Why Does TGV Ridership Overperform Models?

I’ve found some TGV ridership data with which I can check the model I use for high-speed rail ridership projection. The model is trained on Japanese data and has flaws in Japan too, but I’ve wanted to see how well it ports to Europe, where I don’t have as complete a dataset of ridership between pairs of metropolitan areas. Thankfully, I have just found a short Omnil report from 2015 about intercity rail ridership at the Parisian terminals, from which we can extract some information. The TGV overperforms the model substantially; this can be explained with fudge factors, but those fudge factors only work if we assume that the aspects of the TGV that seasoned rail advocates hate don’t matter much.

The model

As a reminder, the model posits that the annual ridership in millions between two metropolitan areas with populations \mbox{Pop}_{A} and \mbox{Pop}_{B} in millions, of distance d kilometers, is,

75000\cdot\mbox{Pop}_{A}^{0.8}\cdot\mbox{Pop}_{B}^{0.8}/\max\{d^{2}, 250000\}

The model is very accurate for ridership between Tokyo and other cities on Honshu; it overpredicts inter-island ridership, but becomes correct if we replace the Japanese air/rail modal splits with European ones, where taking the train over five hours is more normal than in Japan. I would expect that in isolation, European ridership should overperform it, because fares here are much lower, about 0.10-0.11€/kilometer compared with about $0.23/kilometer on the Shinkansen. French ridership significantly overperforms, beyond what the fares alone can explain, as we will see.

We will need to modify the model as written above for the French case anyway. TGV ridership relies on direct through-service from Paris to every city in France, including many that are not on the network of dedicated high-speed lines (called LGVs); trains serve those by diverting from the LGVs to classical lines, on which they travel more slowly. Therefore, while we can apply the model as above for connections that entirely use LGVs, like Paris-Lyon or Paris-Marseille, we need to consider the slower speeds for connections that use classical lines. For those, we assume that trains average 220-225 km/h; this is the rough average speed of the express Shinkansen trains as well as that of the TGVs to Lyon and Marseille. Thus, the model, at travel time t, is,

1.5\cdot\mbox{Pop}_{A}^{0.8}\cdot\mbox{Pop}_{B}^{0.8}/\max\{t^{2}, 5\}

The floor of 500 km, or in this case a trip time of \sqrt{5} \sim 2:14 hours, is empirical in Japan. But then it is clear, from Italian data, that speeding up the trip has a roughly square-law effect on ridership, even within the limit – the growth in ridership on Bologna-Florence is consistent with an even higher elasticity of ridership with respect to average speed. The best way to reconcile these two observations is that in the presence of high-speed rail, the effect of distance cancels out the effect of better competition with the car up to about 500 km, but if the trains are slower, the car is more competitive and this is seen as a square law at all speeds. This is not too relevant to France, but is useful context for medium-distance, medium-speed lines in Germany.

TGV ridership

I have never been able to find city-to-city or station-to-station ridership figures in France. The Omnil report is no exception: it reports ridership at the Paris stations and breaks down where people are going by region of France in the geography of 2015, before the merger of some regions.

The total ridership at the Paris stations, including TGVs, low-speed intercity trains, and other regions’ regional trains (TERs) but not Paris-area regional trains (Transilien), is 443,000/day; of those, the TGVs comprise 239,000 and the slow trains 204,000. The four Parisian terminals with TGVs – Gare de Lyon, Gare du Nord, Gare Montparnasse, Gare de l’Est – have 92% of the TGV ridership in the region, while the other 8% are at suburban stations on bypasses around the city, like CDG Airport. Ridership is asymmetric: two-thirds of those 443,000 daily riders don’t live in Ile-de-France, which is what we should expect of a commuter-heavy ridership profile. Within Ile-de-France, 63% of passengers originate or are destined to Paris itself and another 21% for the Petite Couronne suburbs, showcasing destination centralization – Paris is only 17% of regional population and about 33% of regional employment, but 63% of those interregional and intercity trips go there and not to the suburbs.

There is also a breakdown of where passengers are connecting, by region of France or country. Picardie is increasingly an exurb of Paris, to the point that as France was debating the merger of regions in the early 2010s, one proposal was to detach its southernmost department, Oise, and attach it to Ile-de-France; 19% of the non-Francilien passengers originate there and 10% of Franciliens go there, for a ratio of nearly 4:1. More relevantly to high-speed rail, Rhône-Alpes is 9% of both non-Francilien and Francilien ridership, for a ratio of about 2:1, and a total of about 40,000/day, or around 13 million/year. PACA is 5% of non-Francilien and 7% of Francilien ridership, for a ratio of about 1.4:1 and a total of 25,000/day or around 8 million/year.

So we need to evaluate our model against an observed ridership of 13 million between Paris and Rhône-Alpes, and 8 million between Paris and PACA. Both sets of numbers involve multiple city pairs, with fairly long tails: France is a country of small metro areas, the median person living in a metro area of 330,000, whereas half of Japan lives in the metro areas of Tokyo (37 million), Osaka (18 million), and Nagoya (9 million).

French metro areas and the model

France recently changed its definition of metro areas. The old one, the aire urbaine, was similar in definition to the American metropolitan statistical area; the new one, the EU-wide functional area, generally spits out slightly larger numbers, though it still seems tighter than the Japanese definition. The functional area of Paris, comprising Ile-de-France, about half of Oise, and surrounding communes, has 13.2 million people. The new definition splits Nice and Cannes apart, which is good, since both have TGV service to France.

Metro cityPopulationTrip timePrediction
Lyon2.291:584.586
Grenoble0.723:010.999
Saint-Etienne0.52:580.771
Geneva (French part)*0.443:130.485
Annecy0.33:450.321
Chambéry0.262:520.49
Valence0.262:120.805
Bourg-en-Bresse0.141:500.49
Marseille1.883:072.02
Nice0.625:480.24
Toulon0.584:020.47
Cannes0.395:180.199
Avignon0.342:400.701
Geneva is deemed to have 1.2 million people, and the 0.44 million in the French part are imputed proportionally, rather than counted as a separate metro area, since there is no direct connection to Pari except via Geneva.

The Rhône-Alpes metro regions combine to a predicted ridership of 8.95 million; actual ridership is higher by about 50%. The PACA metro regions combine to a prediction of 3.63 million; actual ridership is higher by a factor of maybe 2.2.

Note that the prediction is already based on some optimistic assumptions. The trip time is the best that can be sustained multiple times a day; the issue of frequency is ignored, so the effective trip time on connections from Paris to cities like Annecy with a train every three hours gets no malus, even though the Japanese city pairs that the model is trained on get multiple express trains per hour. This is relevant, because as we examine fudge factors below to rescue the model, we need to keep ignoring or at best minimizing the malus due to poor frequency and lack of trip spontaneity in the ticketing system.

Fudge factors explaining the overperformance

We need to explain why Rhône-Alpes overperforms by 50%, and PACA by more than 100%.

Fares

The average JR East Shinkansen fare revenue in 2020-1 was ¥23.8/passenger-km (source, PDF-p. 50), and has risen little in the last 10 years. The average TGV fare revenue in 2019 was 0.10€/passenger-km (source, pp. 16 and 20) and has likewise little changed in nominal terms. These differ by a factor of 1.6. The elasticity of high-speed ridership with respect to price varies widely by study; the Italian study linked above says -0.37, one Spanish study says -0.59, and Börjesson’s lit review says -0.59 for non-business trips and -0.72 for business trips. A value of -0.5 explains a factor of 1.27 overperformance by itself; a value of -0.6 explains a factor of 1.33.

In fact, Germany, charging similar average intercity rail fares to France, seems to overperform the Shinkansen model too. I have little data here, only line-wide Berlin-Hamburg and Berlin-Munich, both of which look like they overperform by about 20%. This can result from a 30% overperformance mitigated by the issue of lower speed: the modeled prediction is based on trip times, but when trips are shorter than about 2:15, the model stops seeing the impact of slowdowns – Berlin-Hamburg is 1:44 and Berlin-Leipzig is 1:13, where at Shinkansen or TGV speeds they’d be 1:17 and 0:45 respectively.

Metro area size

French metro area definitions, even with the new functional areas, are somewhat tighter than Japanese ones. The functional area of Berlin has 5 million people, but reckoned the Japanese way (1.5% of the age 15+ population commuting to the central city), practically all of Brandenburg would count, a population of 5.7 million in total. This is likely more significant in PACA, where the above-listed metro area are 80% of the total population, than in Rhône-Alpes, where they are 90%. It’s possible even Paris is a bit bigger than 13.2 million – but only a bit, since Ile-de-France and Oise together only have 13.1 million. This factor can scrounge some extra ridership, but probably no more than 10%, maybe a bit more in PACA.

Leisure travel

Provence is renowned for its tourism, which generates extra trips out of Paris beyond what we should expect from population alone. This should disproportionately affect Nice and Cannes; for what it’s worth, I’m seeing seven weekday trains from those cities and Toulon to Paris, I believe all skipping Marseille, and 14 trains from Marseille; if we take ridership as proportional to the offer, this does show some Riviera overperformance relative to Marseille, though not by much.

Of course, the majority of Paris-PACA ridership comprises Provençals, not Franciliens. But perhaps the 1.4:1 ratio of Provençals to Franciliens is atypically low, and the 2:1 ratio in Rhône-Alpes is more normal of capital-province relations; I have no Japanese numbers on this, and would overall expect to see similar asymmetries in both countries, given their similar level of economic capital-centricity. If 2:1 is typical, then the extra leisure ridership from the capital to make it 1.4:1 adds a total of 14%, which is far less than PACA’s overperformance relative to Rhône-Alpes.

Metro area coverage

The PACA cities have multiple stops. The population distribution in the Riviera is linear, and multiple cities with extensive leisure (like Saint-Tropez) are served by the TGV. Marseille likwise has a second stop at Aix-en-Provence, close by car to its northern suburbs to the point that I’ve heard it called Marseille-bis. If we split metro Marseille’s population 2:1 between Marseille and Aix, then the 0.8 exponent in the model produces a 14% increase in ridership. 14% and another 14% from leisure combine to 31%, which explains the majority of the PACA overperformance relative to Rhône-Alpes.

Competition with air in small cities

The TGV competes with cars and planes; domestic buses are almost a non-factor, and were entirely a non-factor in 2015 (they’re called Macron buses because it was Macron, as minister of economics in 2014-6, who passed the reform that allowed them). In Rhône-Alpes, competition is entirely with the car: Lyon is just close enough to Paris that air travel can’t compete; in PACA, competition is mostly with the plane, especially beyond Marseille.

The population distribution in both Rhône-Alpes and PACA may favor the train. The issue is that the secondary cities of Rhône-Alpes are around three hours from Paris, at which point the train is strongly favored but planes normally still exist, as in Marseille. However, those cities are scattered all over the region, and so there is no single airport that could serve them, except Lyon – and if the choice is to take the train for three hour or to drive an hour to Lyon-Saint-Exupéry, then the train can just demolish air competition.

In PACA, the same is true for the secondary cities. Nice has a strong airport with many flights to Paris, buoyed by the leisure market, but Toulon and Avignon don’t; on the eve of corona, Toulon-Hyères had 500,000 passengers a year, most not bound for Paris.

I believe this effect on air-rail competition is more significant in Rhône-Alpes than PACA. However, air competition is overall more significant in PACA than in Rhône-Alpes, and thus it likely effects a similar boost to TGV ridership in both regions, or perhaps is more significant in PACA, explaining the remainder of its overperformance.

Some conclusions

I don’t think the TGV’s overperformance of the model invalidates the model. Most of the overperformance in Rhône-Alpes can be explained by fares alone, and I think the rest can be explained by the modal split versus air being more favorable than in Japan given the small size of Annecy, Saint-Etienne, and so on. Most of the overperformance in PACA relative to Rhône-Alpes can then be explained by leisure travel and the good metro area coverage of the TGV thanks to Aix and the linear population distribution of the Riviera. However, these fudge factors have implications for rail planning in France, Europe, and beyond.

Connections to smaller cities

The modeled prediction is that Lyon and Marseille comprise little more than half the ridership to Paris from their respective regions. Moreover, the overperformance of TGV riderhip relative to the hinkanssen model likely comes disproportionately from smaller cities, due to their lack of good air connections. This underscores the importance of good service not just to million-plus metro areas but also to the tail of metro areas of half a million, give or take. Those metro areas are less important in rich Asia or the US, but are important throughout Europe.

This service to smaller cities can take the characteritic of TGV-style direct connections to Paris on classical lines. In Switzerland, the Netherlands, Austria, and increasingly Germany, service to smaller cities is provided through timed connections at carefully-chosen nodes; the Swiss network particularly excels at this. But the French system’s ridership is such that it not obviously inferior, and is unlikely to be inferior to the German system at all. Thus, a country like Poland or Britain can safely choose between the French and German system, or even mix them.

The issue of frequency

The low frequency of TGV services to smaller cities – trains run every two to three hours, often timed to just miss regional trains – should be visible as a serious malus to ridership. But it isn’t. Perhaps it exists and countermands the effect of lack of air competition to cities the size and distance class of Grenoble – but Grenoble is not Nice, and air competition there even under more favorable scenarios to planes would be second-order.

At the same time, there are markets where the TGV is visibly much weaker. The TGV’s modal split between provincial regions is not good. Because trains from Paris to Marseille don’t stop at Lyon, and trains from Paris to Lyon don’t continue onward to Marseille, the Lyon-Marseille city pair cannot piggyback on strong connections to the capital the way same-side pairs of provincial Japanese cities can. The dedicated Marseille-Lyon trains have an inexplicable six-hour gap, with frequent service on both sides of it, and the Toulon-Lyon trains are even worse. The modal split is evidently weaker – in 2009, nearly everyone drove betwen Lyon and Toulon (the 2023 number in the link are speculation for what if an LGV is built to Nice), even over a rail-friendly distance of about 390 km, averaging around 130-150 km/h.

So while the system that centers direct trains to Paris is not suspect, the lack of frequency on shorter connections between secondary cities is. This could be resolved with buying rolling stock that makes boarding and alighting faster, with two door pairs per car rather than just one; TGV connections not including Paris run local, and since the trains are not optimized for many stops, those connections have low average speed, which in turn discourages SNCF from providing more frequent local connections.

Liberalization

The EU is increasingly forcing national railways to allow on-rail competition. This is an idea imported from the UK, where John Major’s privatization of British Rail split up operations and infratructure, the latter eventually renationlized; in Japan, privatization broke up JNR into regional JR companies, each responsible for both infrastructure and operations as in the pre-nationalization era of rail, and in the US, the breakup of Conrail likewise restored the pre-nationalization status quo. SNCF resists the mandate for competition in increaingly spiteful ways: it makes up excuses why RENFE can’t operate on its network, and where it does operate, it won’t even let its crew use break rooms at French stations. Eurocrats, even more progressive ones, treat SNCF as public enemy #1.

And SNCF’s anti-competitive monopoly on domestic rail travel generates high rail ridership. Italy and Spain have both seen sharp increases in ridership from the competition mandate. But Madrid-Barcelona, offering worse frequency and a more broken market than the domstic TGVs (domestic TGVs are split just between lower-price OuiGo and higher-price InOui brands; Spanish high-speed trains have more classes of train on thinner markets), don’t perform nearly so well. Madrid-Barcelona riderhip in 2019 was 4.4 million; the modeled prediction is 4.1 million for this city pair alone, and 6 million including intermediate trip to Zaragoza. Riderhip ha risen since the introduction of competition in 2020, and media coverage has been laudatory, and at times depreating of France for failing to liberalize – but the 50% growth in ridership cited in most articles still leave the line barely overperforming the high-fare Shinkansen and strongly unerperforming the TGV.

European media should be less credulous of promises of private-sector efficieny and recognize that the TGV’s model of public-sector monopoly, with integration between infraatructure and service (even if this means shoving direct trains to Paris on trunk line rather than building a Swiss integrated timed transfer system), produces better outcomes than competition. Germany has the same model too and, relative to how slow its trains are, has good outcomes too; Switzerland, the undisputed leader of European rail ridership, resists privatization entirely. Private competition did not invent high-speed rail, and where it has been introduced it has so far failed to produce outcomes on a par with what the TGV has with entirely public operations.

The Regional Connector and Network Coherence

The Regional Connector just opened in Los Angeles. This is a short, expensive tunnel permitting through-running for the city’s main light rail lines, linking the A Line (formerly Blue) with the half of L (formerly Gold) to Pasadena and Azusa and the E (formerly Expo) with the half of the L to East Los Angeles. It’s a welcome development and I only regret that this line cost, in 2022 dollars, $660 million per km. The broader question about the line, though, is that of line pairing and network coherence.

Source: The Source

Network coherence is a nebulous concept. I can best define it as “it looks reasonable on a map,” but then the number of railfan crayons that violate coherence principles is so large that it has to be defined more precisely. Usually I talk about it in two concrete ways: a rapid transit network should avoid reverse-branching and tangential lines.

But there are some more concerns, which are less obvious than those two, and also appropriate to violate in some cases. My previous post on New York alludes to the principle of consistency of local and express trains, which is good to aspire to but may conflict with other priorities. In fact, the current pattern in Los Angeles is coherent: trains go north-south on the new A Line or east-west on the new E Line; the question, posed to me by a supporter (thank you!), is whether it should be. I think this is right, but it’s not obvious (in fact it wasn’t to the reader who posed this question).

The issue is that simple coherence – lines from the south should go north and not east, express trains on one side should go through city center and then be express on the other side, and so on – is a good starting point. But there may be other concerns at play. When SEPTA opened the Center City Commuter Connection in the 1980s, it relabeled its commuter rail lines R1 through R8, skipping R4, pairing each branch on the former Pennsylvania Railroad side with one on the Reading side; the resulting map, for which this was an early plan, was not at all coherent and featured self-intersecting lines, but it matched the lines based on expected ridership and railyard locations.

The New York City Subway, likewise, has some pairings that only make sense in light of railyards. The R train shares local tracks in Manhattan with the W train, but in Queens it goes to the Queens Boulevard Line whereas the N shares tracks with the W to Astoria, because the Queens Boulevard Line has a yard at the end and the Astoria Line does not, and the N has a yard at the Brooklyn end of Coney Island whereas the R does not at Bay Ridge; the R is the longest all-local line in the system and railfans periodically propose to switch it with the N in Queens, but the railyard issue makes it problematic in the current situation.

Matching branches by ridership, though, runs into the problem of inconsistent headways on the trunk. A light rail line running every eight minutes can share a tunnel with another light rail line running every eight, but if one of them runs every six minutes, they will run into each other on the shared section. This problem is magnified if both the shared section and the branches are long, which is usually the case in New York; thankfully, in Los Angeles the shared section is short enough that optimizing on it is less important. At most, it’s possibly to match branches by ridership if the train lengths can be made different.

The final concern that may lead to violations of coherence is origin-and-destination ridership patterns; I think this is what my reader had in mind when querying it. Such patterns may be ethnic – there’s a notable bump in ridership between Anacostia and Columbia Heights on the Washington Metro, and it’s likely that such patterns also exist elsewhere in cities that just don’t track O&D pairings, like New York with its internal Chinatown buses. They may even be classed, in the sense that wealthier Americans are likelier to be working in city center, but on the whole, rich and poor people are mostly traveling to the same places. They may be sporadic – a university may be driving ridership enough that there could be connections to either places students want to go or (for commuter universities like UBC) where students live.

O&D pairings like this should, as a rule, never drive infrastructure decisions. These patterns are too fleeting – for example, Columbia Heights’ gentrification is so rapid that the links with Anacostia are unlikely to last for more than a generation. However, it’s not always bad to look at them when making decisions on service for a given infrastructure network.

The issue, though, is that there’s no real compulsion to connect Long Beach with East Los Angeles or Santa Monica with Pasadena and Azusa. The commute data doesn’t suggest special links (which, to be fair, it also doesn’t in Washington). I am not aware of any other big pattern that would create such ridership. So it’s something that I think Metro should have looked at (and probably did), but ultimately made the right decision on.

Assume Normal Costs: An Update

The maps below detail what New York could build if its construction costs were normal, rather than the highest in the world for reasons that the city and state could choose to change. I’ve been working on this for a while – we considered including these maps in our final report before removing them from scope to save time.

Higher-resolution images can be found here and here; they’re 53 MB each.

Didn’t you do this before?

Yes. I wrote a post to a similar effect four years ago. The maps here are updated to include slightly different lines – I think the new one reflects city transportation needs better – and to add light rail and not just subway and commuter rail tunnels. But more importantly, the new maps have much higher costs, reflecting a few years’ worth of inflation (this is 2022 dollars) and some large real cost increases in Scandinavia.

What’s included in the maps?

The maps include the following items:

  • 278 km of new streetcars, which are envisioned to be in dedicated lanes; on the Brooklyn and Queensborough Bridges, they’d share the bridges’ grade separation from traffic into Manhattan, which in the case of the Brooklyn Bridge should be an elevated version of the branched subway-surface lines of Boston, Philadelphia, San Francisco, and most German cities. Those should cost $8 billion in total, based on Eno’s European numbers plus some recent urban German projects.
  • 240 km of new subway lines, divided as 147.6 km in the city (97.1 km underground) and 92.4 km across the Hudson and in New Jersey (45.2 km underground). Of those 240 km, 147.5 km comprise four new trunk lines, including the already-planned IBX, and the rest are extensions of existing lines. Those should cost $25.2 billion for the city lines and $15.4 billion for the New Jersey lines.
  • Rearrangement of existing lines to reduce branching (“deinterlining“) and improve capacity and schedule robustness; the PATH changes are especially radical, turning the system into extensions of the 6 and 7 trains plus a Hoboken-Herald Square shuttle.
  • 48.2 km of commuter rail tunnels, creating seven independent trunk lines across the region, all running across between suburbs through the Manhattan core. In addition to some surface improvements between New York and Newark, those should cost $17.7 billion, but some additional costs, totaling low single-digit billions, need to be incurred for further improvements to junctions, station platforms, and electrification.

The different mode of transportation are intended to work together. They’re split across two maps to avoid cluttering the core too much, but the transfers should be free and the fares should be the same within each zone (thus, all trains within the city must charge MetroCard/OMNY fare, including commuter rail and the JFK AirTrain). The best way to connect between two stations may involve changing modes – this is why there are three light rail lines terminating at or near JFK not connecting to one another except via the AirTrain.

What else is included?

There must be concurrent improvements in the quality of service and stations that are not visible on a map:

  • Wheelchair accessibility at every station is a must, and must be built immediately; a judge with courage, an interest in improving the region, and an eye for enforcing civil rights and accessibility laws should impose a deadline in the early to mid 2030s for full compliance. A reasonable budget, based on Berlin, Madrid, and Milan, is about $10-15 million per remaining station, a total of around $4 billion.
  • Platform edge doors at every station are a good investment as well. They facilitate air conditioning underground; they create more space on the platform because they make it easier to stand closer to the platform edge when the station is crowded; they eliminate deaths and injuries from accidental falls, suicides, and criminal pushes. The only data point I have is from Paris, where pro-rated to New York’s length it should be $10 million per station and $5 billion citywide.
  • Signaling must be upgraded to the most modern standards; the L and 7 trains are mostly there already, with communications-based train control (CBTC). Based on automation costs in Nuremberg and Paris, this should be about $6 billion systemwide. The greater precision of computers has sped up Paris Métro lines by almost 20% and increased capacity. Together with the deinterlining program, a single subway track pair, currently capped at 24 trains an hour in most cases, could run about 40 trains per hour.
  • Improvements in operations and maintenance efficiency don’t cost money, just political capital, but permit service to be more reliable while cutting New York’s operating expense, which are 1.5-2.5 time a high as the norm for large first-world subway systems.

The frequency on the subway and streetcar lines depicted on the map must conform to the Six-Minute Service campaign demand of Riders Alliance and allies. This means that streetcars and subway branches run very six minute all day, every day, and subway trunk lines like the 6, 7, and L get twice as much frequency.

What alternatives are there?

Some decisions on the map are set in stone: an extension of Second Avenue Subway into Harlem and thence west along 125th Street must be a top priority, done better than the present-day project with it extravagant costs. However, others have alternatives, not depicted.

One notable place where this could easily be done another way is the assignment of local and express trains feeding Eighth and Sixth Avenues. As depicted, in Queens, F trains run local to Sixth Avenue and E trains run express to Eighth; then, to keep the local and express patterns consistent, Washington Heights trains run local and Grand Concourse trains run express. But this could be flipped entirely, with the advantage of eliminating the awkward Jamaica-to-Manhattan-to-Jamaica service and replacing it with straighter lines. Or, service patterns could change, so that the E runs express in Queen and local in Manhattan as it does today.

Another is the commuter rail tunnel system in Lower Manhattan. There are many options for how to connect New Jersey, Lower Manhattan, and Brooklyn; I believe what I drew, via the Erie Railroad’s historic alignment to Pavonia/Newport, is the best option, but there are alternatives and all must be studied seriously. The location of the Lower Manhattan transfer station likewise requires a delicate engineering study, and the answer may be that additional stops are prudent, for example two stops at City Hall and South Ferry rather than the single depicted station at Fulton Street.

What are those costs?

I encourage people to read our costs report to look at what goes into the numbers. But, in brief, we’ve identified a recipe to cut New York subway construction costs by a factor of 9-10. On current numbers, this means New York can cut its subway construction costs to $200-250 million per kilometer – a bit less in the easiest places like Eastern Queens, somewhat more in Manhattan or across water. Commuter rail tunnel costs are higher, first because they tend to be built only in the most difficult areas – in easier ones, commuter rail uses legacy lines – and second because they involve bigger stations in more constrained areas. Those, too, follow what we’ve found in comparison cases in Southern Europe, the Nordic countries, Turkey, France, and Germany.

In total, the costs so projected on the map, $66.3 billion in total, are only slightly higher than the total cost of Grand Paris Express, which is $60 billion in 2022 dollars. But Paris is also building other Métro, RER, and tramway extensions at the same time; this means that even the program I’m proposing, implemented over 15 years, would still leave New York spending less money than Paris.

Is this possible?

Yes. The governance changes we outline are all doable at the state level; federal officials can nudge things and city politician can assist and support. There’s little confidence that current leadership even wants to build, let alone knows what to do, but it’s all doable, and our report linked in the lede provides the blueprint.

How Fast is Necessary?

Swiss intercity rail planning follows the maxim “run as fast as necessary, not as fast as possible.” This is usually uttered in opposition to high-speed rail in the sense of the French TGV network. But what does this slogan really mean? And how does it inform good planning?

The issue of timed connections

The origin of the Swiss planning system is in the 1970s and 80s, as it was refining intercity rail, taking what West Germany was doing with its InterCity brand and going further. Two key elements were present from the start: timed connections, and regular clockface timetables (initially every two hours in West Germany). The clockface timetabling facilitated the timed connections, since it’s easier to figure out how to schedule a timed transfer at (say) Frankfurt if the same train moves happen at the same intervals.

With these elements in place, Swiss Federal Railways (SBB) set up a schedule in which trains would be timed to arrive in Zurich all at the same time, currently just before the hour, and depart all at the same time, currently on or just after the hour.

The issue is what to do at stations other than Zurich. Optimizing for timed connections at Zurich means compromising on the question of what to do at other cities. If trains arrive and depart Zurich roughly on the hour, then the terminal at the other end may have ugly arrival and departure times – for example, a 1:40 trip time, with Zurich arrival :56 and departure :04, would mean that the other end has an arrival time of :44 and a departure time of :16. If it’s an hourly train, it means it’s not possible to time connections there – an arriving bus or regional train would have a transfer time of perhaps 20 minutes, which in a country the size of Switzerland is a large share of the overall trip.

And then there is the looming issue of intercity connections. Zurich is located fairly centrally in Switzerland, but there are some key connections that don’t go anywhere near it, led by Bern-Basel via Olten. Bern’s central location makes it a great node for timed connections as well. However, the Zurich-Bern trips took 1:09, making it impossible to have timed everywhere-to-everywhere connections in both cities.

Speeding up trains to make connections

In the 1980s, the Zurich-Bern trips took 1:09, so connections could only be timed in Zurich, not Bern. At the time, there were plans for a French-style high-speed rail network connecting Zurich, Bern, and Geneva, but those plans were canceled due to high costs relative to Switzerland’s size.

Instead of running as fast as possible, enough to connect Zurich and Geneva in perhaps 1:30, trains would run as fast as necessary, just enough to make connections. The centerpiece of this plan, dubbed Rail 2000, was to speed up Olten-Bern by just enough to shorten Zurich-Bern and Basel-Bern to 0:56 and 0:55 respectively. This way, trains could arrive in all three cities just before the hour and depart just after, facilitating more timed connections.

This system was a resounding success. Swiss rail ridership has been sharply rising in the last 20 years, from an already fairly high level; by all metrics I am aware of, such as modal split, ridership per capita, and passenger-km per capita, it is Europe’s strongest rail network. More speedups are planned, all aiming to add more points where trains can be timed to connect, called knots, till the network looks like this:

The shape of Switzerland

The notion of running trains as fast as necessary is in one sense a global principle. But its surface manifestation of a system designed as an alternative to high-speed rail is a product of Swiss geography; in Japan, the current speed of the Shinkansen is also as fast as necessary. Switzerland’s current knot diagram has to be understood in the context of the following geographical features:

  • Switzerland is small enough that the strongest trunk corridors, like Zurich-Bern, can support just a train every half hour.
  • Switzerland is also physically small enough that the Zurich-Bern-Basel triangle has legs of about 110 km, which an upgraded rail system can connect in an hour minus transfer time, and which it is very hard to speed up to 30 minutes minus transfer time.
  • Switzerland is two-dimensional: there is no central trunk through which all service must pass – the diagram above depicts Zurich-Olten as a key link, but Luzern, though not yet a knot, connects separately to both Zurich and Bern.
  • Swiss train stations have a surplus of tracks, while still functioning as central stations, without the separation into different stations for different directions that Paris and London have.

Switzerland is not unique in having these features. The Netherlands is the same: it’s a small two-dimensional country with demand for many connections to be timed. The Netherlands built a 300 km/h high-speed rail line between Amsterdam, Rotterdam, and the border with Belgium toward Antwerp, Brussels, and Paris, but this line is not important to its intercity rail planning, which instead lives on Swiss-style knots.

However, not even all small countries are like Switzerland and the Netherlands. Israel is one- rather than two-dimensional: all trains pass through Tel Aviv and the Ayalon Railway, and there is no real need for timed connections, as the most important city pair not involving Tel Aviv, Haifa-Jerusalem, passes through Tel Aviv, with through-service to begin as soon as electrification is completed. Taiwan is so linear that a single high-speed rail line connects all of its cities, with higher ridership per capita than in any other country I have data for except maybe Japan; even before high-speed rail, the shape of its mainline network was a single line on the west coast connecting the cities and another on the east coast connecting smaller settlements.

And then in large countries, running as fast as necessary requires choosing a speed. Germany is two-dimensional like Switzerland, but has a string of city pairs 90-100 km apart – Hanover-Bielefeld-Dortmund, or Mannheim-Stuttgart, or the Fulda-Frankfurt-Würzburg triangle, or Würzburg-Nuremberg – for which running a fast as necessary entails a choice between connections in 60, 45, or 30 minutes. Longer-distance city pairs, like Nuremberg-Munich, Berlin-Leipzig, and the Hamburg-Hanover-Bremen triangle, are even less constrained.

Capacity

The Swiss network is based on the idea that near major stations like Zurich, trains should arrive and depart on pulses every 30 minutes. This way, the Olten-Bern line runs eight intercity trains per hour, but they are not equally spaced: they are timetabled in two platoons over a period of seven minutes each, with the off time used for regional trains not participating in the knot system.

Even in the Netherlands, this isn’t quite tenable. Amsterdam-Eindhoven trains come at regular 10-minute interval, each third train requiring a connection at Utrecht and the other two trains in three running directly. In Israel, Tel Aviv-Haifa trains run at 30-minute intervals midday, but for two four-hour peak periods each day this is boosted to four trains per hour – and this is even before electrification has been completed.

The situation in Germany deserves especial mention. Germany is investing in it intercity rail network in a way that sometimes tries to be Switzerland except bigger (such as the Deutschlandtakt and other measures supported by technical advocates) and sometimes tries to build high-speed rail lines and through-stations like Stuttgart 21. The technical advocates dislike Stuttgart 21 and argue that it’s un-Swiss to timetable intercity trains throughout the hour rather than in two pulses with inactivity between them. But Germany has large enough cities that it can’t afford to keep intercity lines out for so much time. Already, with mediocre speeds, the Deutschlandtakt plan for 2030 prescribes 4.5 trains an hour between Frankfurt and Mannheim and between Frankfurt and Cologne. As more high-speed lines come online, demand will grow; Deutsche Bahn projects to double ridership between 2019 and the 2030s, which will force the busiest links to operate a train every five to 10 minutes.

Is Swiss planning useful outside Switzerland and the Netherlands?

Unambiguously, yes. However, it would look different.

The best place to see how different it should be is, naturally, the Northeastern United States, my area of research. None of the features that made Rail 2000 work is present there: the region is large and has huge cities, is one- rather than two-dimensional, and has capacity-constrained stations forcing round-the-hour use of every approach track.

What this means is that there is little optimization from running slower than as fast as possible on the Northeast Corridor trunk line. However, running a fast as necessary remains a solid planning maxim on all the branches that connect to it, with their own timed connections to one another and to local buses in secondary cities like Worcester and Springfield. Most trains between Boston and Washington should run as express as practical based on station track speeds, and the local trains may plausibly only run every half hour, making them ideal for a system of timed half-hourly connections.

Large Urban Terminals are Unnecessary

London and Paris are both famous for their elaborate rail terminals, each with many tracks to store trains connecting the capital with one direction of the country’s provinces. This leads rail advocates to romanticize this type of station; Jarrett Walker said in 2009 that these stations make the arriving passengers feel grander for, in theory, being able to walk directly from platform to station entrance without changing levels. Earlier this month, Ido Klein generalized from London and Paris to the rest of Europe, arguing that the setup in Tel Aviv, with only through-tracks on the Ayalon Railway and only six platform tracks at Savidor Center, is inadequate. Ido is incorrect on this; the Israel Railways network needs continued capital investment in the Ayalon Railway and not in alternatives such as more terminating tracks or a bypass.

What is the situation in Israel?

The Israel Railways network, owing to Israel’s small size and lack of connections to other countries, is in effect an interregional system, in which the most common intercity trips are 60 km (Tel Aviv-Jerusalem) or 90 km (Tel Aviv-Haifa). Due to the country’s narrow width, it is entirely centered on Tel Aviv, where the four stations of the Ayalon Railway – University, Savidor Center, HaShalom, HaHagana – are on pre-corona numbers the four busiest in the country; this leads to much political distaste, as everywhere else in the Israeli economy.

Blue lines are in operation for passenger service; light blue denotes electrified lines as of last year, with a few additional segments having been wired since. Green lines are freight-only. Gray lines are abandoned. Source: Wikipedia.

The Ayalon Railway has three tracks, with plans for a fourth in the design phase; capacity is said to be around 14 trains per hour per track, even with the latest ETCS signals, and I’ve been told of at least one Israel Railways manager who disbelieved that the Munich S-Bahn could push 30 trains per hour per track. The main station in construction is Savidor Center, which has six platform tracks, but in ridership it is second to HaShalom, which is better-located in the Tel Aviv central business district and has only three tracks. Practically all trains run through Tel Aviv – only the trains to Jerusalem terminate in Tel Aviv.

Do trains need to terminate in or bypass Tel Aviv?

No.

There are plans for a bypass railway east of the city, but they’re largely for political reasons, signifying that the state supports decentralizing economic geography away from Tel Aviv. The most important city pair not involving Tel Aviv, Haifa-Jerusalem, does not get any faster via the bypass, and will be direct via the Ayalon Railway as soon as electrification is completed up to Haifa. The city pair that could most gain from a Tel Aviv bypass, Jerusalem-Beer Sheva, has its most direct route passing through Bethlehem and Hebron, which besides being in the Territories (and in Area A) have terrain from hell.

Terminal tracks are likewise useless. Demand from Tel Aviv to points north and south is fairly symmetric on both intercity and commuter rail; there was some asymmetry before the new Tel Aviv-Jerusalem line opened, but at this point Jerusalem’s one station has elbowed its way to the number four position, narrowly ahead of Tel Aviv University, and together with Beer Sheva, Jerusalem forms a fine counterpart to Haifa.

But what about Europe?

Ido’s thread goes over the largest 10 metropolitan areas in the EU and UK, and looks at their intercity rail stations, which he defines as stations serving lines of at least 200 km in length. Among those, he finds that London and Paris only have terminals (their through-stations are for regional rail), and the same is true of Milan, Rome, and Athens. However, from that point, things fall apart.

First, Madrid, Barcelona, and Warsaw have intercity rail through-stations; Ido incorrectly says they are not. Madrid Atocha is a through-station; there’s little to no AVE through-service there or at Chamartín, but they are both through-stations, and the medium-speed Alvias, which are faster than anything in Israel, run through routinely. Barcelona-Sants has AVE through-service between Madrid and France; most trains terminate, but that’s because of asymmetric demand, not because it’s inherently better. Warsaw has a through-tunnel and many through-running intercity trains listed as serving Warszawa Centralna.

Second, Ido skips over Germany. He portrays Berlin as atypical for having a through-station at Hauptbahnhof, but the Berlin way is what Germany wishes were the norm for all cities. Smaller German cities either have through-stations, like Hanover or the cities of the Rhine-Ruhr, or act as pinch-points, with through-trains coming in and reversing direction to continue onward, including Leipzig, Frankfurt, and (until Stuttgart 21 opens) Stuttgart. The pinch-point operations are as efficient as they can be, but still occupy more platform and approach slots than through-trains would.

Third, the actual practice of Paris and London is that it’s assumed people only travel between the capital and a provincial city. France and the UK do not have good everywhere-to-everywhere trains. Paris has a bypass, the Interconnexion Est, but the service quality on it is terrible: the operating paradigm is that trains that bypass Paris make every intermediate stop, which takes 5-10 minutes per station, as the TGVs are not designed for fast boarding and alighting at intermediate points but for nonstop Paris-provincial city trips.

The way forward

Province-province trips are difficult to serve with high frequency. Therefore, the best practice for them is to run through the main city if possible. Israel can do it, using the Ayalon Railway, and once electrification provides through-service from Haifa to Jerusalem, this city pair can piggyback on the higher demand of Haifa-Tel Aviv and Tel Aviv-Jerusalem to serve passengers frequently. Israel is famously small; Haifa-Jerusalem is around 150 km and 1:36 with upcoming speedups planned for electrification and other investments – trains have to run at worst every 20-30 minutes to avoid throwing away ridership, and this can only be supported if they run through Tel Aviv.

London and Paris have many rail terminals because they were huge cities in the steam era and private railroads figured they should connect the capital with one section of the country. This inherited infrastructure is a liability to both of their respective national rail networks, especially that of France, where Paris is centrally located and could be the center of a Lille-Marseille spine. Israel’s newer network lacks this seam, and this should be celebrated and form the basis of further investment.

The most important investment is to ensure that the Ayalon Railway can run at decent capacity. Electric multiple units on regional and interregional rail systems with more complexity than that of Israel Railways do much better than 14 trains per hour on each track: Zurich does 16 and is (I believe) capable of 24, Munich does 30, Tokyo does 24 on some commuter lines with comparable length to Israel’s intercity lines. This is not a problem of the signaling system, which has been upgraded to ETCS Level 2 at the same time as the electrification project. Rather, it’s a problem of how the trains are timetabled, and possibly also of infrastructure on the commuter rail branches, some of which are still single-track.

If it’s possible to cancel the Eastern Railway plans, it should be canceled. There isn’t much that’s being served on the way, and the split in frequency between Haifa-Tel Aviv, Tel-Aviv Jerusalem, and Haifa-Jerusalem trains would seriously hurt ridership on the last of the three city pairs. The fourth track on Ayalon Railway is useful, but long-term plans to go up to six tracks should be shelved – a four-track electrified line could support a large multiple of current traffic.

Can Intercity Trains into Boston Enter from Springfield?

From time to time, I see plans for intercity rail service into Boston going via Springfield. These include in-state rail plans to run trains between the two cities, but also grander plans to have train go between Boston and New Haven via Springfield, branded as the Inland Route, as an alternative to the present-day Northeast Corridor. In-state service is fine, and timed connections to New Haven are also fine for the benefit of interregional travel like Worcester-Hartford, but as an intercity connection, the Inland Route is a terrible choice, and no accommodation should be made for it in any plans. This post goes over why.

What is the Inland Route?

Via Wikipedia, here’s a map of the Northeast Corridor and connecting passenger rail lines:

Red denotes Amtrak ownership, and thus some non-Northeast Corridor sections owned by Amtrak are included, whereas the New Rochelle-New Haven section, while part of the corridor, is not in red because it is owned by state commuter rail authorities. Blue denotes commuter rail lines that use the corridor.

The Inland Route is the rail route in red and black from New Haven to Boston via Springfield. Historically, it was the first all-rail route between New York and Boston: the current route, called the Shore Line, was difficult to build with the technology of the 1840s because it required many river crossings, and only in 1889 was the last river bridged, the Thames just east of New London. However, as soon as the all-rail Shore Line route opened, mainline traffic shifted to it. Further investment in the Shore Line relegated the Inland Route to a secondary role, and today, the only passenger rail at all between Boston and Springfield comprises a daily night train to Chicago, the Lake Shore Limited. More recently, there has been investment in New Haven-Springfield trains, dubbed the Hartford Line, which runs every 1-2 hours with a few additional peak trips.

What rail service should run to Springfield?

Springfield is a secondary urban center, acting as the most significant city in the Pioneer Valley region, which has 700,000 people. It’s close to Hartford, with a metro population of 1.3 million, enough that the metro areas are in the process of merging; this is enough population that some rail service to both New York and Boston is merited.

In both cases, it’s important to follow best practices, which the current Hartford Line does not. I enumerated them for urban commuter rail yesterday, and in the case of intercity or interregional rail, the points about electrification and frequency remain apt. The frequency section on commuter rail talks about suburbs within 30 km of the city, and Springfield is much farther away, so the minimum viable frequency is lower than for suburban rail – hourly service is fine, and half-hourly service is at the limit beyond which further increases in frequency no longer generate much convenience benefit for passengers.

It’s also crucial to timetable the trains right. Not only should they be running on a clockface hourly (ideally half-hourly) schedule, but also everything should be timed to connect. This includes all of the following services:

  • Intercity trains to Hartford, New Haven, and New York
  • Intercity trains to Boston
  • Regional trains upriver to smaller Pioneer Valley cities like Northampton and Greenfield (those must be at least half-hourly as they cover a shorter distance)
  • Springfield buses serving Union Station, which acts as a combined bus-rail hub (PVTA service is infrequent, so the transfers can and should be timed)

The timed connections override all other considerations: if the demand to Boston and New York is asymmetric, and it almost certainly is, then the trains to New York should be longer than those to Boston. Through-running here is useful but not essential – there are at least three directions with viable service (New York, Boston, Greenfield) so some people have to transfer anyway, and the frequency is such that transfers have to be timed anyway.

What are the Inland Route plans?

There are perennial plans to add a few intercity trains on the New Haven-Springfield-Boston route. Some such trains ran in my lifetime – Amtrak only canceled the last ones in the 2000s, as improvements in the Shore Line for the Acela, including electrification of the New Haven-Boston section, made the Inland Route too slow to be viable.

Nonetheless, plans for restoration remain. These to some extent extend the plans for in-state Boston-Springfield rail, locally called East-West Rail: if trains run from Boston to Springfield and from Springfield to New Haven, then they might as well through-run. But some plans go further and posit that this should be a competitive end-to-end service, charging lower fares than the faster Northeast Corridor. Those plans, sitting on a shelf somewhere, are enough that Massachusetts is taking them into account when designing South Station.

Of note, no modernization is included in these plans. The trains are to be towed by diesel locomotives, and run on the existing line. Both the Inland Route and East-West plans assume frequency is measured in trains per day, designed by people who look backward to a mythologized golden age of American rail and not forward to foreign timetabling practices that have only been figured out in the last 50 years.

Is the Inland Route viable as an intercity route into Boston?

No. This is not even a slag on the existing plans; I’m happy assuming best practices in other cases, hence my talk of timed half-hourly connections between trains and buses above. The point is that even with best practices, there is no way to competitively run a New Haven-Springfield-Boston route.

The graphic above is suggestive of the first problem: the route is curvy. The Shore Line is very curvy as well, but less so; it has a bad reputation because its curves slow trains that in theory can run at 240 km/h down to about 150-180 km/h, but the Boston-Springfield Line has tighter curves over a longer stretch, they’re just less relevant now because the trains on the line don’t run fast anyway. In contrast, the existing Northeast Corridor route is fast in Massachusetts and Rhode Island.

The Inland Route is also curvy on the Boston-Worcester stretch, where consideration for slow trains is a must. The main way to squeeze extra speed out of a curvy line is to cant it, but this is less viable if there is a mix of fast and slow trains, since slow trains would be overcanted. This, in fact, is the reason Amtrak trains outside the Northeast are slower than they were in the middle of the 20th century – long-distance passenger trains have less priority for infrastructure design than slow freight trains, and so cant is limited, especially when there are hills. Normally, it’s not a problem if the slower trains are commuter trains, which run fast enough that they can just take the curve, but some curves are adjacent to passenger train stations, where passengers would definitely notice the train sitting still on canted track, leaning to the inside of the curve.

Then, there is the issue of how one gets into Boston. The Providence Line is straight and fast and can be upgraded to provide extra capacity so that fast intercity trains can overtake slow ones if need be. The Worcester Line has a two-track narrows in Newton, hemmed by I-90 with no possibility of expansion, with three stations on this stretch and a good location for a fourth one at Newton Corner. Overtakes are possible elsewhere (one is being designed just to the west, in Wellesley – see my sample timetable here), but they still constrain capacity. It’s comparably difficult from the point of view of infrastructure design to run a 360 km/h intercity train every 15 minutes via Providence and to run a 160 km/h intercity train every 30 minutes via Springfield and Worcester. Both options require small overtake facilities; higher frequency requires much more in both cases.

The Worcester Line is difficult enough that Boston-Springfield trains should be viewed as Boston-Worcester trains that go farther west. If there’s room in the timetable to include more express trains then these can be the trains to Springfield, but if there’s any difficulty, or if the plan doesn’t have more than a train every half hour to Worcester, then trains to Springfield should be making the same stops as Boston-Worcester trains.

Incentives for passengers

The worst argument I’ve seen for Inland Route service is that it could offer a lower-priced alternative to the Northeast Corridor. This, frankly, is nuts.

The operating costs of slower trains are higher than those of faster trains; this is especially true if, as in current plans, the slow trains are not even electrified. Crew, train maintenance, and train acquisition costs all scale with trip time rather than trip distance. Energy costs are dominated by acceleration and deceleration cycles rather than by cruise speed at all speeds up to about 300 km/h. High-speed trains sometimes still manage lower energy consumption per seat-km than slow trains, since slow trains have many acceleration cycles as track speeds change between segments whereas high-speed lines are built for consistent cruise speed.

The only reason to charge less for the trains that are more expensive to operate is to break the market into slow trains for poor people and fast trains for rich people. But this doesn’t generate any value for the customer – it just grabs profits through price discrimination that are then wasted on the higher operating costs of the inferior service. It’s the intercity equivalent of charging more for trains than for buses within a city, which practice is both common in the United States and a big negative to public transit ridership.

If, in contrast, the goal is to provide passengers with good service, then intercity trains to Boston must go via Providence, not Springfield. It’s wise to keep investing in the Shore Line (including bypasses where necessary) to keep providing faster and more convenient service. Creating a class system doesn’t make for good transit at any scale.

Through-Running and American Rail Activism

A bunch of us at the Effective Transit Alliance (mostly not me) are working on a long document about commuter rail through-running. I’m excited about it; the quality of the technical detail (again, mostly not by me) is far better than when I drew some lines on Google Maps in 2009-10. But it gets me thinking – how come through-running is the ask among American technical advocates for good passenger rail? How does it compare with other features of commuter rail modernization?

Note on terminology

In American activist spaces, good commuter rail is universally referred to as regional rail and the term commuter rail denotes peak-focused operations for suburban white flighters who work in city center and only take the train at rush hour. If that’s what you’re used to, mentally search-and-replace everything I say below appropriately. I have grown to avoid this terminology in the last few years, because in France and Germany, there is usually a distinction between commuter rail and longer-range regional rail, and the high standards that advocates demand are those of the former, not the latter. Thus, for me, a mainline rail serving a metropolitan area based on best practices is called commuter and not regional rail; there’s no term for the traditional American system, since there’s no circumstance in which it is appropriate.

The features of good commuter rail

The highest-productivity commuter rail systems I’m aware of – the Kanto area rail network, the Paris RER, S-Bahns in the major German-speaking cities, and so on – share certain features, which can be generalized as best practices. When other systems that lack these features adopt them, they generally see a sharp increase in ridership.

All of the features below fall under the rubric of planning commuter rail as a longer-range subway, rather than as something else, like a rural branch line or a peak-only American operation. The main alternative for providing suburban rapid transit service is the suburban metro, typical of Chinese cities, but the suburban metro and commuter rail models can coexist, as in Stockholm, and in either case, the point is to treat the suburbs as a lower-density, longer-distance part of the metropolitan area, rather than as something qualitatively different from the city. To effect this type of planning, all or nearly all of the following features are required, with the names typically given by advocates:

  • Electrification/EMUs: the line must run modern equipment, comprising electric multiple units (self-propelled, with no separate locomotive) for their superior performance and reliability
  • Level boarding/standing space: interior train design must facilitate fast boarding and alighting, including many wide doors with step-free boarding (which also provides wheelchair accessibility) and ample standing space within the car rather than just seated space, for example as in Berlin’s new Class 484
  • Frequency: the headway between trains set at a small fraction of the typical trip time – neighborhoods 10 km from city center warrant a train every 5-10 minutes, suburbs 20-30 km out a train every 10-20 minutes, suburbs farther out still warrant a train every 20-30 minutes
  • Schedule integration: train timetables must be planned in coordination with connecting suburban buses (or streetcars if available) to minimize connection time – the buses should be timed to arrive at each major suburban station just before the train departs, and depart just after it arrives
  • Pedestrian-friendliness: train stations designed around connections with buses, streetcars if present, bikes, and pedestrian activity – park-and-rides are acceptable but should be used sparingly, and at stations in the suburbs, the nearby pedestrian experience must come first, in order to make the station area attractive to non-drivers
  • Fare integration/Verkehrsverbund: the system may charge higher fares for longer trips, but the transfers to urban and suburban mass transit must be free even if different companies or agencies run the commuter trains and the city’s internal bus and rail system
  • Infill: stations should be spaced regularly every 1-3 km within the built-up area, including not just the suburbs but also the city; slightly longer stop spacing may be acceptable if the line acts as an express bypass of a nearby subway line, but not the long stretches of express running American commuter trains do in their central cities
  • Through-running: most trains that enter city center go through it, making multiple central stops, and then emerge on the other side to serve suburbs in that direction

Is through-running special?

Among the above features, through-running has a tendency to capture the imagination, because it lends itself to maps of how the lines fit together in the region; I’ve done more than my share of this, in the 2009 post linked in the intro, in 2014, in 2017, and in 2019. This is a useful feature, and in nearly every city with mainline rail, it’s essential to long-term modernization; the exceptions are cities where the geography puts the entirety of suburbia in one direction of city center, and even there, Sydney has through-running (all lines go west of city center) and Helsinki is building a tunnel for it (all lines go north).

The one special thing about through-running is that usually it is the most expensive item to implement, because it requires building new tunnels. In Philadelphia, this was the Center City Commuter Connection, opened in 1984. In Boston, it’s the much-advocated for North-South Rail Link. In Paris, Munich, Tokyo, Berlin, Copenhagen, London, Milan, Madrid, Sydney, Zurich, and other cities that I’m forgetting, this involved building expensive city center tunnels, usually more than one, to turn disparate lines into parts of a coherent metropolitan system. New York is fairly unique in already having the infrastructure for some through-running, and even there, several new tunnels are necessary for systemwide integration.

But there are so many other things that need to be done. In much of the United States, transit advocacy has recently focused on the issue of frequency, brought into the mainstream of advocacy by Jarrett Walker. Doing one without the other leads to awkward situations: after opening the tunnel, Philadelphia branded the lines R1 through R8 modeled on German S-Bahns while still running them hourly off-peak, even within the city, and charging premium fares even right next to overcrowded city buses.

This is something advocates generally understand. There’s a reason the TransitMatters Regional Rail program for commuter rail modernization puts the North-South Rail Link on the back burner and instead focuses on all the other elements. But there’s still something about through-running that lends itself to far more open argumentation than talking about off-peak frequency. Evidently, the Regional Plan Association and other organizations keep posting through-running maps rather than frequency maps or sample timetables.

Through-running as revolution

I suspect one reason for the special place of through-running, besides the attractiveness of drawing lines on a map, is that it most blatantly communicates that this is no longer the old failed system. There are good ways of running commuter rail, and bad ways, and all present-day American commuter rail practices are bad ways.

It’s possible to make asks about modernization that don’t touch through-running, such as integrating the fares; in Germany, the Verkehrsverbund concept goes back to the 1960s and is contemporary with the postwar S-Bahn tunnels, but Berlin and Hamburg had had through-running for decades before. But because these asks look small, it’s easy to compromise them down to nothing. This has happened in Boston, where there’s no fare integration on the horizon, but a handful of commuter rail stations have their fares reduced to be the same as on the subway, still with no free transfers.

Through-running is hard to compromise this way. As soon as the lines exist, they’re out there, requiring open coordination between different railroads, each of which thinks the other is incompetent and is correct. It’s hard to sell it as nothing, and thus it has to be done as a true leap generations forward, catching up with where the best places have been for 50+ years.

Quick Note: Heavy Touch and Control

There’s a distinction between light- and heavy-touch forms of management and control. Light-touch systems try to stay out of the details as far as they can; heavy-touch systems do the opposite. American business culture considers light touch to be superior, and I think this is especially prominent in the public sector, which has some ready-made examples of how the light touch approach works better – for example, in the military, it’s called mission command and is repeatedly shown to work better than more centralized command-and-control. Unfortunately, the same does not work for rail infrastructure. Why?

Heavy touch in infrastructure

In practice, a heavy-touch system in infrastructure construction, for example the way Germany, France, and Southern Europe work, has the following features:

  • The state agency maintains control of designs, and even when it outsources something to consultants, it owns the product and may tweak it or assume that future contractors will tweak it. There is little privatization of planning.
  • There is reluctance to devolve decisions to local governments; if SNCF or RENFE lets a regional government get involved in a rail plan, it’s because it’s an unprofitable regional line and the national railroad would rather not know it exists, and even DB happily unloads these same unprofitable regional lines on Land governments while focusing on intercity rail.
  • Regulators are technical and make specific decisions.

This is not the only way to organize things, but it’s the only way that works. The Nordic countries have been moving away from this system in the last generation, influenced by British governing ideology; the sources I reference in the Stockholm case study repeatedly treat privatization as self-evidently good and exhort Nordic agencies to be more like the UK and less like Germany and Switzerland, and meanwhile, in the last 20 years Nordic costs have exploded while German ones have been fairly stable.

I’ve talked about the issue of privatization of the state to consultants many times, most recently a month ago. This post is about something different: it’s about how regulators work, an issue on which Sweden appears little different from Germany to me, and profoundly different from the United States and its can’t-do government.

American light touch elements

To understand how American regulations work, we need to look at the regulators and grant funders, that is, the Federal Transit Administration and Federal Railroad Administration, henceforth abbreviated FTA and FRA. How do they work?

  • The chief regulators are never especially technical. The most sought after background, equivalent to a French grande école degree, is as far as I can tell a law degree from Yale. Engineers and planners always have to have a non-technical generalist watching over their shoulder, and this is the most prominent for the most politically sensitive projects.
  • FTA/FRA reviewers are in some cases not even allowed to probe into the funding package they are to decide on. One of the biggest projects relevant to what we’ve studied at the Transit Costs Project and what I’ve written on this blog has a multi-billion dollar package, one that will almost certainly be only partially funded due to competing priorities, but the regulators are not allowed to see any itemized breakdown to see what partial funding would even do. Nor are regulators allowed to say which priorities to build first if there’s only partial funding.
  • The higher regulators themselves believe that light-touch approaches are better and are reluctant to engage in any direct management – if they’d like to be more involved but are prohibited from doing so by the law or by constitutional interpretations, they have not said so.
  • There is little churn between operations and regulations – in fact, this separation is treated as sacrosanct, even as in all other aspects the governing ideology calls for breaking down silos (and thereby disempowering specialists in favor of generalists). The contrast here is with Sweden, where state planners who worked on Citybanan, a state project, have since moved on to work for Stockholm County on the county-led Nya Tunnelbanan.

None of this works. The people who make the big decisions on funding in this system do not have the ability to make professional judgments, only political ones, and agencies know this and don’t bother with technical soundness.

Heavy touch and expertise

The connection between heavy touch and expertise is, you can’t manage things directly if you don’t have a lot of subject matter knowledge. In this sense, light touch may not necessarily be by itself bad, but rather, like design-build project delivery, it is in practice used to mask incompetence at the top level. The non-technical boss, who is in all but name a political commissar, can make vague proclamations, not get into details, and not feel like they’re out of their league and must defer to the engineer in the room.

The Origins of Los Angeles’s Car Culture and Weak Center

On Twitter, Armand Domalewski asks why Los Angeles is so much more auto-oriented than his city, San Francisco. Matt Yglesias responds that it’s because Los Angeles does not have a strong city center and San Francisco does. I am fairly certain that Matt is channeling a post I wrote about the subject 4.5 years ago (and insight by transit advocates that I don’t remember the source of, to the effect that the modal split for Downtown Los Angeles workers is a healthy 50%), looking at employment in these two cities’ central business districts as well as other comparison cases. In addition, Matt gives extra examples of how Los Angeles is unique in having prestige industries located outside city center: the movie studios are famously in Hollywood and not Downtown, and to that I’ll add that when I looked at high-end hotel locations in 2012, Los Angeles’s were all over the region and most concentrated on the Westside, which isn’t true of other big American coastal cities, even atypically job-sprawling Philadelphia. Because of my connection to this question, I’d like to inject some nuance.

The upshot is that Los Angeles’s car culture is clearly connected to its weak center. I wouldn’t even call it polycentric. Rather, employment there sprawls to small places, rarely even rising to the level of a recognizable edge city like Century City. It is weakly-centered, and this favors cars over public transit – public transit lives off of high-capacity, high-frequency connections, favoring places with high population density (which Los Angeles has) and high job density (which it does not), while cars prefer the opposite because excessive density with cars leads to traffic jams. However, historically, best I can tell, the weak center and the cars co-evolved – I don’t think Los Angeles was atypically weakly-centered on the eve of mass motorization, and in fact every city for which I can find such information, even model transit cities, has gotten steadily job-sprawlier in the last few generations.

How is Los Angeles weakly centered?

There are a number of ways of measuring city center dominance. My metric is the share of metro area employment that is in the central 100 km^2; some gerrymandering and water-hopping is permitted, but the 100 km^2 blob should still be a recognizable central blob rather than many disconnected islands. This is not because this is the best metric, but because my information about France and Canada is less granular than for the United States, and 100 km^2 lets me compare American cities with Vancouver and with the combination of Paris and La Défense; my data on Tokyo is of comparable granularity to Paris and this lets me pick out Central Tokyo plus some adjacent wards like Shinjuku.

As a warning, the fixed size of the central blob means that the proportion should be degressive in city size, which I notice when I compare auto-centric American metro areas of different sizes. It should also be higher all things considered in the United States, where I draw blobs on OnTheMap to capture as many jobs as possible without the blob looking like it has tendrils, than in the foreign comparisons.

I gave many examples in a Twitter thread from 2019, though not Los Angeles. Doing the same exercise for Los Angeles with 2019 data gives 1.6 million jobs in a 500 km^2 blob stretching as far as Culver City, UCLA, Downtown Burbank, and Downtown Pasadena; a 100 km^2 blob gerrymandered to just include Hollywood, West Hollywood, and Century City, none of which can reasonably be called city center, is already down to 820,000, where the roughly same-area city of San Francisco is 770,000, and more like 900,000 when taking its central 50 km^2 plus those of Oakland and Berkeley. A circle of area 100 km^2 centered on Vermont/Wilshire to include all of Downtown plus Hollywood is down to 620,000. This compares with a total of 6.5 million jobs in Los Angeles and Orange Counties, and 8.3 million including Ventura County and the Inland Empire.

The upshot is that Downtown Los Angeles is pretty big, but not relative to the size of the metro area it’s in. On an honest definition of the central business district, it is smaller in absolute job count than Downtown San Francisco, Boston (which has around 830,000), Washington (around 700,000), or Chicago (1 million), let alone New York (around 3 million) or Paris (2 million in the city and the communes comprising La Défense).

Nor are the secondary centers in Los Angeles substantial enough to make it polycentric. Downtown Burbank has around 20,000 jobs, Downtown Glendale around 50,000, Downtown Pasadena including Caltech 67,000, Century City (included in the less honest central 100 km^2) 54,000, UCLA 74,000, El Segundo 55,000, LAX 48,000, Culver City around 20,000, Downtown Long Beach around 35,000. New York, in contrast, has Downtown Newark around 60,000, the Jersey City and Hoboken waterfront around 80,000, Long Island City around 100,000, Downtown Brooklyn around 100,000 as well, Flushing 45,000. Morningside Heights has 42,000 jobs in 1 km^2, a job density that I don’t think any of Los Angeles’s secondary centers hits, and the neighborhood is not at all a pure job center. No: Los Angeles just has a weak center.

I bring up Paris as a comparison because there’s a myth on both sides of the Atlantic, peddled by European critical urbanists who think tall buildings are immoral and by American tourists whose experience of Europe is entirely within walking distance of their city center hotels, that European city centers are less dominant than American ones. But Paris has, within the same area, comfortably more jobs than the centers of Los Angeles and Chicago combined; its central-100-km^2 job share is somewhat higher than New York’s (though probably only by enough to countermand the degressivity of this measure).

Was Los Angeles always like this?

I don’t think so. My knowledge of Los Angeles history is imperfect; the closest connection I have with it is that my partner is developing a narrative video game set in 1920s Hollywood, intended to be a realistic depiction of that era. But Los Angeles as I understand it was not especially polycentric, historically.

Historically polycentric regions exist, and tend to have weaker public transit than similar-size monocentric ones. The Ruhr has several centers, each with decent urban rail within the core city and high car usage elsewhere; Upper Silesia is far more auto-oriented than similar-size metropolitan Warsaw; Randstad has rather low urban rail ridership as people bike (in the main cities) or drive (in the suburbs). All three are truly historically polycentric, having developed as different city cores merged into one metro area as mechanized transportation raised people’s commute range, and in the case of the first two, much of this history involves different coal mining sites, each its own city.

Los Angeles doesn’t really have this history. The city had a slight majority of the county’s population in 1920 (577,000/936,000) and 1930 (1,238,000/2,208,000), only falling below half in the 1940s – and in the 1920s the city was already notable for its high use of cars. The other four counties in the metro area were more or less irrelevant then – in 1920 they totaled 244,000 people, rising to 389,000 by 1930, actually less than the city. Glendale grew from 14,000 to 63,000, Long Beach from 56,000 to 142,000, Santa Ana from 15,000 to 30,000; other suburbs that are now among the largest in the country either were insignificant (Anaheim had 11,000 people in 1930) or didn’t exist (Irvine had 10,000 people in 1970).

Los Angeles did annex San Fernando Valley early, but there wasn’t much urban development there in the 1920s; Burbank, entirely contained within that region, had 17,000 people in 1930, and San Fernando had 8,000. There was a lot of suburbanization in this period, but it did not predate car culture.

This is not at all how a polycentric region’s demographic history looks – in the Rhine-Ruhr, in 1900, Dortmund and both cities that would later merge to form Wuppertal had 150,000 people, Essen had just over 100,000 and would annex to over 200,000 within five years, Duisburg and Bochum both had just less than 100,000 and would soon cross that mark, Cologne had 370,000 people.

The region had an oil-based economy at the time – in the early 20th century the center of the American oil industry was still California and not Texas – but evidently, development centered on Los Angeles and to a small extent Long Beach (in 1930 having about the same ratio of population to Los Angeles’s that the combination of Jersey City and Newark did to New York’s). The same can be said of the various beach resorts that were booming in that era – the largest, Santa Monica, had 37,000 people in 1930, 3% of the population of Los Angeles, at which point Yonkers had 2% of New York’s population.

Boomtown infrastructure

While Los Angeles did not have a polycentric history in the 1920s, it did have a noted car culture. I believe that this is the result of boomtown dynamics, visible in many places that grow suddenly, like Detroit in the same era (in the 2010s, metro Detroit had a transit modal split of about 1%, the lowest among the largest American metro areas, even less than Dallas and Houston). Infrastructure takes time and coordination to build. In a growing region, infrastructure is always a little bit behind population growth, and in a boomtown, it is far behind – who knows if the boom will last? Texas is having this issue with flood control right now, and that’s with far less growth than that of Southern California in the first half of the 20th century.

The upshot is that in a very wealthy boomtown like 1920s Los Angeles (California ranking as the fourth richest state in 1929 and third richest in 1950), people have a lot of disposable income and not much public infrastructure. This leads to consumer spending – hence, cars. It takes long-term planning to convert such a city into a transit city, and this was not done in Los Angeles; plans to build a subway-surface tunnel for the Red Cars did not materialize, and the streetcars were not really competitive with cars on speed. Compounding the problems, the Red Cars were never profitable, in an era when public transit was expected to pay for itself; they were a loss leader for real estate development by owner Henry Huntington, and by the 1920s the land had already been sold at a profit.

Then came the war, and the same issue of private wealth without infrastructure loomed even more. California boomed during the war, thanks to war industries; there was new suburban development in areas with no streetcar service, with people carpooling to work or taking the bus as part of the national scheme to save fuel for the war effort. Transit maintenance was deferred throughout the country (as well as in Canada); after the war, Los Angeles had a massive population of people with very high disposable income, whose alternative to the car was either streetcars that were falling apart or buses that were even slower and had even worse ride quality.

Everywhere in the United States at the time, bustitution led to falling ridership per Ed Tennyson’s since-link-rotted TRB paper on the subject, even net of speed – Tennyson estimates based on postwar streetcar removal and later light rail construction that rail by itself gets 34-43% more ridership than bus service net of speed, and in both the bustitution and light rail eras the trains were also faster than the buses. But the older million-plus cities in the United States at the time had their subways to fall back on. Los Angeles had grown up too quickly and didn’t have one; neither did Detroit, which has a broadly Rust Belt economic and social history but a much more car-oriented transportation history.

The sort of long-term planning that produced transit revival did happen in the Western United States and Canada, elsewhere. In the 1970s, Western American and Canadian cities invented what is now standard light rail in both countries, often out of a deliberate desire not to be Los Angeles, at the time infamous for its smog; those cities have had more success with transit revival and transit-oriented development, especially Vancouver with its SkyTrain metro and aggressive high-rise residential and commercial transit-oriented development. But in the 1920s-40s, there was no such political counter to automobile dominance. Los Angeles did start building urban rail in the 1980s, but not at the necessary scale, and with ridiculously low levels of transit-oriented development: in the 2010s, after the economy recovered from the Great Recession, the 10 million strong county approved a hair more than 20,000 housing units annually, slightly less than the 2.5 million strong Metro Vancouver region.

Co-evolution of transportation and development

Los Angeles was not very decentralized in the first half of the 20th century. It had lower residential density, but none of today’s edge cities and smaller sub-centers really existed then, with only a handful of exceptions like Long Beach. By today’s standards, every American city was very centralized, with people generally working either in their home neighborhood or in city center. The city did have high car ownership for the era, and this encouraged freeway construction after the war, but the weak central business district came later.

Rather, what has happened since the war is a co-evolution of car-oriented transportation and weakly-centered job geography. Cars got stuck in traffic jams trying to get to city center, so business and local elites banded together to build an edge city closer to where management lived, first Miracle Mile and then Century City; Detroit similarly had New Center, where General Motors headquartered starting 1923. New York underwent the same process as businesses looked for excuses to move closer to the CEO’s home in the favored quarter (IBM in Armonk, General Electric in Fairfield), but the existence of the subway meant that there was still demand for ever more city center skyscrapers, even as city residents of means fled to the suburbs.

This story of co-evolution is not purely American. I keep going back to Paul Barter’s thesis, which portrays the urban layout in his example cities in East and Southeast Asia as starting from a similar point in the middle of the 20th century. Density was high throughout, and central sectors in Southeast Asia were ethnically segregated, with a Chinatown, an Indian area, a low-density Western colonial sector, and so on. The divergence happened in the second half of the 20th century, Singapore choosing to be a transit city and Kuala Lumpur and Bangkok choosing to be car-oriented cities. I don’t have job data for these cities, but my impression as a visitor (and former Singapore resident) is that Singapore has a clear central office district and Bangkok has a hodgepodge of skyscrapers with no real structure to where they go within the central areas.

So yes, Los Angeles’s weak center is making it difficult to expand public transportation there now and get high ridership out of it; boosting the region’s transit-oriented development rate to that of Vancouver would help, but Los Angeles is far more decentralized and auto-oriented than Vancouver was in the 1990s. But the historic sequence is not first polycentrism and then automobility, unlike in Upper Silesia or the Ruhr. Rather, a weak center (never true polycentricity) and automobility co-evolved, reinforcing each other to this day – it’s hard to get ridership out of urban rail expansions since city center is so weak, so people drive, so jobs locate where there’s less traffic and avoid Downtown Los Angeles.

Quick Note: Los Angeles Spends $50,000 Per Bus Shelter

I saw a few months ago that Los Angeles was planning to equip its entire bus network with shelter, and rejoiced that such an underrated amenity finally gets the attention it deserves. Unfortunately, the cost of the program is now $50,000 per bus shelter; to lower the cost, the city is experimenting with a substandard shelter providing no protection from the elements for $10,000. In Victoria, as pointed out by one commenter on Twitter, a full shelter is around $15,000 in 2018 Canadian dollars, comparable to 2023 American ones.

Ordinarily, it would be a dog-bites-man story: of course the costs are at a premium of a factor of three over Canada (let alone a place that builds cheaply), it’s an American transit infrastructure project. But this one is significant because bus shelter is small in scope – it’s not a megaproject, and this means that rules of megaprojects do not apply to it.

For example, installing shelter at thousands of stops, as Los Angeles is doing, allows for repetition of a standard design that vendors can figure out on their own. This means that the usual privatization of decisionmaking should not be a problem. (It’s been pointed out to me that design-build works well for municipal parking garages, since they’re so streamlined at this point.)

Moreover, bus shelter, even repeated so many times, is still small enough scope that a small in-house team could handle it. LACMTA has been talking to us about how to scale up their in-house team, and we couldn’t give them concrete numbers, but I believe what they have for design review is in the teens, which is probably not enough for a subway extension but should be for a bus shelter program measured in the tens of millions of dollars or (with the cost blowouts factored in) very low hundreds of millions. Elsewhere, for example in Boston, we’ve seen agencies build small things affordably – Boston’s premium over Berlin for building infill commuter rail stations is a factor of maybe 1.5 – and fail at building large things such as urban rail expansion, because their design review team is sized for small and not large things.

Finally, there’s a problem with politicization, in which when politicians insert themselves into a piece of infrastructure, hoping to make it their legacy, it will end up raising costs with at best no and at worst negative benefit to users. Small items like this ordinarily do not have this problem, as they fly under the radar of politicians as well as surplus-extracting community groups.

So why is this so expensive?

I don’t know the details of this failure – all I know is a handful of links to the current cost. I suspect, judging by the tone used in press coverage, that it’s politicization. Note that the piece linked in the lead paragraph centers the issue of gender equity; this isn’t stupid (as the other link points out, bus shelter has a disproportionately positive impact on women), but it does suggest someone thought about the political implications of this. The piece also quotes the designers as having developed the new substandard option “with input from female riders,” suggesting some waste on community consultation.

If I’m right on this, then this suggests a great peril for any city that looks at small and large projects, finds that the small ones are more cost-effective, and decides to prioritize them over large ones. In a city that builds subway expansion and also installs bus shelter beneath anyone’s notice, the bus shelter may be achievable at reasonable cost. But as soon as the shelter turns into a splashy program, beloved by incompetent managers who figure it’s within their span of control and by politicians (who are by definition incompetent on managerial issues), it will acquire the same problems of politicization usually associated with megaprojects. In this case, it’s waste coming from community consultation. In other cases, it might be demands for neighborhood signatures, or intransigence by abutting property owners or by utilities figuring there is surplus to extract, or any of the other causes of cost overruns.

And in particular, downgrading your city’s investment plan just because the flashiest items have cost blowouts wouldn’t end the cost blowouts. If you can’t build subways, work on making yourself capable of building subways; avoiding that mess and building other things instead of subways would run into the same problems and soon you’d be paying $50,000 for a single bus shelter.