Braess's paradox
A Socratic walk-through of Braess's paradox — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why can adding a new road to a city make every driver's journey longer?
A new road is offered to a city. Nobody is forced onto it. Every driver keeps every route they already had, and gains one more. It seems almost a matter of logic that nothing can get worse: an option you may decline cannot hurt you.
So let us hold that sentence up to the light. "An option you may decline cannot hurt you" is true of a single person choosing alone. Is it still true when four thousand people choose at the same time, and each one's choice becomes part of the conditions the others are choosing under?
Reasoning it through
REASONING #Take the smallest network that can show us anything. Four thousand drivers must get from Start to End each morning. There are two routes, and each is built from one wide segment and one narrow one.
The wide segments do not care how busy they are: a motorway with room to spare takes 45 minutes whether ten cars or four thousand use it. The narrow segments do care. Call the number of cars on a narrow segment T; it takes T/100 minutes. Empty, it is instant. With everyone on it, it takes 40 minutes.
Route one is narrow-then-wide: Start to Junction A across the narrow bit, then A to End on the motorway. Route two is the mirror image: motorway from Start to Junction B, then the narrow bit into End.
Where does traffic settle? Ask what a driver does if the split is uneven. If 2,500 have taken route one, that route costs 25 + 45 = 70 minutes while route two costs 45 + 15 = 60, so drivers peel off until the two costs match. They match at an even split: 2,000 each, and every journey takes 20 + 45 = 65 minutes. Nobody can do better by switching alone. That is what makes it a resting point rather than merely a tidy one.
Now the city builds a short, fast connector from Junction A to Junction B. Call it free — zero minutes — to make the arithmetic clean.
What does one driver think when they see it? A new route exists: narrow bit, connector, narrow bit. Notice it skips both motorway segments. Even if the narrow bits are completely full at 40 minutes each, that is 80 minutes, and that is worse than 65 — but here is the thing to watch. No individual driver is comparing the new route against the old equilibrium. Each one compares it against what the roads look like at the moment they choose, given what everyone else is doing.
So let the drivers move one at a time. The first to try the connector finds two half-empty narrow segments and sails through in well under 65 minutes. That is a real gain, and it is a gain taken partly out of other people's journeys, because their narrow segment now carries one more car. The next driver finds the same argument slightly weakened but still winning. And so on.
Where does it stop? Work out the state where everyone has switched. All 4,000 are on narrow-connector-narrow: 40 + 0 + 40 = 80 minutes. Is anyone tempted to leave? A driver who returns to the old route one pays 40 minutes on the narrow segment — still crowded, since the other 3,999 are still on it — plus 45 on the motorway: 85 minutes. Worse. The same for route two. So nobody leaves.
Sit with what that means. Every driver is doing the best available thing given what everyone else does. Every driver takes 80 minutes instead of 65. And the road that did it is a road they were free to ignore. The collective-action structure is exactly this: the sum of individually correct choices is not an individually correct outcome, and no one acting alone can undo it, because the deviation is punished even though the agreement would be rewarded.
Would they all agree to close the connector? Yes — unanimously. That is the sharpest version of the paradox. A change that every single affected person would vote against is the one they all bring about by choosing for themselves.
The analogy
THE ANALOGY #Think of a restaurant that adds one dish to its menu: superb, and quick to plate when only a few order it, but sharing a single pan with everything else. Each diner reads the menu and picks it, correctly — for them, at that moment, it is the fastest good thing on offer. The pan jams, and the whole room waits longer than it did when the dish did not exist. Ordering something else now is worse still, because the kitchen is already clogged. Everyone would vote to strike the dish; nobody can gain by declining it.
the analogy makes the dish sound like a mistake, whereas the new road really is the fastest option for whoever takes it — what spoils things is not the road's quality but that the cost each driver adds lands on segments other drivers were already depending on.
Clarifying the model
THE MODEL #Three things worth pinning down, because the paradox is easy to over-read.
It is not a claim that building roads makes traffic worse. It needs a specific setup: at least some segments whose delay grows with use, at least some whose delay does not, and drivers routing themselves rather than being assigned. Take away selfish routing — assign the flows centrally — and the 65-minute solution is available again, since the planner simply leaves the connector empty.
The damage is bounded, and the bound has a name. The price of anarchy is the ratio of the cost at selfish equilibrium to the cost of the best possible assignment. Roughgarden and Tardos showed in 2002 that when every segment's delay is a linear (affine) function of its load, that ratio can never exceed 4/3. Our example runs at 80/65, about 1.23 — close to the worst such a network can do, and by design.
And the empirical side deserves a caveat. Road closures in Seoul (the Cheonggyecheon expressway, removed 2003-2005) and in New York (42nd Street, closed for a day in 1990) are often cited as Braess in reverse, and traffic did not seize up as feared. But real cities also change trip counts, times, and modes when roads change, so disentangling Braess's specific mechanism from ordinary demand response is genuinely hard, and the attribution is not settled.
A picture of it
THE PICTURE #How to readthe two original routes each pair one crowd-sensitive segment with one flat motorway and settle at 2,000 drivers apiece for 65 minutes; the connector opens a path using both crowd-sensitive segments and neither motorway, which every driver prefers individually and which leaves all of them at 80 minutes.
What became clearer
WHAT CLEARED #An option nobody is forced to take can still make everyone worse off, because in a shared network your choice is part of my conditions. The paradox is not about roads being bad; it is about what happens when the thing each person optimises and the thing they collectively care about come apart, with no unilateral move back.
Where to go next
ONWARD #- The same structure in packet routing, power grids, and queueing systems, where the operator's counter-move is usually pricing or routing control rather than demolition.
- Congestion charging as a way of making each driver feel the delay they impose on others.
- How the 4/3 bound loosens when delays grow faster than linearly.
Key terms
TERMS #| Term | What it means |
|---|---|
| Nash equilibrium | a state in which no participant can improve their own outcome by changing strategy alone. |
| Latency function | the rule giving a segment's travel time as a function of how many use it. |
| Price of anarchy | the ratio between the cost at selfish equilibrium and the cost of the best central assignment. |
| Affine latency | a delay of the form a*x + b; the class for which the 4/3 bound holds. |
Every term the collection defines is gathered in the glossary.