True that article ignores passengers using only tubes or trains.
But I think the article suggests global congestion may be improved by slowing down certain parts of a system.
Even considering only people using just the tube, the tube station can switch only a finite number of passengers in and out of tubes. If a system has two bottlenecks, improving the narrower bottleneck will result global improvement only until that bottleneck is no longer the narrowest.
It is possible that increasing tube throughput would increase arrival rates at the destination stations.
And if we consider congestion as nonlinear to number of passengers arriving at the station, then this might decrease the throughput at that station (beyond some threshold arrival rate); (I think this transition is similar to a phase change).
That would mean the second bottleneck would effectively becoming slightly narrower in capacity, and thus the global throughput decreases.
The trick would be to improve the narrowest bottleneck; but not too much, that it would cause such phase transitions at other points in the network. At which point, we need to start looking at other bottlenecks.
Even considering only people using just the tube, the tube station can switch only a finite number of passengers in and out of tubes. If a system has two bottlenecks, improving the narrower bottleneck will result global improvement only until that bottleneck is no longer the narrowest.
It is possible that increasing tube throughput would increase arrival rates at the destination stations. And if we consider congestion as nonlinear to number of passengers arriving at the station, then this might decrease the throughput at that station (beyond some threshold arrival rate); (I think this transition is similar to a phase change). That would mean the second bottleneck would effectively becoming slightly narrower in capacity, and thus the global throughput decreases.
The trick would be to improve the narrowest bottleneck; but not too much, that it would cause such phase transitions at other points in the network. At which point, we need to start looking at other bottlenecks.