Video 1, chapter 16

Braess's Paradox

How Braess's paradox lets a free shortcut slow every driver in a traffic network, explained by the Nash equilibrium of selfish routing and how it differs from the social optimum.

  • Video 1, chapter 16
  • When better goes in circles

Video 1, chapter 16

Nothing loads from YouTube until you press play.

Every morning 4000 drivers cross a small road network from the same start to the same finish, and each of them takes whichever route is quickest for them. A new one-way road opens that takes no time at all to drive, and nobody is forced to use it. Yet once the drivers have adjusted, every trip takes 80 minutes instead of 60. How can adding a free shortcut make every driver slower? This page computes the traffic before and after the shortcut, shows why the old pattern cannot last, and compares what each driver chooses with what would be best for all of them together.

Key terms

Braess’s paradox

Definition Braess’s paradox is the fact that in a traffic network where every driver chooses their own route, adding a road can make the travel time of every driver longer once the traffic has settled into equilibrium. It is named after Dietrich Braess, who described it in 1968. The standard example has a start SS, a finish FF and two junctions AA and BB, joined by four one-way roads. The roads S→AS \to A and B→FB \to F are narrow, and each takes x/100x/100 minutes when xx drivers use it. The roads A→FA \to F and S→BS \to B are wide and always take 4040 minutes. Exactly 40004000 drivers go from SS to FF. Each driver is so small a part of the traffic that one driver changing routes changes no travel time, and each driver takes a route with the shortest travel time, given the routes of everyone else. With only the upper route S→A→FS \to A \to F and the lower route S→B→FS \to B \to F, the traffic settles at 20002000 drivers on each, and every trip takes 6060 minutes. Now add a one-way shortcut A→BA \to B that takes 00 minutes, which creates the central route S→A→B→FS \to A \to B \to F. The traffic then settles with all 40004000 drivers on the central route, and every trip takes 8080 minutes. The resolution is that each driver compares routes at the current traffic and ignores the delay that their own choice adds for everyone else on the narrow roads. The old split is still possible after the shortcut opens and would still give everyone 6060 minutes, but it is no longer an equilibrium, because any one driver could then save 2020 minutes by switching. A new road does not always do this harm. In many networks, and in this one with other travel times, it shortens every trip. The example is a model with chosen numbers, and by itself it says nothing about whether a particular real road should be built or closed.

In plain words The shortcut looks like a free gift to whoever takes it, so everyone takes it. But every driver on the central route uses both narrow roads, so the narrow roads fill up until the central route is no faster than the old routes, and by then every driver is slower than before.

Example Before the shortcut, the split of 20002000 and 20002000 gives each narrow road 2000/100=202000/100 = 20 minutes, so each route takes 20+40=6020 + 40 = 60 minutes. When the shortcut opens and the old split is still in place, the central route would take

20+0+20=4020 + 0 + 20 = 40

minutes, which is 2020 minutes faster than either old route, so drivers start to switch. With 20002000 drivers on the central route and 10001000 on each outer route, each narrow road carries 30003000 drivers and takes 3030 minutes. The central route then takes 30+0+30=6030 + 0 + 30 = 60 minutes and each outer route 30+40=7030 + 40 = 70, so switching still pays. Once all 40004000 drivers use the central route, each narrow road takes 4040 minutes and the trip takes 8080. A driver who tries the upper route instead meets 40+40=8040 + 40 = 80 minutes, and the lower route also gives 40+40=8040 + 40 = 80, so no one gains by leaving. Every trip is now 2020 minutes longer than before the shortcut was built.

Traffic network

Definition A traffic network model has points called nodes, one-way roads between them called links, and a demand, a fixed amount of traffic that must travel from a start node to a finish node. A route is a path from the start to the finish that follows links in their allowed direction. The traffic is described by how many drivers take each route, with no negative numbers and with the total equal to the demand. The load on a link is the number of drivers whose route contains it. Each link has a travel time that depends only on its own load and never decreases when the load grows, and the travel time of a route is the sum of the travel times of its links at their loads. In the continuous model used here, the drivers form a flow that can be split in any proportions, so one driver alone changes no travel time.

In plain words The model records who drives where. Each road gets slower, or at least no faster, as it gets busier, and a trip takes as long as the roads it uses, added up.

Example In the network of Braess’s paradox with the shortcut, let uu, ll and cc be the numbers of drivers on the upper, lower and central routes, with u+l+c=4000u + l + c = 4000. The road S→AS \to A lies on the upper and central routes, so its load is x=u+cx = u + c, and the road B→FB \to F has load y=l+cy = l + c. The route times are

Tup=x100+40,Tlow=40+y100,Tmid=x100+y100.\begin{aligned} T_{\text{up}} &= \frac{x}{100} + 40, \\[0.4em] T_{\text{low}} &= 40 + \frac{y}{100}, \\[0.4em] T_{\text{mid}} &= \frac{x}{100} + \frac{y}{100}. \end{aligned}

For u=l=1000u = l = 1000 and c=2000c = 2000 the loads are x=y=3000x = y = 3000, which gives Tup=Tlow=70T_{\text{up}} = T_{\text{low}} = 70 and Tmid=60T_{\text{mid}} = 60. Before the shortcut is built there is no central route, which is the same as requiring c=0c = 0.

Nash equilibrium

Definition In a game, each player chooses a strategy, and each player’s result depends on the choices of all the players. A Nash equilibrium is a choice of strategy for every player such that no player can get a better result by changing only their own strategy while all the others keep theirs. In a traffic network with continuous flow, where one driver alone changes no travel time, this becomes a condition on the traffic. The traffic is in equilibrium when every route used by a positive number of drivers has the shortest travel time among all the available routes, at the current loads. A route that nobody uses may tie with the used routes or be slower, but it is never faster. This form of the condition is also called a Wardrop equilibrium.

In plain words An equilibrium is a pattern of traffic that no driver wants to leave on their own. It is stable against single drivers changing their minds, which is not a promise that it is good for anyone.

Example Before the shortcut, suppose uu drivers take the upper route and 4000−u4000 - u take the lower one. The route times are

Tup=u100+40,Tlow=40+4000−u100.\begin{aligned} T_{\text{up}} &= \frac{u}{100} + 40, \\[0.4em] T_{\text{low}} &= 40 + \frac{4000 - u}{100}. \end{aligned}

They are equal only for u=2000u = 2000, when both are 6060 minutes. A split of 25002500 and 15001500 gives 6565 minutes on the upper route and 5555 on the lower one, so a driver on the upper route can save 1010 minutes by switching, and the split is not an equilibrium. In the same way, any split with more than 20002000 drivers on one route makes that route the slower one, so the equal split is the only equilibrium.

Selfish routing

Definition Selfish routing is the model of traffic in which each driver chooses a route that minimizes their own travel time, taking the travel times produced by everyone else as given and ignoring the extra delay that their own choice adds for other drivers. The traffic it predicts is a Nash equilibrium. When the travel time of every link is a continuous function of its load that never decreases, such an equilibrium exists, and at an equilibrium all drivers with the same start and finish have the same travel time.

In plain words Every driver does the best they can for themselves, and the traffic settles where no one can do better alone. Nobody counts the minutes that their choice costs everyone else.

Example In the network with the shortcut, the upper and central routes share the road S→AS \to A, so their difference is

Tup−Tmid=40−y100.T_{\text{up}} - T_{\text{mid}} = 40 - \frac{y}{100}.

The load yy is at most 40004000, so the difference is never negative, and it is 00 only when y=l+c=4000y = l + c = 4000, which forces u=0u = 0. So whenever some drivers take the upper route, the central route is strictly faster for them. The same comparison, with Tlow−Tmid=40−x/100T_{\text{low}} - T_{\text{mid}} = 40 - x/100, applies to the lower route. Selfish routing therefore ends with all 40004000 drivers on the central route, and this is the only equilibrium of the network with the shortcut.

Social optimum

Definition A social optimum of a traffic network is a way of routing the whole demand, with no negative numbers of drivers, that makes the total travel time of all the drivers as small as possible. Dividing the total by the number of drivers gives the average travel time, so a social optimum also has the smallest possible average. A social optimum need not be a Nash equilibrium, because it can ask some drivers to take a route that is slower for them than another available route. The ratio of the total travel time at an equilibrium to the total at a social optimum is called the price of anarchy. A theorem about selfish routing in the continuous model says that when the travel time of every link has the form ax+bax + b for its load xx, with constants a≥0a \ge 0 and b≥0b \ge 0, as in Braess’s network, this ratio is never more than 4/34/3.

In plain words The social optimum is the traffic pattern that a planner who could assign every route would choose to make the average trip as short as possible. Drivers who choose for themselves need not end up there.

Example With the shortcut open, the total time of all drivers is the sum over the roads of the load times the road’s travel time. With x=u+cx = u + c, y=l+cy = l + c and u+l=4000−cu + l = 4000 - c this gives

total=x2+y2100+40(4000−c).\begin{aligned} \text{total} &= \frac{x^2 + y^2}{100} \\[0.4em] &\quad + 40(4000 - c). \end{aligned}

The loads satisfy x+y=4000+cx + y = 4000 + c. Since 2(x2+y2)−(x+y)2=(x−y)22(x^2 + y^2) - (x + y)^2 = (x - y)^2 is never negative, the first part is at least (4000+c)2/200(4000 + c)^2/200, with equality when x=yx = y, and expanding gives

total≥240000+c2200\text{total} \ge 240000 + \frac{c^2}{200}

minutes. The smallest total, 240000240000 minutes, comes from c=0c = 0 and u=l=2000u = l = 2000, an average of 6060 minutes. So the old split, with nobody on the shortcut, is the social optimum even after the shortcut is built, while selfish routing gives 8080 minutes. The ratio 80/60=4/380/60 = 4/3 is the largest that the theorem allows. A planner who sent 10001000 drivers along the shortcut and 15001500 along each outer route would give the shortcut drivers 5050 minutes and the others 6565, an average of 61.2561.25 minutes, which is worse than leaving the shortcut empty.

Check yourself

  1. Keep the same roads, but let 30003000 drivers make the trip instead of 40004000. Find the travel time before the shortcut and after it.
Show the answer

Before the shortcut, the route times are equal only when 15001500 drivers take each route. Then each narrow road takes 1515 minutes and each trip 15+40=5515 + 40 = 55 minutes, and as before this is the only equilibrium.

After the shortcut opens, the comparison from selfish routing becomes

Tup−Tmid=40−y100T_{\text{up}} - T_{\text{mid}} = 40 - \frac{y}{100}

with yy at most 30003000, so the difference is at least 1010 minutes and the upper route is always slower than the central one. The same holds for the lower route, so all 30003000 drivers take the central route. Each narrow road then takes 3030 minutes and every trip takes 30+0+30=6030 + 0 + 30 = 60 minutes, while either outer route would take 30+40=7030 + 40 = 70. The shortcut again makes every trip longer, here by 55 minutes.

  1. Change the wide roads so that each takes 7070 minutes instead of 4040, with 40004000 drivers and the same narrow roads. Does the shortcut still slow everyone down?
Show the answer

No. Before the shortcut, the equal split gives every trip 20+70=9020 + 70 = 90 minutes. After it opens, the difference between the upper and central routes is

Tup−Tmid=70−y100,T_{\text{up}} - T_{\text{mid}} = 70 - \frac{y}{100},

which is at least 70−40=3070 - 40 = 30 minutes, and the same holds for the lower route. So the outer routes are always slower and every driver takes the central route, which takes 40+0+40=8040 + 0 + 40 = 80 minutes, while an outer route would take 40+70=11040 + 70 = 110. The shortcut now saves every driver 1010 minutes.

The paradox depends on the numbers. With these narrow roads and 40004000 drivers, the same kind of calculation shows that the shortcut makes the trips longer exactly when each wide road takes more than 2020 and less than 6060 minutes.

  1. Let each wide road take 2525 minutes, with 40004000 drivers and the same narrow roads. Before the shortcut every trip takes 20+25=4520 + 25 = 45 minutes. Find the equilibrium after the shortcut opens.
Show the answer

Leaving the shortcut empty is not an equilibrium, because with 20002000 drivers on each narrow road the central route would take 20+0+20=4020 + 0 + 20 = 40 minutes, less than 4545. Sending everyone along the shortcut is not one either, because then the central route takes 40+40=8040 + 40 = 80 minutes, while the upper route would take 40+25=6540 + 25 = 65. So some drivers use the central route and some use an outer one.

A used outer route must tie with the central route. The upper and central routes tie when 25=y/10025 = y/100, so y=2500y = 2500, and the lower and central routes tie when x=2500x = 2500. Both outer routes must be used. If, say, nobody took the lower route, the upper route would be used, so y=2500y = 2500, while x=u+c=4000x = u + c = 4000. The central route would then take 40+25=6540 + 25 = 65 minutes and the unused lower route 25+25=5025 + 25 = 50, which is faster. The same argument with the two outer routes swapped rules out leaving the upper route empty. So x=y=2500x = y = 2500, and from

x+y=u+l+2c=4000+c\begin{aligned} x + y &= u + l + 2c \\[0.4em] &= 4000 + c \end{aligned}

the shortcut carries c=1000c = 1000 drivers and each outer route 15001500. Every route then takes 25+25=5025 + 25 = 50 minutes. Some drivers use the shortcut and some do not, and all of them are 55 minutes slower than before it opened.