Completion Time $\max_i V_i/b_i$. The relaxation allows $-1\le x^i_{uv}\le 1$, introduces $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$0, $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$1, and $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$2, and yields a convex program with affine flow constraints and link-capacity constraints of the form

$$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$3

This converts discrete single-path routing into a flow-splittable multi-commodity formulation. CoRBA solves the relaxed problem, extracts integral routes using maximum-capacity paths, and then recomputes the optimal bandwidth allocation for the chosen routes.

In "Optimal Routing across Constant Function Market Makers with Gas Fees" (Escudero et al., 3 Mar 2026), the original model is mixed-integer because pool activation incurs fixed gas costs through $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$4 and the buy variables satisfy $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$5. The relaxed problem replaces $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$6 by $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$7. The resulting KKT system yields explicit activation thresholds: $$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$8 The paper also derives no-trade conditions and an explicit approximation bound between the relaxed solution and the original mixed-integer solution,

$$\sum_{e\in E}(\mu_e x^k_e-p_e \phi^k_e) + \sum_{r\in R} \sigma_r \alpha^k_r \le N^{th}Y_k.$$9

The same section of the literature therefore uses relaxed routing to couple routing, activation, and resource costs without discarding the original economic or physical constraints.

3. Adaptive, dynamical, and mode-aware routing

In "System Relaxation for Interpretable and Adaptive Network Control" (Ren et al., 21 Sep 2025), Relaxed Routing is not a convex relaxation of an optimization problem but a controlled dynamical system. Structural stress is quantified by unnormalized betweenness centrality,

$\mu_e=\log(1/\gamma_e)$0

and the controller integrates normalized stress into a pressure state: $\mu_e=\log(1/\gamma_e)$1 Shortest-path routing is then recomputed on the adapted node-cost vector $\mu_e=\log(1/\gamma_e)$2. The theory is explicitly non-smooth: the path-set changes discontinuously, and convergence is analyzed through Krasovskii regularization, a Lyapunov-like potential

$\mu_e=\log(1/\gamma_e)$3

and dwell-time arguments. The paper proves existence of a global attractor and practical stability, and reports topology-dependent behavior such as an 83.76% reduction of $\mu_e=\log(1/\gamma_e)$4 with average path length increase of 5.63% in Barabási–Albert graphs.

"Dynamic information routing in complex networks" (Kirst et al., 2015) uses “relaxed routing” in a still different sense: routing is changed by switching the collective dynamical reference state while keeping the structural adjacency and coupling matrices fixed. Information transfer is quantified through delayed mutual information and delayed transfer entropy. Around a phase-locked reference state, fluctuations obey a linearized dynamics

$\mu_e=\log(1/\gamma_e)$5

so state-dependent slopes of the coupling functions change effective directional influence. Routing is therefore realized by switching among synchronized or metastable collective states rather than by recomputing graph paths.

In "Real-time Mode-Aware Dataflow" (Roumage et al., 13 Jan 2025), relaxed routing is the combination of conditional execution with relaxed real-time constraints. RMDF extends PolyGraph using controlled splitters and controlled joiners whose 0/1 production or consumption rates are driven by one-hot mode variables. For a controlled splitter with outputs $\mu_e=\log(1/\gamma_e)$6,

$\mu_e=\log(1/\gamma_e)$7

The model remains analyzable only under explicit structural restrictions: control areas are disjoint, no edge crosses a control-area boundary, all timed actors in a control area share the same frequency, and all internal rates are 0 or 1. Consistency and liveness are guaranteed if every per-mode PolyGraph instantiation is consistent and live.

4. Relaxed continuity and relaxed elementarity in vehicle routing

In "Optimal Solutions for the Moving Target Vehicle Routing Problem via Branch-and-Price with Relaxed Continuity" (Bhat et al., 28 Feb 2026), relaxed routing means relaxing the continuity constraints that couple exact interception points across successive target-windows in the pricing subproblem. The true tour cost $\mu_e=\log(1/\gamma_e)$8 requires a continuous speed-admissible trajectory through all target-windows, but the lower-bound construction allows discontinuity at target-windows and composes segment-to-segment costs by Bellman recursions. Exact continuity is restored only when a complete tour is certified. The lower-bounding pairwise segment problem is solved by a two-point SOCP, and the resulting dominance rule compares feasible upper bounds to optimistic lower bounds segmentwise. This makes exact branch-and-price practical on instances with moving targets and time-dependent costs, with numerical results showing more than an order-of-magnitude speedup on instances with up to 25 targets.

"Optimal Solutions for the Moving Target Vehicle Routing Problem with Obstacles via Lazy Branch and Price" (Bhat et al., 23 Mar 2026) extends the same principle to obstacle-rich motion planning. Lazy BPRC solves the restricted master problem using lower bounds $\mu_e=\log(1/\gamma_e)$9 derived from relaxed continuity and defers exact GCS-based motion-planning costs $\mu_e=\log(1/f_e(t))$0 until needed. The relaxed lower bound is obtained over window segments and shortest collision-free space paths, while exact costs are computed by shortest-path search on a Graph of Convex Sets. The reported speedups are up to 44× versus Non-Lazy BPRC and up to 13× versus No-Affine-Heuristic.

In column generation for capacitated VRP, relaxed routing often means replacing elementary-route pricing by a larger route class. "Local Area Routes for Vehicle Routing Problems" (Mandal et al., 2022) defines Local Area route relaxations through recursively defined special indices. If a customer repeats in a route, the cycle is permitted only if a breaker occurs at a special index and the repeated customer is not an ng-neighbor of that breaker. Formally, Local Area routes satisfy

$\mu_e=\log(1/f_e(t))$1

with the inclusion relation

$\mu_e=\log(1/f_e(t))$2

This strengthens ng-routes while preserving efficient pricing through precomputed local arcs.

"Relaxed Dual Optimal Inequalities for Relaxed Columns" (Haghani et al., 2020) addresses the same pricing regime from the dual side. The pricing subproblem is solved over ng-routes, and Smooth DOI and Flexible DOI are adapted to relaxed columns. Because these DOI are not always valid for the relaxed route set, any DOI variable active at termination is removed and column generation is restarted. On 46 CVRP instances, S-DOI and SF-DOI provide an average speedup of about 20% in time, reducing mean runtime from 1354 s to 1006 s and 999 s, respectively.

5. Relaxing shortest-path, equal-cost, and exact-path assumptions

In communication networks, relaxed routing often means departing from rigid equal-cost shortest-path logic. "Trimming the Multipath for Efficient Dynamic Routing" (Tam et al., 2011) replaces ECMP’s use of all equal-cost paths by a small number of centrally selected end-to-end tunnels. Candidate paths satisfy the relaxed path-length constraint

$\mu_e=\log(1/f_e(t))$3

and the selection heuristic minimizes a per-path maximum-utilization score

$\mu_e=\log(1/f_e(t))$4

The paper reports that with $\mu_e=\log(1/f_e(t))$5 and $\mu_e=\log(1/f_e(t))$6, maximum link utilization closely matches ECMP in irregular and fat-tree topologies, while in irregular graphs the relaxed path set can outperform ECMP.

"Relieving Core Routers from Dynamic Routing with off-the-shelf Equipment and Protocols" (Mamede et al., 2016) moves the relaxation from the path level to the control architecture. The core no longer performs dynamic TE recomputation; instead, it stores many a priori computed paths aggregated into a small number of trees, and a logically centralized controller adapts only edge load splitting. The reduction in forwarding state is explicit: in ATT, 2366 paths are aggregated into 54 trees; in Sprint, 1984 paths into 52 trees.

"Routes Obey Hierarchy in Complex Networks" (Csoma et al., 2017) uses relaxed routing to describe empirical deviations from strict shortest-path behavior. Across air transportation, AS-level Internet, brain, and word-ladder datasets, 20–40% of observed paths are inflated relative to the shortest path by up to 4–5 hops. The paper attributes this to three concurrent routing policies: prefer short paths, conform hierarchy through a unimodal closeness-centrality profile, and prefer downstream moves that avoid the core when possible.

"TrajRoute" (Siampou et al., 2024) relaxes dependence on a traffic-annotated road graph by routing directly over raw trajectory points combined with road-segment points in a unified grid index. Search is performed by A* over trajectory and road neighbors, with movement costs taken from timestamp differences on trajectories and from distance divided by speed limit on road segments. As trajectory coverage increases from 5% to 100%, travel-time MAE decreases significantly and the average number of road segments used drops sharply around 25% coverage.

In "Relaxed Schroedinger bridges and robust network routing" (Chen et al., 2018), the relaxed object is the endpoint constraint rather than the path set. The optimization problem is

$\mu_e=\log(1/f_e(t))$7

so the terminal marginal $\mu_e=\log(1/f_e(t))$8 is penalized for deviating from $\mu_e=\log(1/f_e(t))$9 instead of being forced to match it exactly. The optimal path measure is obtained from a generalized Schrödinger system, and the fixed-point map contracts the Hilbert metric with ratio strictly less than $p_e=0$0. This formulation produces diversified maximum-entropy flows that remain robust under random link failures encoded in the reference transition matrices.

6. Trade-offs, guarantees, and limitations

Across the cited literature, relaxed routing is repeatedly introduced when exact routing constraints are too expensive or too brittle. In quantum networking, exact integer optimization is described as too slow for dynamic, time-varying link conditions (Hu et al., 17 Jan 2025). In relaxed-continuity vehicle routing, solving exact chain SOCPs or obstacle-aware GCS costs at every partial extension is the computational bottleneck (Bhat et al., 23 Mar 2026). In ng-route pricing, the underlying ESPPRC is strongly NP-hard, and relaxed columns are used to control state explosion (Haghani et al., 2020). The recurring design choice is therefore to accept a structured relaxation inside an optimization or control loop, then restore feasibility or exactness only where it is most valuable.

The trade-offs are explicit and parameterized. In the quantum formulation, the configurable noise threshold $p_e=0$1 governs the throughput–fidelity balance: higher $p_e=0$2 increases throughput but lowers average fidelity, while lower $p_e=0$3 does the opposite (Hu et al., 17 Jan 2025). In SRA, increasing hub costs reduces stress peaks and raises high-load throughput, but can slightly lengthen paths; in homogeneous topologies the same mechanism shifts toward latency-for-reliability trade-offs (Ren et al., 21 Sep 2025). In trimmed multipath routing, $p_e=0$4 or $p_e=0$5 controls path stretch versus congestion relief (Tam et al., 2011). In CFMM routing, larger gas fees $p_e=0$6 and smaller activation bounds $p_e=0$7 enlarge the no-trade region and suppress pool activation (Escudero et al., 3 Mar 2026). In TrajRoute, increasing the temporal window $p_e=0$8 reduces reliance on road segments but worsens time relevance and increases travel-time error (Siampou et al., 2024).

The guarantees are similarly varied. Some are exact-recovery results: relaxed continuity in MT-VRP restores exact continuity when a tour is certified (Bhat et al., 28 Feb 2026); Lazy BPRC remains an exact optimal algorithm because lower bounds are used only for pruning and master-problem relaxation (Bhat et al., 23 Mar 2026). Some are dynamical guarantees: SRA has a global attractor and practical stability under dwell-time policies (Ren et al., 21 Sep 2025); relaxed Schrödinger bridges have a unique solution and a fixed-point solver with contraction ratio less than $p_e=0$9 (Chen et al., 2018). Others are analyzability guarantees under structural restrictions, as in RMDF, where per-mode consistency and liveness imply consistency and liveness of the full mode-aware system (Roumage et al., 13 Jan 2025).

The limitations are domain-specific and frequently made explicit. The quantum LP does not include explicit time-slot scheduling variables, synchronization constraints, quantum memory lifetimes, or stochastic link-state uncertainty (Hu et al., 17 Jan 2025). RMDF requires disjoint control areas, equal frequencies for timed actors inside a control area, and 0/1 internal rates to preserve analyzability (Roumage et al., 13 Jan 2025). CFMM routing with gas fees assumes continuously differentiable invariant functions and relies on generalized convexity conditions for sufficient KKT optimality (Escudero et al., 3 Mar 2026). Relaxed continuity in MT-VRP and MT-VRP-O depends on segment discretization and on lower bounds that are optimistic by construction (Bhat et al., 28 Feb 2026). These caveats do not negate the approach; they specify the regimes in which a given relaxation remains principled.

Taken together, the literature presents Relaxed Routing as a general methodological pattern: hard routing constraints are weakened in a controlled way, the relaxed problem is solved or analyzed efficiently, and feasibility, optimality, or behavioral guarantees are recovered through rounding, certification, dominance, mode restrictions, or fixed-point structure. The term is therefore best understood not as a single routing algorithm, but as a family of rigorously designed relaxations that trade exactness for tractability while keeping the trade explicit and analyzable.