Costing Northeast Corridor High-Speed Rail

As our high-speed rail project draws to a close, we need to not just write down what is needed for running the trains but also how much it costs. This post should be viewed as a work in progress, and it will not surprise me if I’m missing things that will make it to the report later this year.

The rule for this post is that costs only matter going forward, not backward. If it’s already committed, it’s not part of the budget; in particular, the $6 billion Frederick Douglass Tunnel, already fully funded and in the design and engineering phase, is not part of the budget. In addition, only infrastructure is costed, not rolling stock (new rolling stock may well have negative cost relative to current plans, through buying standard EMUs and not esoteric trainsets like Massachusetts’ battery train idea or nonstandard LIRR/Metro-North-style EMUs).

Bypasses

All bypasses can be seen on this map, but not all bypasses are part of the plan – in particular, nothing between Stamford and New Haven seems worth it for now.

The main bypass we’re proposing, between New Haven and Kingston, is 120 km in relatively easy terrain, including two constrained river bridges (Quinnipiac and Thames; the Connecticut is easier), but no tunnels. The cost should be in line with non-tunneled high-speed lines in Europe, which in 2024 dollars would be around $5 billion.

The secondary bypass, around Port Chester and Greenwich, is 7 km of complex els crossing I-95 multiple times, and should be costed at the upper end of els, which is high hundreds of millions. Call it $1 billion together with a new bridge across the Mianus. The current projected cost for the Cos Cob Bridge replacement is higher, but it should be easier to rebuild the bridge a bit upstream to straighten the approach curves than to do it in situ; with a short section of 4% grades on each side, it should be possible to clear I-95 west of the river and keep the Riverside station east of the river while also having around 23 meters of clearance below the bridge. (4% grades are routine for EMUs; freight trains are so long that they can ascend these grades just fine, since what matters is the grade averaged over the length of the train.)

Frankford Junction is about 2 km of complex urban el, including a rail-on-rail grade separation; the per km cost is likely high, in the very low three-figure millions, but it’s 2 km and so $300 million should cover it.

The other bypasses are very short and in easy environments, for example easing the curve at Kingston (also discussed here), with costs dominated by the track connections rather than the physical construction of 1-2 km of at-grade track outside urban areas. Call this entire portion $6.5 billion total.

Grade separations

The starting point is that NJ Transit thinks that Hunter Flyover should be $300 million in 2022 prices (source, PDF-p. 151). This is as close as can be to a nonnegotiable element of the program.

At the other end of the New York metro area, there’s Shell Interlocking/CP 216, which must be grade-separated as well, and is even more nonnegotiable. I have not seen recent cost figures; it should be comparable to Hunter or somewhat more expensive given the right-of-way constraints. A $500 million placeholder is probably right.

Further north, the junction with the New Canaan Branch is flat and needs to be grade-separated, at a cost likely similar to Hunter, in a similarly built-up area. The Danbury and Waterbury Branches have flat junctions too, but traffic is low enough that they may be kept so (especially Waterbury), but if not, Danbury seems comparable in difficulty to Hunter and New Canaan.

In Philadelphia, the Chestnut Hill West Line (former R8) has a flat junction with the Northeast Corridor, and there are a variety of proposals for what to do with it; for decades, an advocate wish was the Swampoodle Connection, to have it transition to a closely parallel line letting it enter the city via the Reading side rather than the former Pennsylvania Railroad side that it’s on. It’s largely dropped off the wishlist, and instead a grade separation could be done for a cost comparable to that of Hunter, or maybe less (potentially much less) if it’s possible to abuse the line’s low ridership and close proximity to the Chestnut Hill East Line to have shutdowns to speed up the work.

On the other side of Philadelphia, the junction between the intercity and commuter rail approaches to 30th Street is flat as well, which also incorporates the branch to Media/Elwyn (former R3); this should be grade-separated as well.

In Boston, there are two flat junctions on the Providence Line. Canton Junction separates it from the Stoughton Line, and looks routine to either grade-separate (it’s a low-density area) or, potentially, even turned into a shuttle with timed connections to the Providence Line if absolutely necessary, given the demand mismatch between the two branches. The Franklin Line, farther north, has a similar flat junction around Readville, technically within Boston but in an area with plenty of space, but can be sent over to the Fairmount Line if there are difficulties, and may even preferentially go to Fairmount regardless (the main argument against it is service to Back Bay). The answer to “how much should this cost?” is “no more than around $150 million each or else it’s better not to do it at all.”

In total, these should be around $1.8 billion, with New Canaan and Canton but not Danbury or Readville.

Note that rail-on-rail grade separations for bypasses are already priced in, especially New Haven-Kingston, which is of comparable length to European high-speed lines that have been built, with grade-separated connections to legacy lines.

Portal Bridge

The Hudson Tunnel Project within the Gateway Program is funded, but some tie-ins are not. Most (such as Penn Expansion) are useless, but one is essential: a second Portal Bridge, to ensure four tracks of capacity from New York to Newark. The current favored alternative is a lift bridge, budgeted at $800 million; it is a movable and not fixed bridge, but it is not a causeway and has some clearance below, and would only need to open when a sludge barge comes from upriver, which can be scheduled overnight.

High platforms

Everything that touches the Northeast Corridor needs high platforms at all stations. The definition of “touches the Northeast Corridor” is complicated; for example, in New Jersey, there are 68 low-platform stations on the lines that go through Newark Penn or Newark Broad Street, of which 26 are funded for high-platform conversions for around $23 million each ($683 million/30 stations; the other four are on the Erie lines), but of the 68, only 10 are on the lines that would be using the North River Tunnels after the Hudson Tunnel Project opens (see map in ETA’s report). Even taking all 42 as required, it’s around $1 billion at NJ Transit costs, with nearly all benefits accruing to commuter lines.

In Massachusetts, the definition is easier – everything on the Providence and Stoughton Lines needs to be raised; the TransitMatters report explains that there are eight stations, plus two potential infills, with the eight costing around $200 million in 2020 prices, which should be closer to $250 million in 2024 prices. If Franklin Line work is also desired then it should be another $200 million, split across more stations but with shorter platforms. Note that the second phase of South Coast Rail, if it is built, would extend the Stoughton Line, but as the stations are all new construction, they will already have high platforms.

In Pennsylvania, nearly total separation of intercity traffic from SEPTA is possible from the get-go – the only track sharing is peripheral, in and around Wilmington, at low frequency on SEPTA. If the entire Wilmington/Newark Line is to be upgraded, it’s a total of 12 stations, all in four-track territory; SEPTA’s construction costs for high platforms are lower than those of the MBTA and NJ Transit, but much of its construction has been single-platform stations with shorter trains, and my guess is that those 12 stations are around $200 million total. The seven inaccessible stations on the Trenton Line, which, to be clear, does not need to share tracks with intercity trains at all, should be another $100-150 million (it’s a busier line, so, longer trains, and North Philadelphia is more complex).

In Maryland, two stations on the Penn Line are inaccessible, West Baltimore and Martin State Airport. West Baltimore is being upgraded as part of the Douglass Tunnel program, while Martin State Airport has a separate program, which appears funded.

In total, all of this is around $1.8 billion, with the benefits going to commuters at such rate that state matches would be expected; in Massachusetts at least, there are talks about doing it as part of the Regional Rail program, but no firm commitment.

Electrification

The variable-tension catenary south of New York, as users of the Northeast Corridor were reminded two months ago, is substandard. It’s long been a wish to replace it with constant-tension catenary, to both improve reliability and permit unrestricted speeds, up from today’s 135 mph (217 km/h).

Unfortunately, precisely because it’s a longstanding Amtrak project, the project definitions have been written in a way that is not compatible with any cost-effective construction. For example, Amtrak is under the impression that the catenary poles have to be redone because higher speeds require denser pole spacing; in fact, catenary systems sold routinely by European vendors allow high speeds at spacing that exists already on the legacy Northeast Corridor system.

This makes costing this more difficult; Amtrak’s official figures are of little relevance to a project that has even cursory levels of interest in adopting European practices. With the poles and substations already usable, the wire tensioning should cost less than installing new wires; around half of the cost of new-build electrification is the substations and transformers and the other half is the wires, so take the cost of new-build systems outside the US and Canada, cut in half, and then double back to take into account that it’s a four-track corridor. This is around $3 million/km, so around $1 billion corridor-wide.

Commuter rail lines that touch the Northeast Corridor need to be wired as well, and then it’s a matter of which ones count as touching, as with the high platform item. This includes 25 km of the North Jersey Coast Line, 72 km of the Raritan Valley Line, 31 km of the Morristown Line, 30 km of the Montclair-Boonton Line, 38 km of the Danbury Branch, a few hundred meters of Providence Line siding tracks, 6 km of the Stoughton Line, 34 km of the Franklin Line, and 15 km of the Fairmount Line. Much of the unwired territory is single-track, so lower per-km costs can be expected, on the order of $600 million total.

Together, this is about $1.6 billion.

Total

The sum of all of the above lines is $12.5 billion. It’s possible to go lower than this: the high platform and electrification costs are partly modernizing commuter rail that may not quite use the Northeast Corridor, and the Greenwich bypass may be dropped at the cost of 80 seconds (more, if Cos Cob Bridge speed limits have to be lower than what right-of-way geometry allows). A numerological $10 billion limit can still be met this way.

The Northeast Corridor Should not Host Diesel Through-Service

The vast majority of traffic on the Northeast Corridor comprises captive trains, only running internally to the corridor. However, a noticeable minority of trains run south of Washington, swapping the electric locomotive for a diesel locomotive. Those trains have a certain logic to them today – through-service is valuable, and north of Washington they more or less substitute for Northeast Regional service. But drawbacks to reliability remain, and if the corridor modernizes its operations, they will need to be removed; several elements of modern operations are not compatible with keeping either the Virginia trains or the long-distance ones on the corridor. Instead, these trains should be cut to Washington, with transfers to much faster, more frequent Northeast Corridor trains. Potentially, some Virginia lines could be electrified and then through-service could be offered, if they can fill the same size of train that future Northeast Corridor service could.

Fortunately, this tradeoff still leaves the South with better service than it gets today. The forced transfer considerably speeds up the trip for New York-bound passengers, by more than the average transfer penalty even for passengers with heavy luggage. Nonetheless, a tradition of direct through-service from New York to diesel territory in the South will need to end.

Future Northeast Corridor service

Upgraded service, for example in our ongoing project at Marron for how to blend intercity and commuter rail on the corridor, should have all of the following features:

  • High speed: our current timetables have New York-Washington trains taking 1:53, at a top speed of 320 km/h; a blanket speed restriction to 217 km/h, the upper limit of the catenary today, would only slow down the trains to 2:04, the rest of the difference in speed from current trip times coming from reliability improvements, higher curve speed, higher acceleration, and minor curve fixes.
  • All-EMU configuration: nearly the entire world passenger rail market is electric multiple units rather than separated locomotives and coaches, EMUs outperforming locomotive-hauled cars in every aspect, and the exceptions are the less modern intercity and regional lines.
  • Single-class service: trains may have first- and second-class seats, but the trains should not be differentiated by speed – Spain has trains of different speeds on its high-speed line and charges more for the faster ones, and the resulting hit to frequency and interchangeability of tickets explains why Europe’s longest high-speed rail network has a fraction of the per capita ridership of France, Germany, or Japan, which (largely) lack this misfeature.
  • Affordable fares: the average fare should be in line with French and German norms, around $0.15 per p-km.

A train that does New York-Washington for a bit less than two hours and charges a bit more than $50 one-way on average – the current average is $106 on the Regional and $192 on the Acela – can comfortably expect ridership to quadruple, based on my two usual references on elasticity of high-speed rail ridership with respect to travel time and fares (Börjesson says -1.12 and -0.61 respectively, Cascetta-Coppola say -2 and -0.37 respectively). This forces running more frequency and longer trains. Frequency is a welcome addition provided there is capacity on the tracks for it; fortunately, there is capacity for an intercity train every 10 minutes in a post-Gateway timetable. The trains in question should be as long as possible, with platform lengthening where necessary to support 16-car trains, to maximize capacity.

The incompatibility of diesel trains

Very few diesel trains run on the Northeast Corridor today, and none run by Amtrak. The through-trains to the South run with engine changes: all Amtrak service today is run with locomotives, and at Washington, the trains change between diesel and electric locomotives. Nonetheless, even electric locomotive service as it is conceived today is incompatible with Northeast Corridor modernization, and future changes would still not make it compatible.

First, to the point on capacity: there is no way to run 16-car trains into the South. There isn’t enough demand for such trains. The Silver Star today runs nine coaches and the Silver Meteor runs 10, and on both trains, three coaches don’t sell seats but are used for baggage, lounge, and dining; the Palmetto runs six coaches, the Crescent seven, and the Cardinal five, each including two non-seat selling coaches. Speeding up the Northeast Corridor by an hour and a half can lead to ridership explosion internally to the corridor, but not on trains that take 30 hours today.

And second, there is no reasonable rolling stock for this, even if there were demand for a train with 16 cars or close to it. Locomotive-hauled trains would necessarily run slow, compromising not just top speed but also acceleration and, owing to the current equipment’s problems, reliability. The example train we’re using in our calculations, the Velaro Novo, has a power-to-weight ratio of 20 kW/t and an initial acceleration rate of 0.65 m/s^2. A pair of ACS-64 locomotives dragging 14 Airo coaches gets 14 kW/t but cannot accelerate faster than around 0.25 m/s^2 at any speed. The unpadded trip time for high-speed rail making one stop per state is 1:46; the unpadded trip time with the additional acceleration time of this example train and with a 217 km/h (135 mph) speed restriction is 2:09. If the timetable buffer time is still 7% then the trip time is 2:18, which means the train would be overtaken by about two faster trains, and this in turn would slow the trains further as more schedule contingency would be required for this more complex system. If the ACS-64’s problems or any interface with the freight-run Southern network forces more padding, then make it three overtakes.

The TGV used to couple a diesel locomotive in front of a high-speed trainset to reach Les Sables d’Olonne, before the branch to it was electrified. This option would eliminate the speed difference on the Northeast Corridor, but would also mean that expensive 16-car high-speed trainsets would be spending an entire day going to Florida at low speed and another going back, without being able to make back the cost through intensive operations measured in train-km per day.

Exceptions and the tail wagging the dog

The basic reason for prioritizing the Northeast Corridor’s internal traffic over through-traffic is the large mismatch in travel volumes. In fiscal 2023, the Northeast Corridor got 12,122,466 riders. The Virginia services got 1,300,776, the Carolinian 315,781, and the long-distance trains 1,308,211. A 4:1 ratio should tilt planners toward prioritizing the core over the long-distance trains.

Note that I have not, up until now, talked about Keystone service and trains to Springfield. This is because Keystone trains can run through to the corridor just fine. None of the reasons why the long-distance trains cannot do so applies: the Keystone corridor is electrified all the way to Harrisburg, and New York-Philadelphia is a significant enough portion of it that boosting speeds in the core (and acceleration everywhere) would lead to sufficient ridership increases to justify 16-car trains. Springfield service is currently unelectrified, and Amtrak generally runs shuttles with timed connections because of the mismatch in demand; it should be electrified, and through-service instituted.

On Keystone and the New Haven-Springfield line, the mismatch in capacity actually works in favor of through-service. The New York-Philadelphia section has the most demand, so having one third of the trains branch off to Harrisburg rather than continuing to Washington is a good way of assigning capacity. New Haven is not Philadelphia, but has so much commuter demand to New York that giving New York-New Haven an intercity train every 10 minutes is not so stupid; in contrast, unless a lot more is built, I suspect that 16-car trains running every 10 minutes between New York and Boston would end up emptier than most planners would prefer. Years ago, before I started looking at the track charts and the possible schedules systematically, I even used the greater demand of New York-Philadelphia to argue in favor of diverting some trains not just at New Haven to Springfield, but at Penn Station, to Jamaica and Long Island; as it is, my primary argument against sending intercity trains to the LIRR is timetabling complexity.

With Keystone and Springfield added back in, the traffic on the Northeast Corridor rises from 12,122,466 to 13,680,273. The ridership of the trains to the South that are to be cut from the corridor is 2,924,768, or 21% of the internal ridership; the tail should not wag the dog.

Is this even bad for the trains to be cut?

No. As mentioned above, the trip times would get a lot faster, it’s just that turning a 30-hour trip into a 28-hour one does not lead to a large ridership boom.

The extra transfer is annoying, but should be compared with the time cost of both running a slower train to New York and changing the engine at Washington Union Station. As explained above, the slower train would take a minimum of 2:09 between New York and Washington, stopping once per state. The scheduled time would be at least 2:18, and likely more, maybe 2:28 with 15% buffer time. The engine change takes about 20 minutes judging by southbound schedules on Virginia service trains; the wait time at Union Station is much longer northbound, because the train has to have more schedule contingency on the less reliable freight-owned section to make its slot on the more precisely timetabled Northeast Corridor. The most charitable interpretation, ignoring the extra required schedule padding, is that making passengers change trains in Washington would save them 45 minutes minus the wait time for the next train (at most 15 minutes).

The transfer penalty is extensively studied in the modal choice literature. For example, studying intercity trips in the Netherlands, de Keizer-Kouwenhoven-Hofker find that the penalty is 23 minutes, which already incorporates an imputed additional waiting time of 15 minutes. This penalty rises by seven minutes if the transfer is not cross-platform; a cross-platform transfer at Union Station would require the through-tracks to be upgraded with high platforms, as they are currently low-platform. It rises by a further seven minutes for passengers with heavy luggage. Even with all of those penalties, 23+7+7 = 37 < 45. And 45 is in a way a best-case scenario; there is a lot of padding involved in making a long-distance or even Virginia train make a specific slot on the corridor, as opposed to guaranteeing passengers a seat on the next available train, and this adds on the order of half an hour, counting both Alexandria-Washington and on-corridor buffer times.

The upshot is that while trains cannot run through from Virginia to a modernized Northeast Corridor, little is lost in making passengers transfer at Union Station. The transfer penalty is real but limited, even with luggage, and the speed gain from letting such passengers transfer to a faster train is noticeable, if small compared with the total length of a night train trip. It would break tradition, but offer a modest improvement in the quality of rail service on the long-distance trains using the corridor and the Virginia trains, in conjunction with the much larger improvement in the quality of internal Northeast Corridor service.

More on American Incuriosity, New York Regional Rail Edition, Part 2: Station Dwell Times

This is the second part of my series about the Regional Plan Association event about expanding capacity at Penn Station. Much of the presentation, at least in its first half, betrays wanton ignorance, with which area power brokers derive their belief that it is necessary to dig up an entire block south of Penn Station to add more station tracks, at a cost of $16.7 billion; one railroad source called the people insisting on Penn Expansion “hostage takers.” The first part covered casual ignorance about the history of commuter rail through-running in Europe, including cities that appear in the presentation. This part goes over the core claim made in the presentation regarding how fast trains can enter and exit Penn Station. More broadly, it goes over a core claim made in the source the presentation uses to derive its conclusion, a yet-unreleased consultant report detailing just how much space each train needs at Penn Station, getting it wrong by a factor of 5-10.

The issue is about the minimum time a train needs to berth at a station, called the dwell time. Dwell times vary by train type, service type, and peak traffic. Subways and nearly all commuter trains can keep to a dwell time of 30 seconds, with very few exceptions. City center stations like Penn Station are these exceptions; the RER and the Zurich S-Bahn both struggle with city center dwell times. The Berlin S-Bahn does not, but this is an artifact of Berlin’s atypically platykurtic job density, which isn’t reproducible in any American city. That said, even with very high turnover of passengers at central train stations, the dwell time is still usually measured in tens of seconds, and not minutes. In the limiting case, an American commuter train should be able to dump its entire load of passengers at one station in around two minutes.

The common belief among New York-area railroads is that Penn Station requires very long dwell times. This is not made explicit in the presentation; Foster Nichols’ otherwise sober part of the presentation alludes to “varying dwell times” on pp. 23 and 26, but documents produced by the railroads about their own perceived needs go back years and state precise times; for through-running, it was agreed that the dwell times would be set at 12 minutes in the Tri-Venture Council comprising Amtrak, the LIRR, and New Jersey Transit. The consultant report I reference below even thinks it takes 16 minutes. In truth, the number is closer to 2-3 minutes, and investments that would precede Penn Expansion, like Penn Reconstruction, would be guaranteed to reduce it below 2 minutes.

Dwell times in practice

Before going into what dwell times should be, it is important to sanity-check everything by looking at dwell times as they are. It is fortunate that examples of short dwell times abound.

As mentioned in my previous post, I have just returned from a trip to Brussels and London. My train going out of Berlin was late, so at Hauptbahnhof, the dwell time was just three minutes. The train, which had departed Ostbahnhof almost empty, filled almost to seated capacity at Hauptbahnhof, where there is no level boarding. DB routinely turns trains in four minutes at terminal stations that are located mid-line, like Frankfurt and Leipzig, but this time I observed such dwells at a station with almost complete seat turnover. In Japan, where there is level boarding and two door pairs per car rather than one, the dwell times on the Nozomi are a minute, even at Shin-Osaka, where through-trains transition from JR Central to JR West operation.

On commuter rail, dwell times are shorter, even though the trains are much more crowded at rush hour. The reason is a combination of higher toleration for standees, and higher toleration of mistakes – if passengers get on the wrong train or miss their stop, they will get off at the next stop in a few minutes rather than ending up in the wrong city.

As mentioned in the introduction, Penn Station is a limiting case on commuter rail, since it’s the only station in Manhattan for any possible through-trains today; a future tunnel to Grand Central, studied over 20 years ago as Alternative G and recurrently proposed since in various forms (for example, in the ETA writeup, or in this post of mine from last year), would still leave trains that use the preexisting North River Tunnels running through the East River Tunnels and not making a second Manhattan stop. Thus, the best comparison cases need to be themselves limiting cases, as far as possible.

For this, we need to go to Paris, especially its busiest lines, the RER A and B. The RER B has two central stations: Gare du Nord, Les Halles; Gare du Nord isn’t really in the central business district, but is such a large travel hub that its RER and Métro traffic levels are the highest in both systems. The theoretical dwell time (“stationnement”) is 30 seconds on the RER. In practice, at rush hour, it’s higher – but it’s still measured in tens of seconds. In the 2000s, the RER B reached 70-80 second dwell times at Gare du Nord at peak, before new work reduced the average to 55 seconds. I timed dwell times while living in Paris and riding the RER B regularly to IHES, and at rush hour, the two central stations and Saint-Michel-Notre-Dame were usually 50-60 seconds. This is optimized through signaling as well as wide platforms and single-level trains with four door pairs per car, though the internal configuration of the corridor of the RER B rolling stock still leaves something to be desired, especially if there are passengers with luggage (which there often are, as the line serves CDG Airport).

The RER A has four central business district stations: Les Halles, Auber, Etoile, La Défense; a fifth station, Gare de Lyon, is like Gare du Nord a transport hub with very high originating ridership. A report from the early 2010s lamenting that the theoretical throughput of 30 trains per hour was not achieved in practice blames a host of factors, including high dwell times due to traffic, reaching 50 seconds in the central section. The RER A rolling stock is bilevel with three triple-wide door pairs per car, and for a bilevel its internal circulation is good, but it’s still a bilevel train, and getting through a crowded rush hour car to disembark takes a lot of shuffling.

Is Paris a good comparison case?

Yes.

Part 1 of this series goes over the history of the RER, and points out that in 2019, the RER A had 1.4 million weekday trips, and the RER B 983,000. This compares with a combined LIRR and New Jersey Transit ridership of about 600,000 per weekday. About 67% of LIRR ridership is at rush hour; on SNCF-operated Transilien and RER lines, at the suburban stations, the figure is 46%, and my suspicion is that the RER B is somewhat lower than Transilien.

The higher peakiness in New York evens things up somewhat. But even then, peak hourly traffic into Penn Station from New Jersey was 27,223 passengers in 2019, per the Hub Bound report (Appendix III, Section C), and peak hourly traffic from the four-track East River Tunnels was 33,530; in contrast, the RER A’s peak hourly traffic last decade was 50,000.

Now, Paris does have multiple central stations, whereas there is only one in Manhattan on the LIRR and NJ Transit. That said, this only evens things up. My table on this only includes the SNCF-operated portion, and only includes boardings at a resolution of four hours, not one hour; thus, all central RER A stations are missing. From the table, we get the following maximum boarding counts between 4 and 8 pm and between 6 and 10 am on a work day:

StationLineTrains/hourBoardings (pm)Boardings (am)
Penn StationLIRR3773,4304,920
Penn StationNJ Transit2056,6647,838
Gare du NordRER B (both directions)2048,98954,137
Gare du NordRER D (both directions)1234,51228,073
Châtelet-Les HallesRER D (both directions)1228,5866,877
Gare de LyonRER D (both directions)1249,39217,158
Haussmann-Saint LazareRER E1645,38310,719

The numbers represent single-line trips, so people transferring cross-platform between the RER B and D at Gare du Nord count as boardings. The reason for including both morning and afternoon peak traffic is that afternoon boardings are largely symmetric with morning alightings and vice versa, and so the sum represents total on and off traffic on the train at the peak.

Peak traffic per train in a single direction occurs at Saint-Lazare on the RER E, which only began through-running in May of this year; the counts are from the mid-2010s, when the station was a four-track underground terminal. At the through-stations, total ons and offs per rush hour train are slightly lower than at Penn Station on NJ Transit and slightly higher than on the LIRR. Even taking into account that at Penn Station, 40% of the peak four hour traffic is at the peak hour, and the proportion should be somewhat smaller in Paris, the difference cannot be large. If Gare du Nord can support 60 second dwell times, Penn Station can support dwell times that are not much higher, at least as far as the train-platform interface is concerned.

Gantt charts

A yet unreleased consultant report for the Penn Station Capacity Improvement Project (PCIP) details the tasks that need to be done for a through-running train at Penn Station. This is shown as a pair of Gantt charts, both for a future baseline, the second one assuming dropback crews and station scheduling guaranteeing that trains do not berth on two tracks facing the same platform at the same time. All of this is extravagant and unnecessary, and could not be done by people who are familiar with best practices in Europe or Japan.

This is said to be turn time in the chart and dwell time in the description. But the limiting factor is the passenger path and not the crew path, and for that, it doesn’t matter if a train from New Jersey then goes to Long Island or Stamford and a train from Long Island or Stamford goes to New Jersey or if it’s the other way around.

To be clear, 16 minutes is insanely long as an unpadded turn time, let alone a through-dwell time. The MBTA can do it in 10; I think so can Metro-North at the outer ends. ICE trains turn in four minutes at pinch points like Frankfurt Hauptbahnhof, with extensive rail passenger turnover. So let’s go over how to get from 16 down to a more reasonable number.

Passenger alighting

Alighting does not take 6.5 minutes at Penn Station, even at rush hour, even on trains that are configured for maximum seats rather than fast egress. The limiting factor is not the train doors – the RER D runs bilevels with two door pairs per car and narrow passageways, and would not be too out of place on NJ Transit. Rather, it’s the narrow platforms, which have fewer egress points than they should and poor sight lines. This was studied for the Moynihan Station project, which opened in 2021. The project added new staircases and escalators, and now the minimum clearance time is at most 2.03 minutes, on platform 9, followed by 2.02 minutes on platforms 4 and 5. The expected clearance time, taking into account that passengers prefer to exit near the 7th Avenue end but the egress points are not weighted toward that end, peaks at 4.83 minutes on platform 4 – but passengers can walk along the platform while the train is moving, just as they do on the subway or on the RER.

What’s more, Penn Reconstruction, a project that may or may not happen, but that is sequentially prior to the Penn Expansion project that the slide deck is trying to sell, is required to install additional vertical circulation at all platforms, to reduce the egress times below 2 minutes even in emergency conditions (one escalator out). This is because NFPA 130 requires evacuation in 4 minutes assuming every track that can be occupied is, which given timetabling constraints means both tracks facing each platform other than the single-track platform 9. Responding to Christine Berthet’s questions about through-running, the agency even said that Penn Reconstruction is going to bring all platforms into compliance, but still said dwell times would need to be 8 minutes.

Passenger boarding

Alighting and boarding peak at different times of day. As the above table shows, reverse-peak traffic at Penn Station is only 12% of the combined peak and reverse-peak traffic on NJ Transit, and only 6% on the LIRR. In any circumstance in which the alighting time needs to be stretched to the maximum (again, only somewhat more than 2 minutes), the boarding time can be set at 30 seconds, and vice versa.

Moreover, because the access points to the platforms include escalators, not all running in the peak direction, and not just staircases, reverse-peak traffic consumes capacity that is otherwise wasted. Even the 30 seconds for additional boarding time in the morning rush are generous.

Conductor walk time for safety review

This is not done in Europe. Conductors’ safety review comprises checking whether passengers are stuck in the gap between the platform and the train, which is done after boarding, and takes seconds rather than minutes, using CCTV if the sight lines are obstructed.

Door opening and closing

These do not take 30 seconds each; the total amount of time is in the single digits.

Engineer operating position set-up, and engineer/conductor job briefing

Crews switch out in 1-2 minutes at boundaries between train operating companies in Paris and Shin-Osaka. The RER B is operated by SNCF north of Gare du Nord and by RATP south of it, and they used to switch crews there – and the operating position had to be changed, since the two companies’ engineers preferred different setups, one preferring to sit and the other to stand. It took until the early 2010s to run crews through, and even then it took a few years to unify the line’s dispatching. It does not take 3 minutes to brief the engineer on the job.

Total combined time

On a through-train, using alighting times in line with the current infrastructure at Penn Station, the minimum dwell time is 2-3 minutes, provided trains can be timetabled so that no two tracks facing the same platform have a train present at the same time. If there are four through-platforms, then commuter trains can run every 5 minutes to each platform, which is borderline from the perspective of egress capacity at 7th Avenue but does work.

Intercity trains make this easier to timetable: they have lower maximum capacity unless standing tickets are sold, which they currently are not, and even if Amtrak runs 16-car EMUs, they’ll still have fewer seats than there are seats plus standing spaces on a 10-car NJ Transit train, and not all of them turn over at Penn Station. Potentially, platform 6 can be dedicated to intercity trains in both directions, and then platforms 4 and 5 can run eastbound, alternating, and platforms 7 and 8 can run westbound. Using the timetable string diagram here, the local NJ Transit trains on the Northeast Corridor would have to share a platform, running every 5 minutes, while the express trains can get a dedicated platform running every 10; the local trains are likely to be less crowded and also have more through-passengers, first because usually through-service is more popular in inner suburbs than in outer ones, and second because the likely pairing in our Northeast Corridor plan connects those trains to Long Island City and Flushing while the express trains awkwardly turn into local Metro-North trains to Stamford.

Note that intercity trains can be scheduled to dwell for just 2-3 minutes too, and not just commuter trains. That’s actually longer than Shinkansen express dwell times (involving a crew change at Shin-Osaka), and in line with what I’ve seen with full turnover in Berlin. The Avelia Liberty has better circulation than the ICE 3, since it has level boarding, and any future trainset can be procured with two door pairs per car, like the Velaro Novo or Shinkansen, rather than just one, if dwell times are a concern.

The incuriosity of consultant-driven projects

I spoke to some of the people involved about my problems with the presentation, and got very good questions. One of them pointed out that I am talking about two- and three-minute dwell times in big European cities, and asked, how come experienced international consultants like Arup and LTK, which prepared the Gantt chart above, don’t know this? What’s missing here?

This is a question I’ve had to face with the construction cost comparisons before, and the answer is the same: consultants are familiar with projects that use consultants. Anglo consultants like Jacobs, AECOM, Arup, and WSP have extensive international experience, with the sort of projects that bring in international consulting firms to supervise the designs. The bigger Continental European and East Asian countries have enough in-house engineering expertise that they don’t really bring them in.

This can be readily seen in two ways. First, getting any detailed information about rail projects in France and Germany requires reading the local language. Practically nothing gets translated into English. I almost exclusively use French sources when writing about the RER, which can be readily seen in this post and in part 1. My German is a lot less fluent than my French, but here too I have to rely on reading technical German to be able to say anything about the Berlin or Munich S-Bahn or the ICE at greater depth than English Wikipedia (for one example, compare English and German on switches). A lot of the information isn’t even online and is in railfan books and magazines. This is not an especially globalized industry, and a consultancy that works in English will just not see things that are common knowledge to the experts in France or Germany, let alone Japan.

And second, the few Continental European projects that are more globalized turn into small reference pools for American agencies looking to compare themselves to others. Woody Allen portrays a Barcelona with the works of the only architect his American audience will have heard of. The MTA compares its per-rider costs to those of the not-fully-open Barcelona Metro L9/10, MassDOT uses L9/10 to benchmark the North-South Rail Link (again with the wrong denominator), and VTA uses L9/10 as a crutch with which to justify its decision to build a single-bore San Jose subway. L9/10 is an atypically large project, and atypically expensive for Spain; it also, uniquely, uses more privatization of planning than is the norm in Spain, including design-build project delivery, whence the line from the one of the consultants I’ve had to deal with in the US, “The standard approach to construction in most of Europe outside Russia is design-build” (design-build to a good approximation does not exist in Germany, Spain except L9/10, or Italy, and is uncommon in France and done with less privatization of expertise than in the US).

To take these two points together, then, the elements of foreign systems that are likeliest to be familiar to either American railroaders or English-primary consultants are the biggest and flashiest ones. This can even include elements that are not consultant-driven, if they’re so out there that they can’t be missed, like a high-speed rail network: rail consultants know the TGV exists, even if they’re not as familiar with how SNCF goes around planning and building lines, and can sometimes imitate design standards. Commuter rail infrastructure that’s similarly flashy gets noticed, so the presentation mentions the RER and Munich S-Bahn, even while getting their histories wrong and fixating on the new station caverns that even a tourist on a short trip can notice.

Commuter rail operations are not flashy. The map of RER or S-Bahn lines is neat, which is why rail activists talk about through-running so much – it’s right there posted at every station and on every railcar. But the speed at which people get on and off the train is not as obvious, and it requires looking into detailed reports to do an even rudimentary comparison, none of which in the case of Paris is available in English or easy to find on Google (the word “stationnement” usually means “parking,” in the same manner that the word “dwell” usually means “to live in a place”).

The upshot is that consultant reports written by serious people who absorb the knowledge of the railroaders of the Northeastern US with some British sanity checks can still say things that are so wrong to make the entire report useless. The same process that produces the whopper that the Munich S-Bahn, built 1965-72, took 46 years to build, can produce a Gantt chart that has a combined boarding and alighting time with conductor check that’s more than five times longer than what Penn Station in its current configuration is capable of and more than 10 times longer than what Gare du Nord achieves with similar peak ridership. Based on this false belief regarding dwell times, the agencies are then convinced that through-running is difficult and, separately, many additional tracks at Penn Station are required to fully use the capacity of the under-construction Gateway tunnel, building which would waste $16.7 billion.

Tradeoffs in Reliability and Shutdowns

I am writing this post riding trains between Brussels and Berlin. My connection in Cologne was canceled as the connecting train was moved to depart earlier than my first train’s arrival time, and somehow, it is faster to stay on the train until Frankfurt and connect there, the trains between Cologne and Berlin are so disturbed this summer. Cologne-Berlin, normally a direct hourly connection in 4-4.5 hours, is slowed to 5.5 hours every two hours this summer. It got me thinking about something Jon Worth said last month about the importance of public transport being there, including at night, because it reminded me of how there are always tradeoffs. Train service cannot literally run 24/7 without changes; maintenance windows are required. So it’s a question of tradeoffs – when service must run less reliably, or not at all. Deutsche Bahn has unfortunately chosen a grossly wrong side of the tradeoff, leading to summertime shutdowns and slowdowns that its French and Japanese peers simply do not have. Those shutdowns, in turn, are, these days, leading to catastrophic levels of popular mistrust in DB.

The tradeoffs

I wrote six weeks ago about the problems of summer maintenance in Germany. But, more generally, there is a tradeoff between span of service on a railway and how consistently service can be delivered. A railway that runs overnight will not have regular maintenance windows, and therefore have to pick some low-traffic period for a special disturbance. On the New York City Subway, this is the weekend: New York City Transit exploits its four-track mainlines and high levels of redundancy in most of the city to shut down individual sections of track on weekends and tell passengers to use alternatives. In Europe, it’s more common for this to be the summer period, when local travel is lower as people go on vacation; unfortunately, in Germany, this extends to intercity rail, during the high season of travel.

Jon says that, “That 5am train with a dozen building workers on it, or the last train home in the evening matter for the trust and reliability of the system, even if those individual trains make heavy losses and are largely empty.” But the point is that knowing that I can book a train in July and have it run as expected without being rerouted onto the slow line is, like the 5 am train, a matter of trust and reliability too. It’s just a matter of which matter of reliability is easier to compromise on.

Then there is a tradeoff of all of this against maintenance efficiency. It is more efficient from the perspective of minimum total gross hours of shutdown to have a long continuous period of shutdown, such as the four-month period planned for the Riedbahn. Nighttime shutdowns require an hour of preparation and disassembly at each end, so that a five-hour nighttime shutdown only yields three hours of maintenance work. Some systems don’t make that work even with regular nighttime shutdowns, such as the London Underground or American systems that are not New York; notably, the Berlin U-Bahn manages to avoid this even with overnight service on weekends.

The situation in Germany

DB’s response to the tradeoffs outlined above is to attempt to run all day, including occasionally at night. There are night trains between Hamburg and southern Germany on the Frankfurt-Cologne high-speed line, so even this line, without any nighttime freight (the grades are far too steep), does not have the regular maintenance windows that LGVs and Shinkansen lines have. As a result, last month, the line was shut for maintenance, and trains were diverted to the old line, taking an hour longer. Right now, the same diversions apply to Cologne-Berlin trains, slowing them by about an hour.

These are not peripheral connections. Frankfurt-Cologne is not quite the busiest intercity line in Germany – that would be the Riedbahn – but it’s a fairly close second, with the same planned traffic level in the Deutschlandtakt of six trains per hour in each direction. It’s the primary connection between the Rhine-Ruhr and not just Frankfurt but also all of southern Germany. Then, Berlin-Cologne connects the two largest metro areas in Germany; the Rhine-Ruhr is close in population to Ile-de-France, while Berlin and Brandenburg have more people than Rhône-Alpes or PACA, which has implications for how much traffic this connection would have if it were fast and reliable, which it is neither (government officials fly between Berlin and Bonn instead of relying on DB).

Is this unavoidable?

No. France has none of these daytime shutdowns on its main lines. Neither does Japan.

German rail advocates sneer at France and ignore Japan, finding all manners of reasons to avoid learning from countries that, on this point, are Germany’s superiors. A common line from within Germany is that its secondary lines are in better shape than France’s, so there is nothing to learn from France. But then, the reason there are routine hour-long delays (or longer) in the summer on the main lines is not that DB runs better service to a city like Siegen or Münster or Jena than SNCF does to their French peers.

The path forward has to be, at the technical level, to institute regular nighttime maintenance windows, and stop trying to make night trains happen. At infrastructure level, it must be to avoid building dual-use infrastructure, and build passenger-dedicated high-speed lines; if freight capacity is needed that the old lines with just slow regional trains can’t provide, then build a separate freight line, based on the needs of freight, at costs that are going to be lower than the long tunnels required for dual-use lines.

But the most important change has to be at the level of governance and culture. Germany believes itself to be at the top of the world. To borrow a joke about Japanese technological stagnation, there is an element here that visiting a German infrastructure system in 2005 had a futuristic vibe like visiting the year 2015, and visiting it today is still like visiting the year 2015. There’s a slew of problems in Germany for which the solution really is “be less German and more French,” and this is one of them, no matter what people who think all French people are unemployed rioters think.

More on American Incuriosity, New York Regional Rail Edition, Part 1: European History

The Regional Plan Association ran an event 2.5 days ago about New York commuter rail improvements and Penn Station, defending the $16.7 billion Penn Station Expansion proposal as necessary for capacity. The presentation is available online, mirrored here, and I recommend people look at the slides to understand the depth of the ignorance and incuriosity of area decisionmakers about best practices displayed in the first half of the presentation; the second half, by Foster Nichols, is more debatable. I hope to make this a series of two or perhaps three posts, focusing on different aspects of why this is so bad. But for now, I’d like to just talk about what the presentation gets wrong about the history of commuter rail improvement in Europe, on pages 17-19. Suffice is to say, the extent of error that can be crammed into a single slide with little text astounded me. With such incuriosity about best practices, it’s not surprising that regional power brokers are trying to will the unnecessary Penn Expansion project into being, never mind that it has no transportation benefits despite its extravagant cost.

The rub is that the presentation on pp. 18-19 says that commuter rail through-running is really hard. Here is page 18:

  • Regional metro systems comprise a targeted portion of regional rail networks centers of population, employment, business or major attractions like airports that support frequent, fast service
  • Regional metro systems typically do not operate within original historic train sheds
  • They operate in new tunnels, shoulder stations adjacent to existing major stations, and separate, simpler interlockings that facilitate frequent service

Then, page 19 shows maps of the RER, Munich S-Bahn, Elizabeth line, and Thameslink, quoting the length it took to build them as, respectively, “30 years,” “46 years,” “2001-2022,” and “1970s-80s, 2009-2020.” The conclusion is “Systems take decades to implement, usually in stages.”

And all of this is a pack of lies.

In fact, commuter rail through-running systems routinely reuse legacy stations, even fairly major ones: both Berlin and Munich Ostbahnhof were incorporated into their respective S-Bahns, and several Parisian train stations were reused for the RER, for example Gare d’Invalides or Luxembourg, with varying levels of modification. New stations are built from scratch underneath surface stub-end terminals like Gare du Nord and Gare de Lyon as depicted in the presentation, but if the station already has through-tracks then it can be used as-is, like Munich Ostbahnhof, and in some cases even stub-end stations are at such grade that their infrastructure can be used. If Boston chooses to build the North-South Rail Link, then, since North and South Stations are both large at-grade terminals, the link will have to include new underground platforms at both stations. But Penn Station is an existing through-station below grade; Amtrak already runs through, and so could commuter rail, without adding platforms.

And as for the lines about the systems having taken 30 and 46 years to build, this is so painfully wrong that it is perhaps best to go over their actual histories. The actual length of time it took depends on one’s definitions, especially for Paris, but the maximum one can support for Paris is 16 years; for Munich, it is seven years.

The history of the RER

The RER and Transilien are, together, the largest commuter rail network in Europe by ridership, with around 1.1 billion annual riders. Globally, only four systems surpass them: Tokyo, Seoul, Osaka, Mumbai; the first two are integrated metro-commuter rail networks to the point that it’s hard to distinguish which mode they are, Osaka is several competing companies none with the ridership of the combined Paris system, Mumbai runs with practically no metro accompanying it. The RER’s history, as I will shortly explain, also makes it a good prototype for modern commuter rail operations, of the same type that is called S-Bahn in Germany. New Yorkers would do especially well to understand this history, which has some parallels to the administrative situation in New York today.

The topline of this is that since the 1960s, Paris has connected its legacy commuter and intercity rail terminals with new through-tunnels, called the RER, or Réseau Express Régional. There are five lines, dubbed A through E. Métro operator RATP runs most of the RER A, and the RER B south of Gare du Nord; national railway SNCF runs the rest plus commuter train networks stub-ending at most of the historic terminals, called Transilien, signed with letters from H to R.

A rough before-and-after map can be found below, as used in the ETA report on through-running:

But the history of the RER goes back further – and none of it can be said to have taken 30 years. In short: the Métro was built, starting in the 1890s and opening in 1900, to be totally incompatible with mainline rail – for one, where mainline trains in France run on the left, the Métro runs on the right. This was on purpose: city residents in the Belle Epoque already looked down on the suburbs and worried that if the Métro were compatible with the mainlines, then it might be used to connect to the suburbs and bring suburbanites to their city. The stop spacing, separately, was very tight, even tighter than on New York local subway trains, let alone the London Underground. By the time the system reached the inner suburbs in the 1930s, it was clear that it could not by itself connect the growing suburbs to the city, it would be too slow.

Various proposals for investment in commuter rail go back to the 1920s, but little happened, with one exception: the Ligne de Sceaux, shown as the blue line on the first image entering the city from the south, was acquired by the forerunner of RATP, CMP, in 1938, as the rest of the French mainline network was nationalized. CMP was attracted to the line because of its atypically good penetration into the center of Paris – the other lines terminated farther from the historic center, for example at Gare du Nord or Gare de Lyon. The line was also not useful for SNCF as it was being formed, due to its isolation from the rest of the network. The line was electrified as it was acquired, and run as a regional line, still isolated from all others.

More serious plans for commuter rail through-running began in the 1950s, as postwar growth and suburbanization put more pressure on the system. Gare Saint-Lazare was especially under pressure, first because of growth in the western suburbs, and second because the Paris CBD had been creeping west, making its location more attractive for commuters. In 1956, Marc Langevin proposed an eight-line network; in 1959, RATP and SNCF began collaborating, planning east-west and north-south lines. As late as 1966, there were still plans for two separate north-south lines (for example, see here, p. 244), of which only one has been built and the other is no longer seriously proposed.

In the 1960s, the plans got more serious. Construction began in 1961, starting with the east-west axis, still with an uncertain alignment. Eventually, RATP would take over the Ligne de Vincennes (the eastern red line in the before map) in 1969 and the Ligne de Saint-Germain-en-Laye (the southernmost of the western red lines) in 1972, and connect them with a new tunnel, opening in 1977. Over the 1960s, the plans still had to be refined: it was only in 1963 that it was confirmed that the Ligne de Vincennes’ Paris terminal, Bastille, was too small to be used for this system, and therefore the new tunnels would have to begin farther east, to Nation, which opened in 1969 and is thus already depicted on the before map.

The Ligne de Vincennes was simultaneously modernized, starting in 1966. The entire systems had to be redone, including new platforms and electrification. Nation had to be built underground, starting 1965, complete in 1967 and opening with the rest of the line in 1969.

On the west, the cornerstone was laid in 1971, and construction began shortly later, starting with La Défense. Shuttle trains run by RATP opened between La Défense and Etoile in 1970, and extended to Auber in 1971. In 1972 the line was connected to the Ligne de Saint-Germain-en-Laye.

At the same time, deepening SNCF-RATP integration meant that the planned alignment within the city would need to change to connect to SNCF’s train stations better. Originally, the east-west axis was supposed to run as an express version of Métro Line 1, stopping at Etoile, Concorde, and Châtelet; this was modified to have it swerve north, replacing Concorde with Auber, which is connected to Saint-Lazare. East of Châtelet-Les Halles, the alignment swerves south to connect to Gare de Lyon instead of Bastille.

In 1977, the Nation-Auber section opened, finally offering through-service; the appellation RER A dates only from then. Simultaneously, the north-south axis that was actually built half-opened, connecting the Ligne de Sceaux onward to Les Halles, with cross-platform transfers from the south to the west. On the same date that the central section opened, RATP also inaugurated an entirely greenfield branch of the RER A to the east, initially to Noisy-le-Grand, eventually (by 1992) to the new Marne-la-Vallée development, where Eurodisney was built. Contemporary media reports called Les Halles the biggest metro station in the world, and President Valéry Giscard d’Estaing (center-right) spoke of public transport for everyone, not just the poor. The cost of this scheme was enormous: it cost 5 billion francs (update 8-9: see Alain Dumas’s comment below – it’s 5 billion FRF for the entire RER A, not just the Nation-Auber section), which would make it about $1 billion/km $350 million in 2023 prices, inflation since then more or less canceling out the franc:USD exchange rate. The RER B cost 400 million francs between Luxembourg and Les Halles, a distance of 2.3 km, and 1.6 billion to get to Gare du Nord and connect to the SNCF network to the north (opened 1981), a distance of 3.5 km.

The RER C then opened in 1979, as a second east-west line, on the Left Bank. Missing all of the main centers within Paris, it has always had far lower ridership than the RER A; it was also much easier and cheaper to build – all that was required was a short tunnel connecting Invalides on the west, previously a subsidiary commuter rail-only stop on the same lines to Montparnasse and Saint-Lazare, and Gare d’Orsay on the east, a commuter rail-only extension of the line to Austerlitz. This was built quickly – the decision was made in 1973, and the line opened within six years. This required a total rebuild of Gare d’Orsay with new underground platforms; Invalides required reconstruction as well, but could use the same station and track structures.

Subsequently, the system has added new lines and branches – the RER D opened from the north to new Gare de Nord platforms in 1982, was extended in 1987 along the same tracks used by the RER B to Les Halles but serving dedicated platforms at both stations, and was extended along a new tunnel to and beyond Gare de Lyon in 1995; the RER A acquired new western branches in 1988 to be operated by SNCF, requiring dual-voltage trains since those branches use 25 kV 50 Hz AC and not 1.5 kV DC like the RATP lines; the RER C acquired a new branch also in 1988 taking over part of the Petite Ceinture; the RER E was opened as a stub-end extension of lines from the Gare de l’Est network to a new underground station at Saint-Lazare in 1999, and was finally extended to the west with some through-service this year.

So in a sense, it’s taken 63 years to build the RER, starting 1961, and the work is not yet done. But the core through-running service opened in 1977, within 16 years, with some decisions made midway through the works. The total required work greatly exceeded anything New York needs to do – just what opened through 1977 includes 16 km of double-track central tunnel on the RER A, 3 km on the new branch to Noisy plus 6 km of new above-ground line, 2 km of tunnel on the RER B, and around one km of tunnel on the RER C, inaugurating eight new underground stations, all on the RER A. The RER A’s ridership reached 1.4 million per workday by 2019, and the RER B’s reached 983,000 – and a great majority of the work on both was done by 1981.

The history of the Munich S-Bahn

The Munich S-Bahn is not the oldest or busiest S-Bahn system in Germany; Berlin and Hamburg both have prewar systems, and Berlin’s ridership is considerably higher than Munich’s. Nonetheless, precisely because Berlin and Hamburg built so much of their infrastructure in the steam era, some lessons do not port well to cities today. In contrast, Munich’s entire system has been built after the war – in fact, the construction of the S-Bahn took place over just seven years, from the decision of 1965 to opening in 1972, timed with the Olympics.

As in Paris and many other cities, the history of proposals for rapid urban mainline rail in Munich stretches back decades before the decision was made. The first proposal was made in 1928, and there was more serious planning in Nazi Germany, as the Nazi Party had been founded in Munich and was interested in investing in the city due to that history; by 1941, there were plans for a three-line system, comprising a north-south, an east-west, and a circular tunnel. But little was built, and during the war, the resources of Germany toward rail were prioritized in a different direction.

After the war, Munich grew rapidly. It was not much of an industrial city in the early 20th century; early industrialization in Germany was mostly in the Ruhr and Saxony, while the professional services economy was centered on Berlin, whose metropolitan area in the 1930s was of comparable size to that of Paris. After the war, things changed, at least in the West: the Ruhr’s coal and steel economy stagnated, while southern Germany grew around new manufacturing of cars and chemicals; decentralization dispersed the professional services economy, and while most went to Frankfurt and Hamburg, a share went to Munich (for example, Siemens’ headquarters moved there from Berlin right after the war). The city’s wartime peak population was 835,000; it would surpass 1 million in 1957 and is 1.5 million today. The region, Oberbayern, comprising essentially the metro areas of Munich and Ingolstadt, would grow from 2 million at the beginning of the war to 2.8 million by 1960 and 4.8 million today, and is the richest region in the EU at this scale, with per capita income from work approaching that of New York.

This small size of Munich in 1900 means that it never had as extensive a rail network as Paris or Berlin. It had just two major urban stations: Hauptbahnhof, a terminal with a station throat leading to points west, and Ostbahnhof, a through-station with tracks leading east, south, and the west, the western tracks looping back south of city center to reach Hauptbahnhof. To this day, area railfans would like this loop to be incorporated into a regional S-Bahn system avoiding city center – but Munich is still a rather monocentric city. There was no U-Bahn, unlike in Berlin or Hamburg.

By 1961, the number of suburban commuters into Munich reached 114,000. The undersize rail network relative to the city’s current importance and the rapid growth in wealth meant that car ownership was high, leading to traffic congestion. The trams were slowed down by traffic, to the point of not running faster than walking in city center.

To resolve these problems, both an U-Bahn network and an S-Bahn network were planned. Early planning began in the 1950s, with the federal government taking over the wartime plans in 1956, but as in Paris, the extent of the system to be planned was up in the air: both an east-west axis and a north-south line were desired, and only in 1963 was the decision finalized that the north-south axis should be a municipal U-Bahn tunnel and not an S-Bahn. The study period began in 1961, with the plan approved in 1965 for the construction of a single east-west S-Bahn tunnel between Hauptbahnhof and Ostbahnhof, and a separate U-Bahn system with three branched trunk lines.

Construction was done on a tight timeline, since Munich was awarded the 1972 Olympics in 1966, and delays were not considered acceptable; the first U-Bahn line, U3/U6 running north-south, opened 1971, and the S-Bahn opened 1972, in what is described as a “record time.”

During the seven years of construction, other projects had to be done in parallel. Commuter rail lines had to be extensively upgraded: the project included 143 km of electrification, and 115 stations outfitted with new high platforms at a level of 760 mm mostly 210 meters long. Simultaneously, most of what has become the standard for good timetabling was invented, out of necessity on a network that had to share tracks and systems with other trains on its outer margin, most importantly the clockface schedule – the system was designed around a 20-minute Takt on each branch from the outset, with outer tails running every 40 minutes.

The central tunnel itself, the Stammstrecke, comprises six stations from Hauptbahnhof to Ostbahnhof of which all except Ostbahnhof are underground, and three have Spanish platforms. Ostbahnhof itself is used as a pinch point for some trains, reversing direction depending on branch. The Stammstrecke in total was built for 900 million DM, or $2.8 billion in 2023 PPPs; the overall line included 4.1 km of tunnel and about 7.3 more km of above-ground connections. (Update 8-9: cost fixed – I originally stated it to be 900 DM.)

There has been further investment adding new branches and upgrading the system. The new signal system LZB was installed in the central section experimentally when it opened in 1972, but it was not used on all trains, and was taken out of service in 1983, only returning in 2004 when its capacity was needed, boosting throughput from 24 trains per hour to 30. However, as in Paris, the core of the system’s high ridership, now about 900,000 per workday, comes from infrastructure that was there from the start, and thus it’s most correct to say that the system took not 46 years to build but seven.

Some lessons for New York

By the standards of Paris and Munich, New York has practically everything it needs to run through-service. The electrification systems on its three commuter railroads are not compatible, but multivoltage trains not only are routine, but also already present in New York; the current configurations all have one problem or another, but fundamentally, ordering multivoltage trains is a solved problem. Only a handful of outer branches need to be electrified, and all can be deferred, running with forced transfers until they are wired as is current practice on the Raritan Valley Line and for the most part also the outer Port Jefferson Branch. The LIRR and Metro-North are entirely high-platform and New Jersey Transit’s Manhattan-facing lines only have 68 low-platform stations of which 26 are already funded for high platform conversions.

By far the biggest missing element for New York by cost is the Gateway Program and its Hudson Tunnel Project, which is budgeted at $16 billion and is funded and beginning construction, with the New Jersey land tunnel contract just awarded. Even before the new tunnel opens, it can run some through-service after Penn Station Access opens from the Hell Gate Line, pairing it with some New Jersey Northeast Corridor trains.

On top of that, some surface improvements are prudent, such as some grade separations of rail junctions, the most expensive costing on the order of hundreds of millions (Hunter is $300 million on the budget, maybe $400 million by now); much of that is already getting funds from the Bipartisan Infrastructure Law or likely to get them in the near future, since the infrastructure is also used by Northeast Corridor intercity trains.

But it does not need to do anything that area railroaders have convinced themselves they need, especially not new tracks at Penn Station. Nor are decades of prep work needed – rapid installation of high platforms is completely feasible, as was done not just in Munich in the 1960s and 70s but also in suburban New York in the same period and in the 1980s and 90s, converting the LIRR and Metro-North to full high-platform operations and doing the same on the Northeast Corridor in New Jersey.

All that is needed is a modicum of curiosity about the world, curiosity that is not seen in the presentation with its whoppers about the timelines of the RER and Munich S-Bahn, or its belief that new underground tracks are always required as if Penn Station is the same as the surface Gare du Nord. I find myself having to explain to journalists who interview me that all of this can be done, but the people in charge of the railroads around New York cannot do it.

Why New York-New Haven Trains Crawl

Between New York and New Haven, a distance of 120 km (from Penn Station) or 116 km (from Grand Central), the two fastest intercity trains of the day take 1:35 to travel, an average of 75 km/h. Most do the trip in about 1:40, averaging about 72 km/h. Commuter trains to Grand Central do it in about 1:40 three times a day, averaging 70 km/h, but the vast majority of even the rush express trains are slower, a few doing it in 1:52 and most in about two hours, averaging 58 km/h. This is not normal for a primary intercity corridor; the Acela averages about 120 km/h between New York and Washington and between New Haven and Boston, which is typical for non-high-speed intercity lines in Europe, while high-speed ones usually average 200 km/h or more. I’ve been asked by some big names in online transit content creation why this is so, and hope to explain why the trains are slow, and what it would take to reduce 40 minutes from the one-way trip time.

The contrast should be with the high-speed rail proposal that I’m working on at Marron, which cuts the intercity trip time between New York and New Haven to about 52 minutes, on the existing right-of-way, and the express commuter rail trip time to Grand Central to about 1:16. The result is not high-speed rail, but is a fast upgraded intercity rail line, on a par with the faster British and Swedish lines. Changes in right-of-way geometry, including buyouts of houses in expensive suburbs in Connecticut, could reasonably cut the intercity trip time to about 45 minutes; these are mapped here, the 52-minute trip corresponding to the alternatives that stay on the existing right-of-way and the 45-minute one to the alternatives that use the bypasses where they exist.

The primary culprit for the slow trip times today is poor scheduling practices. Those practices, in turn, come from mutual abuse between Amtrak and the commuter rail operators, in this case Metro-North and the Connecticut Department of Transportation, both of which display terminal incompetence on all matters related to rail. The state of the tracks contributes to the slowness, and thus the second most important issue is poor maintenance practices leading to unreliable infrastructure, which then feeds into poor scheduling. Metro-North and CTDOT are again especially bad even by American standards. Physical infrastructure problems add minutes here and there, but the most important interventions are cheap and for the most part can only work with better timetabling rather than on their own.

Of note, it is common to blame the low speeds on curves. However, the curves are not especially onerous – few restrict trains to slower speeds than about 150 km/h given good operating practices. In fact, the Northeast Corridor gets if anything curvier east of New Haven until after it crosses into Rhode Island, but the speed there is higher, as there is less dense commuter traffic complicating the schedule, and Amtrak’s level of incompetence is bad but less bad than that of CTDOT.

Timetable padding

Every rail timetable has to include contingency or buffer time. This takes into account primarily the need for trains to recover from delays, and secondarily suboptimal driver behavior, such as starting to brake a little too early. Switzerland pads its timetables 7%; the TGV network can only do about 10-13%, and the ICE network about 25%. What I and others have seen on Amtrak and Metro-North trains as well as what train drivers have told me suggests that the buffer time between New York and New Haven is 25% or even maybe 30%.

More complex networks require more padding, since delays on one train cascade to others. The ICE network mixes intercity trains together with much slower regional ones on the same tracks, all over Germany, and delays can cascade across the entire country, to the point that some people have begun to advocate that Germany build a separate high-speed rail network, not for speed (which activists here don’t care much about), but for the reliability of having a fast network and a slow network rather than one mixed network. The more segregated TGV network thus does better; the almost entirely dedicated-track Shinkansen system does even better, and JR East suggested 4% padding in its review of California High-Speed Rail. Switzerland is like Germany in having a single mixed-speed network, but it has more systematic processes for avoiding delays, such as strategic investment in bypasses around known bottlenecks.

The Northeast Corridor is not an especially complex network. It is a single line with branches, rather than a two-dimensional mesh like the German rail network. There is little freight traffic, which makes it possible to control freight through regular slots, with the number of potential slots greatly exceeding actual traffic so that if a train misses its slot, it can wait 10 or 15 minutes for the next one. Passenger traffic is high on all lines serving the corridor, and thus there is no need to cut corners on reliability (such as signals, or platforms) on any of the branches. It is a mixed-speed line, but nearly all of it has four tracks, and where commuter trains share tracks with intercity trains, they run express and the speed difference is not large. In the timetables we developed at Marron with Devin Wilkins, express commuter trains do Stamford-Grand Central in 28 minutes if they run as today, stopping only at Harlem-125th, and in 29 if they also stop at New Rochelle; intercity trains do Stamford-Penn Station in 25 minutes, on a marginally longer route into New York. Slotting intercity and express commuter trains on the same tracks between Stamford and New Rochelle is annoying, but is not an objectively hard scheduling problem.

This does not mean that Amtrak and Metro-North could just shave minutes off of the existing timetables, change nothing else, and run trains to the faster schedules. Other elements of the schedule would make the trains too unreliable. But it is possible to realign the schedules appropriately and cut the trip time by a factor of about 1.3/1.07 = 1.2.

Timetable complexity

The ideal schedule is one with as few variations as possible. This way, planners can write one schedule, ensure that it works, and, if there are problems with it, then develop an infrastructure program that builds around the bottlenecks. Switzerland, as usual, sets the standard, with its all-day repeating clockface timetable, or Takt. Swiss trains repeat regularly every hour, and on the busy lines every half hour; planners need to make sure one pattern works and then repeat it all day. It’s the planning equivalent of economies of scale in manufacturing.

New York planning, relative to the ideal, represents the list of what not to do, and it’s worse on busier lines such as the New Haven Line than on less busy lines. In effect, the New Haven Line schedule is the planning equivalent of rules for writing prose that illustrate each rule by breaking it – remember to not split infinitives, the passive voice should be avoided, eschew obfuscation, and so on – except that it is meant to be taken seriously. It has all of the following problems:

  1. Where good planning begins with one peak hour and repeats it all day, the New Haven Line has few repeating patterns, and practically none at the peak.
  2. Where good planning aims to have trains make consistent stops for legibility and for ease of planning around bottlenecks, the New Haven Line has bespoke stopping patterns – not counting branches, there are 16 trains entering Grand Central at the peak hour, which make 13 distinct stopping patterns.
  3. Where good regional rail planning keeps the peak-to-base ratio low – Switzerland is almost 1:1, and even very large cities that need a huge volume of commuter trains at rush hour like Paris or Tokyo do not exceed 2:1 (and London is well below it) – the New Haven Line has, with branches, 20 trains entering Grand Central at the peak hour and 4 entering each off-peak hour.
  4. Where good planning runs more or less the same service on weekends as in the off-peak on weekdays, the New Haven Line’s midday off-peak and weekend schedules are different even as they run the same number of trains (two express and two local per hour).
  5. Where good planning aims to use the timetable for a prolonged period of time to reduce the need to redo the schedule, for example updating annually as in Switzerland, New York-area practice is to update several times a year, in what looks like a 3-6 month period.
  6. Where good planning keeps the trains spaced far enough based on signal system constraints by default, Metro-North timetables somehow have trains on the shared trunk between Harlem and Grand Central sometimes arriving within less than the 2 minute minimum on the same track, requiring special speed restrictions, even with unimpressive traffic levels by urban commuter rail trunk standards.
  7. Where good maintenance is done when trains are not running, that is, at night, in order to avoid disturbing weekday traffic, American planning assumes that daytime maintenance will always take some track out of service; the New Haven Line’s track renewal program has been so mismanaged that at no point since it began in the 1990s have all four tracks between New York and New Haven been operable along the entire line – some section is always shut down. Daytime maintenance is also a problem in Germany, and is a factor behind the poor schedule reliability here.

The constant tweaks to the timetable are also a feature of the New York City Subway, with its substantially simpler stopping patterns. There, the services are consistent, and change at a rate of a handful per decade (most recently, when Second Avenue Subway opened; the previous time was during the 2010 service cuts). However, frequency is micro-targeted based on crowding guidelines, so the planners never have time to optimize one schedule; moreover, with 24/7 service, daytime closures for maintenance are unavoidable. This way, where planners at healthy railroads write schedules, planners at American passenger railroads write service changes. The New York City Subway at least has the partial excuse of 24/7 service; Metro-North has no such excuse. The maxim that the Northeast Corridor is held together with duct tape, and is managed by people who are unfamiliar with any more advanced tools than duct tape, also applies to timetabling.

In contrast with today’s morass, the schedule we’ve been writing aims to simplify whenever possible. Branches are slotted into windows that could be used by local or express main line trains depending on the desired service pattern. From New Haven south, everything is on a repeating 10-minute Takt. The New Haven Line is reduced to four stopping patterns – local Stamford-Grand Central, local Stamford-Penn Station, express New Haven-Grand Central, intercity New Haven-Penn Station – each running every 10 minutes. It took weeks to find a pattern that worked with all the constraints of the right-of-way and allowed some future desired infrastructure changes, and even that required some track changes detailed below. Off-peak, the commuter train patterns could run every 20 minutes instead, using every other slot; the timetable should not be tweaked further.

It is particularly important to avoid timetable complexity beyond local and express trains east of Stamford. The line has four tracks, and could be run with commuter trains on the local tracks, making all stops before transitioning to the express tracks at Stamford, and intercity trains on the express tracks, running nonstop between Stamford and New Haven. In theory, this means this section could be run with less than 7% schedule padding, for example the Shinkansen’s 4%, but in practice, I suspect it cancels out with the more complex situation between Stamford and New Rochelle, so 7% is the best that can be squeezed with maximally simple schedules.

Speed zones and curves

The New Haven Line is rather curvy, having been built in the 1840s. But its speed limits are still too low for its curves. I wrote here about cant and cant deficiency, and am not going to repeat myself too much. But, in brief, the speed on curves is governed by the formula

v^2 = ar

where v is speed, a is lateral acceleration in the horizontal plane, and r is curve radius. The value of a is usually expressed not in units of acceleration, but in units of distance, scaled so that, on standard-gauge track, 150 mm (of cant) correspond to 1 m/s^2 lateral acceleration. Typical maximum regulatory limits on cant range between 160 and 180 mm; the US permits 7″, but nowhere is more than 6″ used, and the New Haven Line’s curves mostly range between 3″ and 5″ cant. Cant deficiency limits depend on the train – regular passenger trains typically do 130-150 mm at the relevant speeds, but in the US, the normal practice is to limit commuter trains to 3″ cant deficiency, and only use 5″ on Amtrak Regional trains (the Acela tilts and is capable of 7″ today, with the new trains rated for 9″).

The curves on the New Haven Line are, for the most part, built to a standard of 2° radius, or, in metric units, r = 873. The most aggressive common cant and cant deficiency limits, 180 and 150 mm respectively, allow a = 2.2, and thus v = 43.82 m/s = 157.77 km/h; our timetables limit commuter trains to 150 km/h, and there are surprisingly few curves with tighter limits. In contrast, current practice restricts a to about 1.2, which means trains take the same curves at a speed of about 116 km/h, which is rounded down to 70 mph.

The slowdowns also affect intercity rail more than is required. While Amtrak trains are cleared for 5″ cant deficiency, Metro-North prefers to timetable all trains at its own trains’ speed on curves. Then, because there are so few opportunities under current standards for trains to run faster than 70-75 mph within CTDOT territory, the entire line from the state line to New Haven is maintained to those standards, and thus even on relatively straight sections, there is no opportunity to gain speed. East of New Haven, the curves are if anything tighter, but Amtrak dominance means the tracks are cleared for 100-125 mph, cant is higher, and cant deficiency is higher as well.

All of these restrictions can be lifted. The work required to redo a line from 110 km/h to 160 km/h or even more is rather routine, as long as it can be done within the right-of-way. The standards for track irregularity get tighter as speed increases, but all of this can be handled with track laying machines, which use the track itself to do the work, at a pace of about 0.5 km/h, or about 1.5 km in a three-hour nighttime work window; the entire New Haven Line can be regraded in about a year this way.

Unfortunately, Metro-North is used to manual track inspections rather than modern machinery. It finally bought a track laying machine on the model of Amtrak, but appears not to use it very well; the productivity I hear quoted is one tenth what was expected. But what is hard for Metro-North and CTDOT is not objectively hard, and even other Northeastern American railroads are often capable of it.

Supportive infrastructure

Infrastructure construction and timetabling work in tandem normally. Swiss practice is to use insights from the timetable in theory and in practice to inform where to build new tracks. American practice does no such thing – for one, Metro-North is allergic to systematic track improvement, so over the generations, the timetable has diverged from the infrastructure that could support it.

In fact, a very high-frequency peak schedule requires eliminating at-grade conflicts whenever it is even remotely feasible. Shell Interlocking at CP 217, just south of New Rochelle, is a flat junction on which trains from the north can go to either Grand Central or Penn Station. Grade-separating the junction was occasionally on the wishlist for Northeast Corridor improvements, but Metro-North is not currently asking for it, even though it is especially important as Penn Station Access is about to open. The junctions with the branches farther north – New Canaan, Danbury, Waterbury – are flat as well, for which the solutions can be a forced transfer (as is sometimes practiced with Waterbury, the weakest of the three) or grade-separation. This does not cost a large amount of money – New Jersey Transit is applying for money for its equivalent of Shell, Hunter Flyover connecting the Raritan Valley Line to the Northeast Corridor, and the budget is $300 million in the plan and, I’ve been told, $400 million with recent inflation and perhaps some small cost overrun.

Then there is the issue of the Grand Central approaches. The current throat limits trains to 10 mph on the last mile into the station. In other words, the last mile takes six minutes. It should take about two, based on actual throat and turnout geometry; the turnouts are #12 until around 700 meters from the end of the platform, and in Germany, a 1:12 switch is 60 km/h, and closer to the platforms, the turnouts are #7 and (on one cluster of tracks) #6.5, where in a Germany, a 1:7 is 40 km/h. Even with bumper tracks, the last mile has no reason to take longer than two minutes, saving all Metro-North travelers to Grand Central four minutes. The turnouts would need to be regraded to tangential standards, but this can be done within their existing footprint; the cost of a new turnout in a selection of European countries and also on American freight railroads is around $250,000 in the prices of the 2010s, whereas Metro-North’s switches cost perhaps five times much in the same era.

Finally, the movable bridges impose certain speed restrictions. Those are the biggest projects currently in planning for speeding up the New Haven Line. In truth, the slowdowns imposed are secondary (though our timetables still assume they are fixed). They are also extremely expensive – one of them is currently slated for in situ replacement for $1 billion, for a span of 220 meters from tower to tower, on a river about 100 m wide. CTDOT rail projects are generally absurdly expensive even by American standards – infill stations on the Hartford Line are coming in at $50 million or more, twice the cost of suburban Boston and more than twice that of suburban Philadelphia – for which the culprit must be poor project management and lack of in-house expertise.

Conclusion

The New Haven Line is a busy railroad at the peak, but nothing about it is special. It is old, but no older than faster sections of the Northeast Corridor or fast legacy intercity main lines in parts of Europe, especially the United Kingdom. It is busy, but its total ridership is unimpressive by European S-Bahn standards – the single trunk line in Munich with its seven branches on each side generates about 900,000 daily riders, perhaps a bit more than all three New York-area commuter railroads combined. It is branched, but the branching is simpler than on the busier systems, and the graph of the Northeast Corridor overall is acyclic, simplifying planning.

The reason the trains are slow is not the infrastructure. The elements of the infrastructure that need to be fixed to shorten the trip times from about 1:35 intercity and 2:00 commuter to 0:52 intercity and 1:16 commuter are cheap. Rather, the reason is that the line is managed not just by Americans, which is usually bad enough, but specifically by Metro-North and CTDOT. The schedules are designed not to work; the maintenance is designed not to work either and is too expensive.

New York Has Too Few Subway Countdown Clocks

When I was visiting New York in June-July, I was stricken by how hard it was to figure out when the next train would come. Every subway station is equipped with countdown clocks, the A Division (numbered lines) and L trains having older installations than the rest of the B Division (lettered lines). However, the B Division stations that I used did not have many countdown clocks, and I found myself having to walk long distances along hot platforms to figure out which train to take. I counted the number of clocks at a few stations, and asked ETA members to do the same; now back in Berlin, I’ve done some counts here as well, confirming that it’s not just me – New York’s B Division platforms have fewer and harder to find countdown clocks than the standard on the Berlin U- and S-Bahn platforms, even though New York’s more complex subway network requires if anything more clocks as passengers have multiple options. Based on what I’ve seen in Berlin, I recommend that New York install a minimum of four overhead clocks per B Division platform, with the screen going in both directions.

The situation in Berlin

The U-Bahn platforms seem standardized to me. The traditional norm was that stations were built cut-and-cover, right underneath a major street, with an entrance at each end of the central island platform. Nowadays almost all stations have elevators and there are plans for retrofitting the rest, which BVG estimates will be completed in 2028, the date having been pushed later over the years I’ve lived in the city. The elevators always connect two levels, with opposite side doors for the two levels, so that wheelchair users don’t have to turn.

There are, at the stations I use, two overhead countdown clocks for each platform face. Nearly all platforms are islands, and each direction has separate countdown clocks. The clocks display the times on both sides, and are typically located at the quarter points of the station, so that passengers are never more than a quarter of the platform length from a clock, with good sight lines; the platforms are 100-110 meters long.

The S-Bahn is less standardized. A full-length eight-car train is 150 meters is long. The countdown clocks are double-sided and overhead as on the U-Bahn, and each platform face has a separate clock even when the tracks are in the same direction (as at Ostbahnhof), but the number is inconsistent; there are stations with just one, but Friedrichstraße on the North-South Tunnel has three.

The situation in New York

The A Division has overhead countdown clocks, connected to the train control system (automated train supervision, or ATS), installed in the early 2010s; the L has countdown clocks of the same provenance. The number of clocks per station is not fixed, but ranges between two and four per track. The B Division’s train control system let the control center know where trains were but not which train was which – that is, which train on the same track is an A, which is a D, and so on – and therefore the same system was not installed at the time. Years later, a different system was installed, with nicer graphics and a different connection to the control center, which is sometimes less accurate.

This newer system on the B Division has a combination of overhead clocks, often single- rather than double-sided, and floor-mounted clocks facing sideways, toward the tracks rather than toward the front and back of the platforms. The floor-mounted clocks are difficult to read unless I’m standing right there. The platforms are obstructed so it’s hard to tell from a distance where the clock is. Worse, many floor-mounted installations look identical from a distance to the clocks, but instead display advertisements or service changes but no information about the next train.

What’s more, there just aren’t a lot of these clocks. At 2nd Avenue on the F, heading downtown toward Marron, I counted a single clock, but six boards displaying system maps or ads. ETA’s Alex Sramek checked several stations in Lower Manhattan, including Chambers on the A/C/E and on the J/Z, Fulton Street, Cortlandt Street on the R/W, and Broad Street, and found one to three clocks, always a mix of overhead and floor-mounted – and the floor-mounted clocks sometimes would only show the next train and not the subsequent ones, even for platforms serving multiple routes.

There should be more clocks in New York than in Berlin. The platforms are much longer – the A Division platforms are 155 meters, the L and J/Z platforms are 145 meters, the other B Division platforms are 185 meters. The extensive branching means that even while waiting on the platform, regardless of what information is displayed outside the station, it is important to know when each service using the station will come, to plan out which line to take. I made mistakes on trips from Brooklyn to Queens just because I wasn’t sure what to do when transferring at West 4th, where, having just missed the E, I needed to make a decision on whether to wait for a delayed F or try to make the B/D and transfer to the E at 53rd, opted for the latter, and missed the E at 53rd.

If a Berlin U-Bahn station has two double-sided clocks, and a major S-Bahn station has three, then New York should have four per B Division platform. These should be overhead and double-sided – the floor-mounted screens are difficult to see from a distance along the direction relevant to most passengers, and easily confused with ads, ensuring that their utility is marginal.

The Meaning of Construction Costs Per Rider

I’ve written a lot about urban rail construction costs per kilometer, but from time to time, my colleagues and I have been asked about what happens if we compare costs, not per kilometer, but per rider. There’s an intuition among people in transportation advocacy (including anti-transit activists who prefer cars) that the construction costs of urban rail lines per rider are a meaningful measure of cost-effectiveness. This intuition is true, and yet, it must be interpreted delicately.

First, modes of transit with different operating cost structures should tolerate different levels of capital costs; in particular, the current practice in which subways are built at higher cost per rider than light rail, which in turn is built at higher cost than bus lanes, reflects real differences in operating costs and does not mean there is overinvestment in subways and underinvestment in buses. And second, costs per rider can be too low, in a sense – if a city’s construction costs per rider are very low, indicating a very high benefit-cost ratio, then it shouldn’t be lauded for its fiscal prudence but scolded for not having built these lines long ago and for not building more today. In truth, places with healthy decisionmaking about infrastructure expand their networks to the limit of cost-effectiveness, which means that costs per rider averaged over an entire region vary less than costs per kilometer, and this just reflects that cities build what they can, so low-cost cities can afford to build lines to lower-ridership areas, which higher-cost cities would reject as too expensive for the benefit. This way, costs per rider are not too different in New York and in cities that build for an order of magnitude lower cost per km than New York.

The meaning of cost per rider

In the remainder of this post, the meaning of “cost per rider” is “cost divided by the ridership on a working day.” In Europe, workers get around six weeks of paid vacation, and tend to take them in the summer, leading to depressed ridership around July or August, depending on the city; daily counts usually avoid this period, so for example Stockholm specifies that daily ridership figures are taken in winter. This, as I will explain shortly, does not unduly make European lines look more cost-effective than they actually are.

The cost per rider is best understood as a cost-benefit measurement. All benefits of public transportation scale with ridership, generally linearly: higher ridership indicates tighter economic and social ties if it comes from more travel, and better environmental outcomes if it is at the expense of car travel. What’s more, raw ridership measured in trips is better at capturing these benefits than passenger-km. The issue is that focusing on p-km overrates the success of extremely suburban systems, which have low environmental benefits for their p-km (the users are typically park-and-riders and therefore drive extensively, just not to their city center jobs) and usually also high net operating costs since they are peaky and tend to charge low per-p-km fares. Conversely, the short-hop trip is a net profit to the system – even subways with distance-based fares charge degressive rather than linear fares – and comes from dense networks that cut out car-based travel entirely. These effects roughly cancel out to the point that ridership is a good proxy for actual benefits.

That said, all outcomes need to be scaled to regional or even national incomes. Economic benefits are usually measured relative to worker wages anyway; in some business case analyses, such as that of the United Kingdom, the economic benefit is even scaled to rider income rather than regional or national income, which favors lines built to rich neighborhoods over lines built to poor ones, and isn’t really how cities need to think about their public transit networks. Social benefits are usually taken on a willingness-to-pay basis, and the same is true of health benefits including reduced air and noise pollution from cars and reduced car accidents.

The next step is then to compare the cost per rider with GDP per capita, which is not perfect but is good enough as a proxy for incomes. This also takes care of the issue of Europe’s synchronized summer troughs in local travel: those six weeks of paid vacation are visible in reduced GDP per capita, so the apparent bonus to the European system of using cost per daily trip where “day” means “workday outside the summer vacation season” rather than cost per annual trip cancels out with reduced annual GDP per capita.

The rough rule of thumb I use is that the absolute limit of cost-effectiveness for a subway or commuter rail line is when the cost per rider is equal to GDP per capita. This is a coincidence: a one-time cost has no reason to be equal to an annual income – this just follows from Börjesson-Jonsson-Lundberg’s estimate of the Stockholm Metro’s benefit-cost ratio compared with its cost per rider relative to the GDP per capita of 1960s’ Sweden. In practice, infrastructure is never built down to a benefit-cost ratio of 1, due to construction risks; in countries that make decisions based on benefit-cost analyses, the minimum is usually 1.2 or 1.3. In this schema, the United States can afford to build up to an envelope of $85,373/1.3 to $85,373, which is $65,000-70,000/rider in 2024 prices. The frontier lines, like the Interborough Express, are fairly close to this limit already; in practice, there’s a range, with some lines in the same city built well over the limit for political reasons (often airport connectors) and others built far below it.

Cost per rider by mode

The above analysis works for subways and commuter rail. It does not work for trams or buses. The reason is that surface transit never achieves the same low operating costs as metros, so in practice, the total cost to be truly comparable needs to be incremented by the additional operating costs.

To be clear, this is just a rule of thumb. There are different metro lines, even with the exact same technology in the same city, with different projected operating cost profiles; for example, in Vancouver, the Broadway extension of SkyTrain toward UBC was projected in the 2010s to reduce net operating costs as many buses would be replaced by fewer, larger trains, but the outward extension of the same system deeper into Surrey and Langley is projected to increase net operating costs. There are different ways to interpret this – for example, the Surrey extension is in a more auto-oriented area, with more likely car-to-train switchers (this is still much denser than an American park-and-ride); on net, though, I think the differences are not huge and could to an extent even be folded into the notion of cost per rider, which is substantially better on Broadway than in Surrey and Langley.

That said, metros consistently have much lower operating costs than light rail and buses in the same city; here are American cost profiles. As far as I can tell from CoMET data, most European and Asian metros cluster toward the bottom end of the American cost profile (such as the Chicago L; the New York City Subway is the top end among the big systems); bus operating costs are more or less proportional to driver wages times operating hours throughout the developed world. Here we need to briefly switch to cost per p-km, since mature urban rail networks use buses as short-hop feeders – the counterfactual to a bus-based network for New York isn’t people riding the same bus routes as today but at higher intensity, but people riding longer bus routes, so the cost would roughly scale to cost per p-km, not per passenger.

In rich Asia, metros are profitable. In Europe, it depends – the London Underground operationally broke even in the early 2010s, and the Berlin U-Bahn was said to do the same in the late 2010s. In healthy European systems, it’s never reported directly, since there’s fare integration across the region, so financial data are reported at metropolitan scale without much breakdown between the modes, but the farebox operating ratios in at least Germany and Scandinavia, and probably also Paris (which has much higher ridership density than London or Berlin, comparable costs per car-km, and higher fares than pre-2022 Berlin), suggest that metros and the inner sections of commuter rail systems can break even, and then the subsidies go to the buses and to suburban extensions.

Individual bus systems can be profitable, but never at metropolitan scale, not in the first-world cities I’m aware of. In New York, the buses between New Jersey and Manhattan are profitable and run by private companies, but that’s one specific section of the system, and on net the bus system in New Jersey, including not just these cross-tunnel buses but also internal buses within the state, loses money, covered by New Jersey Transit subsidies, and the financial performance of buses within New York is, frankly, terrible.

One potential complication is that BRT infrastructure is usually installed on the highest-performing individual routes, and those can have rather low operating costs. But then, the operating costs of the buses on Broadway in Vancouver are extraordinarily low, and still the projections are for the SkyTrain extension that would replace them to, on net, reduce systemwide operating subsidies. If your city has a bus corridor so strong that ordinary BRT would be profitable, the corridor has high enough ridership for a subway.

Light rail is essentially a via media between metros and buses: higher operating costs than metros, in theory lower ones than buses. I say in theory, because in the United States, light rail as a mode comprises different things, some behaving like lower-efficiency subways with shorter cars like the Boston Green Lines, and others running as mostly grade-separated urban rail in cities like the Los Angeles and Portland cities with extremely low ridership and high resulting operating costs. But a light rail system with serious ridership should comfortably obtain better operating outcomes than buses, if worse ones than metros.

Costs per rider can be too low

In New York, as mentioned above, the current urban rail extensions under construction (Second Avenue Subway Phase 2) or discussion (Interborough Express) have costs not far from the frontier relative to American incomes. In Berlin, the extensions instead are far cheaper; U8 to Märkisches Viertel was projected to cost 13,160€ per daily rider in 2021, which is a fraction of Germany’s GDP per capita.

This does not mean Berlin builds cost-effectively. It means Berlin builds too little. A line that costs less than one third the country’s GDP per capita should have been built when the GDP per capita was one third what it is now. If there are a lot of such possibilities in the city, it means there was a crisis it’s only now recovering from or there has been too much austerity, or both, in the case of Berlin.

Healthy construction environments – that is, not Germany, which has normal costs per kilometer and chooses to barely build intercity or urban rail – will instead build to the frontier of what’s cost-effective. In New York, it’s Second Avenue Subway; in Madrid, it’s extensions into deep suburbia making the system almost as long as that of New York, on one third the metro area population. Rational yes/no decisions on whether to build at all can coexist with good construction practices or with deeply irrational ones.

More American Station Construction Extravaganza

Los Angeles has plans to extend its urban rail network. They’re taking forever, because construction costs are extremely high – and Nick Andert just pointed out one reason: the station caverns are huge. The overage on the proposed northern extension of the K Line, also known as the Crenshaw/LAX Line, is a good deal larger than on Second Avenue Subway, making it the most wasteful station construction that I am aware of. This is driving up the construction cost estimate to, depending on which alignment the K Line is to take as it goes north into West Hollywood, around $1 billion per kilometer.

Nick provides some station footprint diagrams. The K Line’s stations are designed for a maximum train length of three cars, or about 81 meters in total. The stations on the proposed extension, however, start at 124 meters when there is no crossover, or 50% overage, and most are 300 meters with crossovers. In other words, an underground light rail extension with trains less than half the length of New York City Subway trains is proposed to have stations about as long as those on Second Avenue Subway, which are already about twice as long as they need to be by New York standards. (In the low- and medium-cost countries for which I have this information, the overage is not 50%, but ranges between 3% and 17%.)

The 50% overage without crossovers is completely unjustifiable. But the crossovers, which turn the 50% into nearly 300% overage, are equally unjustifiable. It is not normal to build bored tunnel subways with so many crossovers, precisely because it’s expensive to blast caverns for them. Marco gives the example of Milan M4, which, counting the soon-to-open extension, has four crossovers in 15.5 km and 21 stations.

To this I can add that the Copenhagen Metro, built with bored tunnel with blasted caverns for crossover, has on the original line just two underground crossovers; the City Circle Line has only two as well, plus two on the M4 branch. There are more crossovers above ground, where it’s not so expensive to build them, but still less than one per station. This is a system designed for 24/7 operation; crossovers are required to allow trains to run on a single bidirectional track at night to permit maintenance, one track at a time. Without this constraint, even fewer crossovers are needed – only at the ends of the line, which includes the end of each operating segment if the line opens in stages.

If the K Line extension’s split between station and tunnel costs is similar to that of Second Avenue Subway Phase 1, 3:1, then shrinking the stations to the normal overage of a few percent would slash their cost by a factor of close to four, which would cut the line’s cost by more than half. This is what the extravaganza of crossovers is doing to Los Angeles and its ability to build mass transit infrastructure.

What it Means That Miri Regev Wants to Cancel Congestion Pricing

Minister of Transportation Miri Regev is trying to cancel the Tel Aviv congestion pricing plan, slated to begin operations in 2026. Congestion pricing is still planned to happen, but 2026 is already behind schedule due to delays on past contracts required to set up the gantries. The plan is still to go ahead and use the revenue to help fund the Tel Aviv Metro project, to comprise three driverless metro lines at regional scale to complement both the longer-range commuter trains and the shorter-range Light Rail system, which opened last year with a subway segment after several decades of design and construction. But Regev has wanted to cancel it since she became transportation minister early last year, and her latest excuse is the war, never mind that usually during war one raises taxes and aims to restrain private consumption such as personal vehicle driving.

I bring this up partly to highlight that Regev has not been a good minister; the civil servants at the ministry quickly found that she routinely bypasses them, makes decisions purely with her own political team, and sometimes doesn’t even inform them before making public announcements. More recently, she’s been facing corruption investigations, since much of the above behavior is not legal in Israel, a country where one says “the state” with a positive connotation that exists in French and German but not in English.

But more than that, I bring this up to highlight the contrast between Regev and New York Governor Kathy Hochul, who outright canceled the New York congestion pricing plan last month on no notice, weeks before it was about to debut rather than years. At a personal level, Regev is a worse person than Hochul. But Regev’s ability to cause damage is constrained by far stronger state institutions. The cabinet can collectively decide to cancel congestion pricing, in the same way a state legislature could repeal its own laws, but that would involve extensive open debate within the coalition. Thus, the ministry of finance already said that if the ministry of transportation is bowing out, it will have to take over the program, since it’s necessary for financing the metro, which is still on-budget; the civil servants at finance have long drawn ire from populists over their control over the budget, called the Arrangements Law, and unless the metro is formally canceled, the money will have to come from somewhere.

A formal repeal is still possible, but it cannot be done on a whim. Netanyahu, an atypically monarchical prime minister in both power within the coalition and aspirations, might be able to swing it if he wants, but he’d still have to persuade coalition partners. His power derives from long-term deals with junior parties that are so widely loathed they feel like they have to stick with him, and from over time turning Likud into a party of personal loyalists; at the same time, he has to govern roughly within the spectrum of opinions of the loyalists, and while their opinions on the biggest issues facing the state align with his or else they would not be in this position, on issues like transportation they may have different opinions and express them. At no point does a loyalist like Regev get absolute control of one aspect of policy; the coalition gets a vote and absent a formal repeal, the legislation creating congestion pricing still binds.

In other words, Israel is a functioning multiparty parliamentary democracy, more or less. Mostly less these days, let’s be honest. But much of the “less” comes from a concerted attempt to politicize the civil service, which Regev is currently under investigation for. In the United States, it’s fully politicized; one governor can announce the cancellation of a legislatively mandated congestion pricing program on a whim, the MTA board (which she appointed) will affirm that she indeed can do it and will not sign a statement saying the state consents to congestion pricing, and the question of whether it’s legal will be deferred to courts where the judges were politically appointed based on governor-legislative chief deals. Israel can make long-term plans, and a minister like Regev can interfere with them, but would need to do a lot of work to truly wreck them. The United States, as we’re seeing with New York congestion pricing, really can’t.