Braess' Paradox in Networked Systems

• Braess' paradox is a counterintuitive phenomenon in which adding extra capacity to a network, such as an additional road or link, degrades overall performance due to altered equilibrium flows.
• The phenomenon has been demonstrated across various domains using methods like linear volume-delay functions, dynamic microsimulation, and analyses of AC power grids and quantum transport.
• Understanding topology, traffic information regimes, and decentralized routing is crucial for mitigating the paradox in both transportation and coupled infrastructure networks.

Braess’ paradox is the counterintuitive phenomenon in which adding an extra road, strengthening an edge, or otherwise increasing network capacity degrades equilibrium performance rather than improving it. In its classical traffic form, the added link changes route incentives so that selfish route choice produces a worse user equilibrium; in later work, the same logic has been formulated for generalized traffic networks with arbitrary linear volume-delay functions, microscopic exclusion-process traffic, flows over time, conservative supply networks, AC power grids, wireless SINR games, several quantum-transport settings, open chemical reaction networks, coupled power-transport systems, and tandem-running ants (Zverovich et al., 2012, Bittihn et al., 2016, Macko et al., 2010, Manik et al., 2022, Schäfer et al., 2022, Mou et al., 13 Dec 2025, Bairagya et al., 27 Mar 2026).

1. Canonical traffic-network formulation

The canonical setting is the four-node Braess network with three origin-destination routes,

$P_1=a-b-d,\qquad P_2=a-c-d,\qquad P_3=a-b-c-d,$

where the augmented network $N^+$ includes the middle link $(b,c)$, and the reduced network $N$ removes it (Zverovich et al., 2012). In the Wardrop or Nash-equilibrium language, the user optimum is the state in which all used routes have equal travel times and no driver can improve by switching unilaterally; a microscopic traffic treatment defines the system optimum as the state minimizing the maximum travel time of all used routes (Bittihn et al., 2020).

A standard linear example uses

$T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$

For $N=6$ cars, the 4-link network has a pure user optimum with 3 cars on each of the two routes and

$T^{(4)}_{\mathrm{puo}}=83,$

whereas the 5-link network has a pure user optimum with 2 cars on each of the three routes and

$T^{(5)}_{\mathrm{puo}}=92.$

The mixed-strategy values are likewise larger with the added link,

$$\langle T^{(4)}_{\mathrm{muo}} \rangle = 88.5,\qquad \langle T^{(5)}_{\mathrm{muo}} \rangle = 93.6923.$$

The paradox is therefore not restricted to pure equilibria (Bittihn et al., 2020).

A generalized asymmetric treatment replaces each effective link by a linear volume-delay function

$t_{ij}(f_{ij})=\alpha_{ij}+\beta_{ij}f_{ij},$

with $N^+$0 and $N^+$1. In that setting, Braess’ paradox occurs when the augmented-network equilibrium travel time exceeds the reduced-network equilibrium travel time,

$N^+$2

This generalization removes the symmetry assumptions used in earlier analyses and shows that the paradox is not tied to the perfectly symmetric textbook network (Zverovich et al., 2012).

2. Equilibrium structure, topology, and time dependence

In static Wardrop routing, Braess’ paradox is tightly linked to topology: a network admits the paradox if and only if it is not series-parallel, equivalently if it contains the Wheatstone network as a topological minor (Macko et al., 2010). That characterization changes in the model of congestion games with flow over time introduced by Koch and Skutella, where each edge has a capacity $N^+$3, a free-flow transit time $N^+$4, queues form at edge tails, and the model is FIFO. In that dynamic setting, a feasible flow over time is a Nash equilibrium if and only if flow is sent only along currently shortest paths (Macko et al., 2010).

The flow-over-time literature shows that networks can exhibit Braess’ paradox even when they are series-parallel. A family $N^+$5 is series-parallel, in fact extension-parallel, yet admits the paradox in the dynamic model; for suitable capacities and transit times, the Braess ratio can be made arbitrarily close to

$N^+$6

This is substantially larger than the classical static bound $N^+$7 for networks with $N^+$8 vertices. The dynamic paradox is also not symmetric under transposition: there are networks $N^+$9 such that $(b,c)$0 admits the paradox but the transpose $(b,c)$1 does not (Macko et al., 2010).

A further refinement is the pseudo-paradox of generalized static networks, where

$(b,c)$2

over an interval of demand values rather than at a single isolated point. That phenomenon is distinct from Braess’ paradox proper, but it is structurally important because it identifies demand regimes in which the added link changes equilibrium flow patterns without changing equilibrium cost (Zverovich et al., 2012).

3. Microscopic, stochastic, and dynamic traffic realizations

A major line of work replaces linear cost functions by microscopic traffic dynamics on networks of totally asymmetric exclusion processes (TASEPs). In these models, each road is a one-dimensional lattice, each cell can hold at most one particle, particles move stochastically to an empty site, and the single-edge travel time under periodic boundary conditions is

$(b,c)$3

The exclusion principle makes travel times nonlinear in density and introduces jammed states and fluctuation effects that do not appear in classical linear models (Bittihn et al., 2016).

In a stochastic TASEP Braess network with route choice governed by turning probabilities, four regimes appear. At low densities, the added edge lowers travel times. At slightly higher densities, the classical Braess phenomenon occurs. At intermediate densities, strong fluctuations dominate because links enter a domain-wall state. At high densities, the added link again lowers travel times (Bittihn et al., 2016). The intermediate regime is especially notable because route travel times become broad rather than sharply concentrated, making the identification of user and system optima less straightforward.

When the same network is studied with fixed individual route choices rather than stochastic turning, the picture changes. Stable travel times can then be measured throughout the phase diagram, the domain-wall phase vanishes, and gridlock states occupy a large part of phase space. In that formulation, travel time reduction due to the new road occurs only at really low densities, roughly

$(b,c)$4

and Braess’ paradox dominates the largest part of the phase diagram (Bittihn et al., 2018).

Dynamic microsimulation yields a closely related conclusion. A 2-by-3 grid network studied in the Flow framework with the SUMO microscopic simulator reproduces Braess-like behavior under vehicle-level route choice. The added path does not increase total network output flow; depending on geometry and speed limits, output-flow reduction ranges from $(b,c)$5 to $(b,c)$6 as edge length increases from 50 m to 400 m, from $(b,c)$7 to $(b,c)$8 as speed limit decreases from 35 m/s to 10 m/s, and reaches $(b,c)$9 in the most attractive shortcut case of 400 m and 10 m/s. The average travel time from node 1 to node 8 becomes larger in the added-path network at sufficiently high demand, with reported increases up to $N$0, and the added-path network has almost the same capacity as the baseline network despite its extra physical space (Zhuang et al., 2022).

4. Traffic information, route guidance, and decentralized routing

The modern-information question is whether the paradox survives when route choice is not externally fixed but instead based on realistic traffic information. Two information regimes have been studied in a microscopic Braess network. In the personal-historical regime, each agent remembers travel times from the last $N$1 rounds and uses the mean of those remembered times. In the public-predictive regime, edge densities

$N$2

are converted into predicted edge travel times

$N$3

and route estimates are formed by summing predicted edge times, in direct analogy to app-like traffic guidance (Bittihn et al., 2020).

Under personal historical information, user optima are realized in all four representative test states of both the 4-link and 5-link networks, and Braess states remain Braess states: the realized 5-link user optimum still has higher travel times than the corresponding 4-link optimum. Under public predictive information, the low-density user optima are realized stably, while at higher density the 5-link network often reaches the expected user optimum only on average and with fluctuations. In Braess-phase states, however, the 5-link travel times remain higher than those of the 4-link system. The explicit conclusion is that modern traffic information systems are not able to resolve Braess’ paradox (Bittihn et al., 2020).

A related study examines mixed populations of commuters using either personal-historical or public-predictive information. It uses five information splits, from $N$4 personal and $N$5 public to $N$6 personal and $N$7 public, with parameter values

$N$8

Across all five splits, the 4-link network reaches route travel times close to attainable user optima, whereas the 5-link network remains worse than the 4-link network. Higher fractions of public-predictive information produce more route switching and more visible oscillations, but do not steer the system toward the attainable mixed user optimum (Bittihn et al., 2020).

The routing-control literature also describes a distinct response: utility design rather than route-information refinement. In a multi-agent routing model, router agents using the Ideal Shortest Path Algorithm suffer from side-effects across space and time, and load balancing is likewise suboptimal as far as global cost averaged across time is concerned. A Wonderful Life Utility construction within the Collective Intelligence framework yields a COIN routing algorithm that almost always avoids Braess’ paradox in the experiments, and in some cases per-packet costs under ISPA are up to $N$9 higher than under COIN (Tumer et al., 2011).

5. Supply networks and power-grid formulations

Outside traffic proper, Braess’ paradox has been formulated as a generic property of conservative supply and transport networks. In a graph $T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$0 with node potentials $T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$1, edge strengths $T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$2, and flows

$T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$3

an edge $T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$4 is Braessian if an infinitesimal increase in its strength increases the magnitude of the maximum-flow edge. The perturbation problem can be mapped exactly to currents in a resistor network driven by a dipole current source across the strengthened edge, with effective conductances

$T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$5

This yields a computational criterion and a topological heuristic, rerouting alignment, for identifying Braessian edges (Manik et al., 2022).

The rerouting-alignment heuristic predicts an edge to be Braessian if it is aligned by rerouting to the maximum-flow edge. It was tested on a $T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$6 square lattice, a Voronoi tessellation from random points, and the IEEE 300-bus power-grid test case, each with 200 random source/sink input configurations; the reported conclusion is that the classifier performs reasonably well across all three cases. The same framework also shows that deliberately weakening Braessian edges can, in some situations, mitigate overload rather than exacerbate it (Manik et al., 2022).

In AC power grids, the paradox becomes operationally critical because adding a transmission line or upgrading an existing one can increase loading on non-upgraded lines and thereby threaten synchrony or trigger blackouts. A laboratory demonstration uses a 4-node AC grid with two physical synchronous generators and two inverter-based nodes acting as virtual synchronous machines. Reducing the reactance of one line from

$T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$7

to

$T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$8

upgrades that line, yet the measured current on line 3 decreases while the current on line 2 increases. A topological predictor based on flow alignment identifies about $T_1=T_3=10n,\qquad T_2=T_4=50+n,\qquad T_5=10+n.$9 to $N=6$0 of Braessian line extensions correctly, with roughly $N=6$1 to $N=6$2 false predictions (Schäfer et al., 2022).

Control can modify this picture. In a controlled second-order Kuramoto-like AC-grid model, secondary control adds a restoring term to the phase angle and changes the effective injection to

$N=6$3

Linearization gives the stability condition

$N=6$4

for all modes $N=6$5. Numerical investigations on an eight-node network show that when secondary control is applied to all nodes, the studied topology changes no longer destroy the operating state and Braess’ paradox is reliably avoided. When only generators are controlled, the outcome is topology-dependent: in one case uncontrolled failure occurs at $N=6$6, whereas generator-only control causes earlier failure at $N=6$7; in another case generator-only control completely prevents the paradox (Tchuisseu et al., 2018).

6. Nonclassical analogues in wireless, quantum, and chemical systems

Wireless networks provide a game-theoretic analogue in the SINR model. There the “improvements” are interference cancellation, power control, combined power control plus interference cancellation, and lowering the SINR threshold $N=6$8. In each case the optimum feasible set can improve while equilibrium throughput worsens. The degradation is a constant factor for interference cancellation, power control, and improved $N=6$9, and $T^{(4)}_{\mathrm{puo}}=83,$0 when power control is combined with interference cancellation; matching upper bounds show these examples are essentially tight (Dinitz et al., 2013).

Several quantum-transport settings exhibit direct analogues. In a finite-width quantum ring with an attached central horizontal branch, adding the extra channel does not necessarily improve transmission: for sufficiently narrow branch width $T^{(4)}_{\mathrm{puo}}=83,$1, the extra channel is energetically inaccessible and acts as an additional scatterer, whereas for larger $T^{(4)}_{\mathrm{puo}}=83,$2 the reduction in transmission is caused by destructive interference between the branch and the ring arms (Sousa et al., 2013). In chaotic open quantum dots, adding an over-capacity lead can suppress current flow; the weak-localization contribution develops a minimum, with a critical hyperflow condition

$T^{(4)}_{\mathrm{puo}}=83,$3

which marks the onset of the quantum Braess effect (Barbosa et al., 2014). In a phase-coherent Y-shaped beam splitter, a coherent link between the two outgoing branches reduces total transmission, and when the outgoing leads are superconducting the differential conductance spectrum $T^{(4)}_{\mathrm{puo}}=83,$4 becomes the proposed observable of the effect (Zhitlukhina et al., 2017).

A distinct quantum-network formulation replaces roads by shared entangled states and travel time by average concurrence. In a four-node entanglement-swapping network, adding a maximally entangled Bell-state edge between the intermediate nodes creates a new route that is locally preferred but globally worse. For the explicit example $T^{(4)}_{\mathrm{puo}}=83,$5 and $T^{(4)}_{\mathrm{puo}}=83,$6, the original network has

$T^{(4)}_{\mathrm{puo}}=83,$7

while the modified network with the added Bell-state edge has the Nash-equilibrium value

$T^{(4)}_{\mathrm{puo}}=83,$8

The paradox is therefore realized as a decrease in end-to-end entanglement after the addition of a maximally entangled resource (Banerjee et al., 2021).

Open chemical reaction networks exhibit a chemical analogue in which the relevant performance quantity is the steady-state load

$T^{(4)}_{\mathrm{puo}}=83,$9

For the basic four-species scheme, the effect of an extra reaction path is measured by

$T^{(5)}_{\mathrm{puo}}=92.$0

When $T^{(5)}_{\mathrm{puo}}=92.$1, the extra path increases the steady-state load, and the corresponding parameter region is called the BP zone. The same qualitative pattern appears in networks of varying complexity, including a biochemical scheme of uric acid degradation, where some spontaneous decomposition steps are argued to have a functional role in trimming the load (Banerjee et al., 2014).

7. Coupled infrastructures and cooperative systems

Electrified transportation introduces cross-system versions of the paradox because route choice determines charging demand and charging demand feeds back into locational marginal prices and power dispatch. In a coupled graph model, the transportation social cost $T^{(5)}_{\mathrm{puo}}=92.$2, the power social cost $T^{(5)}_{\mathrm{puo}}=92.$3, and the coupled cost $T^{(5)}_{\mathrm{puo}}=92.$4 support four basic paradox types: transportation expansion can worsen transportation performance (T-T) or power performance (T-P), and power expansion can worsen transportation performance (P-T) or power performance (P-P). For general coupled systems, efficient detection is based on a generalized user equilibrium characterized as the unique optimizer of a convex quadratic program, and adaptive or static charging-price policies are proposed to mitigate the paradoxes (Mou et al., 13 Dec 2025).

This coupled-infrastructure formulation shows that the classical road-only paradox is a special case. Even when the isolated transportation or power subsystem would not exhibit the paradox on its own, the charging-mediated feedback loop can create a new indirect externality. A road widening can shift EVs toward a costly charging location and increase dispatch cost, while a transmission upgrade can change locational prices and induce traffic to move toward more congested routes (Mou et al., 13 Dec 2025).

A recent biological study shows that selfishness is not a necessary condition. In tandem-running nest relocation by Diacamma indicum, adding a bridge creates a shorter geometric route $T^{(5)}_{\mathrm{puo}}=92.$5, and leaders strongly prefer it. Yet the bridge increases congestion on the narrow segments, so the colony relocates more slowly:

$T^{(5)}_{\mathrm{puo}}=92.$6

The experiments use 16 colonies, and the exploration-exploitation model yields

$T^{(5)}_{\mathrm{puo}}=92.$7

with 95\% confidence interval

$T^{(5)}_{\mathrm{puo}}=92.$8

The reported conclusion is that evolutionary forces selecting for shortest-path identification can force suboptimal global states, so Braess’ paradox can emerge in a highly cooperative system without individual selfishness (Bairagya et al., 27 Mar 2026).

Taken together, these developments show that Braess’ paradox is best understood as a structural equilibrium effect rather than a peculiarity of urban traffic. Across traffic, supply, power, wireless, quantum, chemical, and biological settings, the recurring mechanism is that a locally attractive new path or enhanced edge reshapes global flow patterns, and the induced redistribution can worsen the relevant collective performance metric even when the added resource is intrinsically beneficial (Manik et al., 2022, Schäfer et al., 2022, Mou et al., 13 Dec 2025).