Category: High-Speed Rail

The Different National Traditions of Building High-Speed Rail

I’ve written five pieces about national and transnational traditions of building urban rail: US, Soviet bloc, UK, France, Germany. I’m about to continue this series with a post about Japan, but yesterday I made a video on Twitch jumping ahead to different national traditions of high-speed rail. The video recording cut two thirds of the way through due to error on my part, so in lieu of an upload, I’m writing it up as a blog post. The traditions to cover are those of Japan, France, Germany, and China; those are the world’s four busiest networks, and the other high-speed rail networks display influences from the first three of those.

The briefest description is that the Shinkansen is treated like a long-range subway, the TGV like an airplane at flight level zero, and the ICE like a regional rail (and not S-Bahn) network. China doesn’t quite fit any of these modes but has aspects of all three, some good, some not.

But this description must be considerably nuanced. For example, one would expect that airplane-like trains would have security theater and a requirement for early arrival. But the TGV has neither; until recently, platforms were completely open, and only recently has SNCF begun gating them, not for security but for ticket checks, with automatic gates and QR codes. Likewise, until recently passengers could get to the train station 2-3 minutes before the train’s departure and get on, and only now is SNCF requiring passengers to show up as long as 5 minutes early.

Tabular summary

TraditionJapanFranceGermanyChina
SummarySubwayAirplaneRegional railMixed
InfluencedKorea, TaiwanSpain, Italy, Belgium, MoroccoNorthern Europe
FrequencyVery highLowMediumHigh
Seat turnoverMediumLowHighMedium
PricingFixedDynamicMixedFixed
Approximate fare/km$0.23$0.14$0.15$0.10
EgressVery fastVery slowMediumFast
Integration with slow trainsMediumPoorGoodPoor
Average speed (major cities)HighHigh, except BelgiumMixed high, lowVery high
Timed connectionsNoNoYesNo
One-seat ridesLimitedExtensiveCommonCommon
Security theaterNoOnly in SpainNoYes
Platform access controlYesIncreasingly yesNoYes
Major city stationsCentralHistoric, Paris has 4CentralOutlying
Terminal turnaroundsFastSlowMixedSlow
Minor city stationsMixedOutlying, “beet fields”Usually legacyUsually outlying
FreightNoNoYesNo
Grades1.5-2%3.5%1.25%, max 4%1.5-2%
TunnelsExtensiveRareExtensiveRare
ViaductsExtensiveRareRareExtensive
Construction costsHighLow or mediumMediumHigh

For more detailed data on costs and tunnel and viaduct percentage, consult our high-speed rail cost database.

The Shinkansen as a subway

The Shinkansen network has very little branching. Currently there is none south of Tokyo; a short branch to Nagasaki is in planning but will not open anytime soon. To the north, there is more branching, and the Yamagata and Akita Mini-Shinkansen lines, the only legacy lines with Shinkansen through-service, split trains, with one part of the train continuing onward to Shin-Aomori and Hokkaido and another part splitting off to Yamagata or Akita.

Source: Wikipedia

Going south of Tokyo, the off-peak frequency to Shin-Osaka is four express Nozomi trains an hour, at :00, :09, :30, :51 off-peak; two semi-express Hikari, at :03, :33; and one local Kodama, at :57. The 21-minute gaps are ugly, but on a train that takes around 2.5 hours to get to Shin-Osaka, they’re not too onerous. Thus, there is a culture of going to the train station without pre-booking a ticket and just getting on the next Nozomi. The ticketing system reinforces this: there is no dynamic yield management, but instead fixed ticket prices between pairs of station depending on seat class. What yield management there is is static: the Nozomi has a small surcharge, to justify excluding it from the JR Rail Pass and so shunt tourists to the Hikari.

This is not literally the headway-management system seen on some unbranched subway systems, like the Moscow Metro and Paris Métro; Moscow keeps time by distance from the preceding train, and not by a fixed schedule. But this is fine: some subway systems are timetabled, like the U-Bahn in Berlin and the Tokyo subway. Tokyo even manages to mix local and express trains on some two-track subway lines with timed overtakes. To the scheduler, the fixed timetable is of paramount importance. But to the passenger, it isn’t – people don’t time themselves to a specific train.

Another subway-like characteristic includes interior layout, designed around fast egress. Shinkansen cars have two door pairs each and platforms are 1,250 mm high with level boarding, enabling 1 minute dwell times even at very busy stations like Shin-Osaka. Trains make multiple stops in the Tokyo and Osaka regions, and even Nozomi and equivalent fastest-train classes on other lines stop there, to distribute loads. There is no cafe car, and luggage is overhead, to maximize train seating space: a 25 meter car has 18-20 seating rows with 1-meter pitch, which is greater efficiency than is typical in Europe.

Station location decisions, finally, are designed as far as practical to be in city centers. Stations with Shin- before their names are new stations, like Shin-Osaka and Shin-Yokohama, but they tend to be sited close to city centers, at intersections with subway and commuter rail lines.

The main drawback of Japan is that the construction costs are very high. This comes from a political decision to build elevated lines rather than at-grade liens with earthworks, as is common in Europe. This preponderance of els has been exported to South Korea, Taiwan, and China, all of which have high costs relative to the tunneling proportion; the KTX, essentially a Shinkansen adapted to an environment in which the legacy trains are standard-gauge too, is notable for having low tunneling costs, as is common in Korea, but high costs on lines with moderate amounts of tunneling thanks to the high share of construction on bridges.

East Asia has high population density, which lets it get away with high costs since the ridership is high enough to compensate – THSR is at this point returning around 4% on very high costs. But in any other environment, this leads to severe problems. China, with lower incomes and fares than in Japan, Korea, and Taiwan, already has trouble paying interest on lines other than the Beijing-Shanghai system. India, building a turnkey Shinkansen as recommended by Japanese consultants, who were burned by Taiwan’s mix of European and Japanese technology on an operationally-Japanese system, is spending enormous sums of money: the Mumbai-Ahmedabad corridor is around PPP$50.6 billion, for 508 km, $100 million/km on a line that’s only 5% in tunnel and even those tunnels could have been avoided by running on broad gauge and using existing a widened legacy right-of-way in Mumbai.

The TGV as flight-level zero air travel

As detailed in New Departures by Anthony Perl, the history of the TGV differs from that of the Shinkansen in a key aspect: the TGV was built after the postwar decline of rail travel (as was the ICE), whereas the Shinkansen was built before it (as was to some extent CRH). The Shinkansen was built in 1959-64: there was no decline in rail evident yet, with only 12 cars/1,000 people in Tokyo in 1960, and the system was designed to deal with growing ridership. In contrast, the TGV was planned after the 1973 oil crisis, in a then-wealthier and more motorized country than Japan, aiming to woo passengers back to the train from the car and the plane.

Previously, SNCF had been engaging in experiments with high speed and high-voltage electrification, inventing 25 kV 50 Hz electrification in the process, which would be adopted by the Shinkansen and become the global standard for new electrification. It also experimented with running quickly on ballasted track – without modifications, the trains of that era kicked ballast up at high speed, there was so much air resistance. But investment had gone to legacy intercity rail, driving up the average speed of the electrified Mistral to 130 km/h and the Aquitaine to 145 km/h. Nonetheless, competition with air was fierce and air shuttles in that era before security theater attracted many people in competition with four-hour trains from Paris to Lyon and Bordeaux.

The TGV’s real origin is then 1973. The crisis shocked the entire non-oil-exporting world, leading to permanently reduced growth not just in rich countries (by then including Japan) but also non-oil-exporting developing countries, setting up the sequence of slow growth under import substitution and then the transition to neoliberalism. France reacted to the crisis with the slogan “in France, we have ideas,” setting up the nuclearization of French electricity in the 1980s, reduced taxes on diesel to encourage what was then viewed as surplus fuel rather than as a deadly pollutant, and the construction of the electric TGV.

Despite the ongoing growth of the Shinkansen then, there was extensive skepticism of the TGV in the 1970s and early 80s. The state refused to finance it, requiring SNCF to borrow on international markets. The LGV Sud-Est employed cost-cutting techniques including 3.5% grades and high superelevation to avoid tunnels, at-grade construction with cut and fill balancing out to avoid surplus dirt, and land swaps for farms that would be split by the line to avoid needing to build passageways.

Construction costs were only 5.5M€/km in 2021 euros. Unfortunately, costs have risen since and stand at 20M€/km, or even higher on Bordeaux-Toulouse. But the LGV network remains among the least tunneled in the world thanks to the use of high grades; in our database the only less tunneled network, that of Morocco, is a turnkey TGV, built at unusually low cost.

As in Japan, the line was built between the two largest cities: Paris and Lyon. Also as in Japan, Lyon could not be served at the historic center of Perrache, but instead at a near-center location, Part-Dieu, which then became the new central business district, as the LGV Sud-Est was built concurrently with the Lyon Metro and nearby skyscrapers, as is typical for a European city wishing to avoid skyscrapers in historic centers. But everything else was different. There were no real intermediate stops the way that the express Shinkansen have always stopped at Nagoya and Kyoto: the LGV Sud-Est skipped Dijon, which instead was served on a branch, and the two intermediate stops on the line, Le Creusot and Mâcon-Loché, are on the outskirts of minor towns and only see a few trains per day each.

Moreover, relying on France’s use of standard-gauge, there was, from the start, extensive through-service beyond Lyon, toward Marseille, Geneva, Saint-Etienne, and Grenoble. Frequency was for the most part low, measured in trains per day. There was little investment in regional rail outside the capital, unlike in Germany, and therefore there was never any attempt to time the connections from Saint-Etienne and Grenoble to the TGV at Part-Dieu.

At the other end, Paris did not build a central station, unlike German or Japanese cities. The time for such a station was, frustratingly, just a few years before work began on the TGV in earnest: RATP was building the RER starting in the 1960s and early 70s, including a central station at Les Halles, which opened 1977. But this was designed purely for urban and suburban use, and the TGV stayed on the surface. The last opportunity for a Paris central station was gone when SNCF extended the RER D from Gare de Lyon to Les Halles. Thus Paris has four distinct TGV stations – Lyon, Montparnasse, Nord, and Est – with poor connections between them.

This turned the TGV into a point-to-point system. Were there a central station, trains could have gone Lille-Paris-Lyon-Marseille. But there wasn’t, and so for Lille-Lyon service, SNCF built the Interconnexion Est, bypassing Paris and also serving Disneyland and Charles-de-Gaulle Airport. When the LGV Atlantique opened, Tours kept its historic terminal, and thus trains went either Paris-Tours or Paris-Bordeaux bypassing Tours. When the LGV Sud-Est was extended south with the LGVs Rhône-Alpes and Méditerranée, trains did not go via Part-Dieu, even though it had always been configured as a through-station for points south, but rather via a bypass serving Lyon’s airport; trains today go Paris-Lyon, Paris-Marseille, or at lower frequency Lyon-Marseille, but not Paris-Lyon-Marseille.

Of note, Japan’s subway-like characteristic is partly the outcome of its linear geography along the Taiheiyo Belt, making it an ideal comparison also for the Northeast Corridor in the United States. But Lille, Paris, Lyon, and Marseille are collinear, and yet the service plans do not make use of that geography. There is no planning around seat turnover: if a train makes an intermediate stop, it’s one with very low ridership, like Mâcon, with no attempt to have seats occupied by Paris-Lyon passengers and then by Lyon-Marseille ones.

Over time, this led to a creeping airline-ization of the TGV. Airline-style dynamic yield management was introduced, I believe in the 1990s. This was after SNCF had spent the 1980s marketing the TGV as 260 km/h for the same fare as 160 km/h; the overall fares on legacy intercity trains and TGVs are similar per p-km, but TGVs have opaque pricing, and are designed to maximize fares out of Paris-Lyon in particular, where air competition vanished. The executives at SNCF are increasingly drawn from the airline world, and, perhaps out of social memory of the navettes competing with 4-hour trains in the 1970s, they think that trains cannot compete with air travel if they take longer than 3-3.5 hours, even though they do successfully on such city pairs as Paris-Toulon.

Having skipped Germany’s InterCity revolution and its refinements in Switzerland, Austria, and the Netherlands, the TGV network has stagnated in the last decade. Ridership is up since the pre-Great Recession peak but barely, only by around 10%. The frequency is too weak for inter-provincial links, where people mostly drive, and in the 1990s and 2000s the TGV network grew to dominate the Paris-province market; there isn’t much of a remaining market for the current operating paradigm to grow into.

While some regional links are adopting takt timetables, for example some of the Provence TERs, SNCF management has done no such thing. Instead, it has spent the last 15 years pursuing airline strategies, including imitation of low-cost airlines, first iDTGV and then OuiGo. A generalist elites of business analysts believes in market segmentation and price discrimination, which do not work on a mode of travel where a frequent, flexible timetable is so paramount.

Among the countries influenced by France, Spain is notable for realizing that it has a problem with operations. In an interview with Roger Senserrich, ADIF head Isabel Pardo de Vera spoke positively of Spain’s efficient engineering and construction, but centered ADIF and RENFE’s problems, including the poor operations. Like Italy and Belgium, and more recently Morocco, Spain learned the concept of high-speed rail from France; also like Italy and Belgium, it mixed in a few German elements, which in the 1980s meant Germany’s more advanced LZB signaling, but at the time, there was no Switzerland-wide takt yet, and the inferiority of French operations and scheduling was not yet evident. But Spain self-flagellates – this is how it learns – whereas France is just a hair too rich to recognize its weaknesses and far too proud for its elite to Germanize where needed.

The ICE as long-distance regional rail

Germany came into the 1960s with some of the most advanced legacy rail in the world, with technology that would be adopted as a Shinkansen standard. This goes back to the 1920s, when Deutsche Reichsbahn was formed from the merger of the state-level railways in the wake of the post-WW1 German Revolution. The new railway regulation, dating to 1925, promoted new kinds of engineering now completely standard, such as the tangential switch. DRB would also experiment with 200 km/h diesel express trains in the 1930s. Even in the 1960s and early 70s, when the most advanced rail tech was clearly in Japan, Deutsche Bundesbahn kept up with rail tech, much like SNCF, inventing LZB signals.

But unlike Japan and France, Germany never built a complete high-speed rail network. The InterCity network, dating to 1971, was designed around fast legacy trains, at slightly lower speeds than available on the express French legacy trains. The key was that city pairs would be served every two hours, with timed connections at intermediate points boosting many to hourly. This was from the start based on a regular takt and turnover, with more expansive service to smaller cities.

High-speed lines in Germany were delayed, and often built on weird alignments. The most important reason is that in the formative period, from 1971 to 1990, there was no such country as Germany. The country was called West Germany, and, much like Japan, had a fairly linear population distribution from the Ruhr upriver to Cologne, Frankfurt, Mannheim, and finally either Karlsruhe or Stuttgart and Munich; but the largest city proper, Hamburg, lay outside this corridor.

The north-south orientation of West Germany contrasted with the rail network it inherited. Until the post-WW1 German Revolution, the rail networks were run by the states, not by the German Empire, and thus interstate connections were underbuilt. Prussia had an east-west orientation, and therefore north-south lines were relatively underbuilt (see for example the 1896 map), and to top it off most north-south routes crossed the Iron Curtain.

To solve many problems at once, but not to solve any of them well, Germany’s first high-speed line connected Hanover, Göttingen, Kassel, Fulda, and Würzburg. Getting to more substantial cities like Hamburg and Frankfurt requires onward through-service at lower speed. The LGV Sud-Est had a minimum curve radius of 3.2 km, and usually 4 km, and can squeeze 300 km/h out of it now, without any tunnels; the Hanover-Würzburg line has a minimum radius of 5.1 km and a maximum grade of 1.25% and is limited to 280 km/h (service runs at 250 km/h), as it was built as a mixed freight-passenger line.

Subsequent lines have, like Hanover-Würzburg, not been complete connections between major cities. Here the difference with France, Italy, South Korea, and China is evident. All are standard-gauge countries, like Germany, and all employ through-service to various degrees. But France opened a complete Paris-Lyon high-speed line in 1981-3, and only the last 30 km into Paris were on legacy trains (since reduced to 8 km with the Interconnexion Est), and likewise Italian, Chinese, and Korean high-speed lines connect major cities all the way. In contrast, this never happens in Germany at longer distance than Cologne-Frankfurt, a 180 km connection. There are always low- or medium-speed segments in between. The maximum average speed between major cities in Germany is either Cologne-Frankfurt or Berlin-Hamburg, a 230 km/h line with tilting trains, both averaging around 180 km/h; the Tokaido Shinkansen, with legacy 2.5 km curves, squeezes 210 km/h out of the Nozomi, and LGVs routinely average 230-250 km/h between Paris and major secondary cities.

Nor are the lower speeds in Germany saving money. The mixed passenger/freight lines have heavier tunneling than they would need if they had 3.5-4% grades. Hanover-Würzburg cost 36M€/km in 2021 euros thanks to its 37% tunneled alignment. German construction costs are not high relative to the tunneling percentage, unlike Chinese or Taiwanese costs, let alone British ones, but the tunneling percentage is in many cases unnecessarily high. This is thankfully not exported to every Northern European country that learned from the InterCity, but the Netherlands, as NIMBY-ridden as Germany, built an unnecessary tunnel on the HSL Zuid and had very high costs even taking that into account; Italy, with an otherwise-French system, likewise overbuilds, as pointed out by Beria-Albalate-Grimaldi-Bel, with viaducts designed to carry heavy freight trains even where there is no such demand.

So the bad in Germany is that the lines have very shallow grades, forcing heavy tunneling, and the costs are so high that the system is not complete. Is there good? Yes!

The InterCity system’s focus on high frequency enables decent service between major cities. Berlin-Munich trains, compromised by the Erfurt detour and subsequent descoping of much of the line, do the trip in 4.5 hours where they should be taking 3 and even 2.5 hours. But it’s not the same as the 4 hours of the pre-TGV Mistral to Lyon or Aquitaine to Bordeaux, the latter of which averaged the same speed as most Berlin-Munich trains today. The Aquitaine ran as a single daily Bordeaux-Paris-Bordeaux round-trip, and another train, branded the Etendard, ran the same route daily but Paris-Bordeaux-Paris. In contrast, DB today connects Berlin-Munich roughly every hour. It’s far more flexible, and the connections to other intercity trains are better.

And just as the TGV’s inexpensive construction has been perfected in Spain while France has slouched on cost control, so has the interconnected system of Germany been perfected on the margins of its sphere of influence, especially in Switzerland. Swiss connections are never fast: the country is too small for 300 km/h trains to make large differences in door-to-door trip times. The average speed on the workhorse Swiss lines connecting the Zurich-Bern-Basel triangle is around 110-120 km/h. But they run on a half-hourly takt, and other lines run on an hourly takt, and connections at the major cities are timed. European urbanism has a long tail of small cities, unlike American or Asian urbanism, and the Swiss takt connections those small cities to one another through regular timed transfers, with investments to prioritize punctuality.

This leads to a false belief among German rail advocates in a tradeoff between French or Spanish speed and Swiss or Dutch or Austrian connectivity. The latter set of countries have higher rail ridership per capita, and even Germany has recently overtaken France’s intercity rail ridership (though not yet per capita), and thus activists in Germany think investing in high speed is a waste. But what is actually happening is that the countries of Europe that look up to France have built high-speed rail, and the countries that look down on France have not; the Netherlands has HSL Zuid but it’s peripheral to the national network and its system is otherwise rather Swiss. Germany absolutely can and should complete its network. It just needs to understand that in certain aspects, countries it is used to stereotyping as spendthrift have done a more prudent job than it has.

Already, the younger rail advocates I meet, like Felix Thoma, seem interesting in applying the Deutschlandtakt concept to a high-speed rail network, rather than to a medium-speed one as the previous generations called for. But Germany is a NIMBY country. NIMBYs blocked French levels of energy nuclearization in the 1970s and 80s, creating the last generation’s Green Party (current leader, Annalena Baerbock, is 40 and came of age after those fights); NIMBYs sue projects they dislike on frivolous grounds until the politicians lose interest, much as in the US with its government-by-lawsuit, and thus high-speed rail on the Hamburg-Hanover line has been stuck in limbo for a generation.

Besides the political deference to NIMBYs, who as in the US are not as powerful as either they or the state thinks, the main problem then is unwillingness to merge French and German planning insights where they work. I might also add Japanese insights – the Shinkansen is far more efficient with platforms than any European railroad – but they’re less important here or in France than in the UK, which is a ridiculously high-cost version of French planning.

China as a mixture of all modes, some good, some awful

When I started planning this video and now post, I was puzzling over where to slot China. Other systems seemed fairly easy to slot as Japanese, German, or French, with the occasional special feature (insanely high UK costs, HSL Zuid in an otherwise Swiss intercity takt system, Korean standard-gauge adaptations). But China is its own thing. It makes sense: on the eve of corona, China had 2.3 billion annual high-speed rail riders, comfortably more than than the rest of the world put together; Japan, the second busiest network, had 436 million. In Europe, only France has more high-speed rail ridership per capita, by the smallest of margins.

Historically, the system should be viewed as having borrowed liberally from other systems in richer countries that built out their networks earlier. Among the three prior traditions, the one most similar to what CRH has converged on is the Shinkansen, and yet there is significant enough divergence I would not class CRH as a direct Shinkansen influence the way I do the KTX and THSR. This also mirrors the situation for rapid transit: China displays clear Soviet influences but has diverged sufficiently that it must be viewed as a separate tradition now.

The most important feature is that CRH evolved on the cusp of the decline of rail in favor of cars and planes, a decline that has been more complete in Western countries. In the 1980s and early 90s, China was already growing very quickly; this was from a very low base, so it was not noticed in richer countries, but it was enough that there were already motorization and domestic air travel competing with China Railway. This led to a multi-phase speed-up campaign, announced in 1993 and implemented from 1997 to 2007.

At this point, construction was on legacy alignments to legacy stations. In the North China Plain, the railroads were straight thanks to the flat topography, and so what was needed was investment in the quality of the physical plant – the sort of investments figured out in midcentury France and Germany, adapted by the Shinkansen. This was not trivial, not in a then-low-income country like China, but it was not enormously expensive either. At the same time, there was growing electrification in China, using 25 kV 50 Hz, leading to higher and higher train classes, all charging premium fares over the third-world tickets for traditional trains. At the apex was the D class, covering 200 km/h EMUs; the one time I rode a train in China, a day trip from Shanghai to Jiaxing and back in 2009, the way back was on a D class train, which had the comfort level and speed of the Northeast Corridor, topping at 170 km/h and averaging maybe 110. This investment has continued, and as of 2019, 72% of the network is electrified.

But China was already looking for more. In 2008, the Beijing-Tianjin high-speed line opened, as the world’s first 350 km/h line. In the financial crisis’s aftermath, China rapidly built out the network as fiscal stimulus, and by 2011, ridership overtook the Shinkansen’s as the world’s largest. Without legacy considerations, the system is built for 380 km/h, even though trains run at 350 km/h, and express trains average 280-290 km/h.

Like the United States and unlike Japan or most of Western Europe, China has an extensive freight rail network. Its approach is the opposite of Germany’s: high-speed lines are dedicated to passengers, and some are officially called passenger-dedicated lines, or PDLs, to make this clear. Freight trains go on the legacy network. Regional rail in China is very weak; the few lines that exist are new-builds, rather like long-range subways, and frequency is often lacking, the Beijing lines branded as S-Bahn barely running off-peak. With nearly all intercity rail having moved over to CRH, the legacy network is relatively free for freight use, even coal trains, which are slow and care little for reliability improvements for higher-end intermodal cargo.

However, the passenger-only characteristic of CRH’s system does not mean it’s employed French cost-cutting techniques. Rather, lines run almost exclusively on viaducts and have shallow grades, raising construction costs as in the rest of East Asia. Stations are newly-built at high expense: Beijing South cost 7 billion yuan, which in today’s PPP dollars is around $3 billion. There are many tracks and no economization with fast turnarounds as in Japan, and station layouts are comparable to airports, with some security theater.

Beijing South is at least just outside the Second Ring Road. Other stations are farther out. This is not just the beet field stations that characterize TGV service to small cities like Amiens or Metz, but also outlying stations in major centers. Shanghai Station only sees high-speed trains on the local line to Nanjing, providing a dedicated track pair equivalent to Kodama service while Nozomi-equivalent trains continue on to Beijing on their own tracks. The trains to Beijing get a separate Shanghai station, Hongqiao, colocated with the city’s domestic airport. The connecting subways tend to be better than at true beet field stations in France, which miss regional rail connections, but those stations are still well outside city center.

China is moreover exporting the bad more than the good. Chinese-funded projects in Africa are not fast – the average speeds are perhaps midway through China’s speed-up campaign, predating CRH. But they do have oversize, airport-like stations located well outside city centers. This happens even when right-of-way to enter city center exists, as in Nairobi.

On mixing and matching

Understanding these four distinct traditions is important for high-speed rail planning, in those four countries as well as elsewhere, such as in the UK and US. It’s important to understand the tradeoffs that these traditions made, and drawbacks that are not so much tradeoffs as things that didn’t seem important at the time.

Most notably, Britain has oversize stations, spending billions on new terminals such as in Birmingham. This comes from the low efficiency of most European turnaround operations, because most European cities have huge rail terminals from the steam era with a surplus of tracks. When trains need to turn fast, they do: German trains running through Frankfurt, which is a terminal, turn in 3-4 minutes to continue to their onward destination. In Tokyo, where space is at a premium, JR East learned to turn trains in 12 minutes even while giving them a cleaning, and with such tight operations, Britain should be able to fit traffic growth within existing station footprints.

It is also desirable to learn from students who have surpassed their old teachers. Korea has lower construction costs than Japan, Spain has lower construction costs than France and greater understanding of the need to integrate the timetable and infrastructure, Switzerland has perfected the German system to the point that German rail advocacy calls for reimportation of its planning maxims.

In the same way that Taiwan built infrastructure to European specs but is running Japanese trains on it, to its profit and to Japan’s chagrin, it may be advisable to build infrastructure in the French (or, better yet, Spanish) way but then run trains on it the German (or better yet, Swiss) way. But it’s more nuanced than this conclusion, due to important contributions from China and Japan, and due to the focus on having a central station, which France chose not to build in Paris to its detriment.

But in general, I think it behooves countries to learn to implement the following from those four traditions:

  • Japan: the best rolling stock, high-efficiency turnaround operations, reliable schedules; avoid excessive viaducts and Japan’s increasing demand for turnkey systems.
  • France: passenger-dedicated infrastructure standards (supplemented by Cologne-Frankfurt), land swap deals for at-grade construction, cost control (in the Spanish version – France is deteriorating); avoid TGV rolling stock and airline-style pricing.
  • Germany: takt (especially in the Swiss and Dutch versions), open station platforms, integration between timetable and infrastructure, seat turnover, decent rolling stock; avoid empowering NIMBYs and building mixed lines with freight.
  • China: separation of passenger and freight operations, very high average speeds; avoid airline-style outlying stations and excessive viaducts.

High-Speed Rail Costs and Presentation

We have a database of high-speed rail construction costs up.

Separately, because of Noah Smith’s opinions about high-speed rail, today there is going to be an event featuring me and him in which we are going to discuss the issue in an American context, alongside a presentation of the database and what lessons can be drawn from it. You can register here; it’s at 13:00 Eastern US Time, or 19:00 Berlin time.

A few notes regarding our database, because I’m being asked on Twitter, and also because it’s relevant for our research:

This is a well-studied topic

Literature on comparative HSR costs already exists, and some of our internal cost references are to studies on the subject. This is not like subway costs, where the biggest databases I know of prior to ours are a Flyvbjerg paper and a Spanish analysis each with a number of items in the teens. This should not in a way be surprising: the costs and impact of megaprojects are analyzed more than those of smaller projects, and subways are megaprojects of greater size than surface transit or street reconstruction but HSR is of yet greater size. Thus, subways are significant enough that we have been able to find largely complete costs from trade and mass media and government reports, which task is far harder for bus lanes or bike lanes, whereas with HSR, not only is it possible to find complete costs, but also there is extensive public debate and analysis.

I believe our contribution to the discussion, then, is not the database itself, but two new points:

  • Contrary to the World Bank report on the subject (see here, starting printed page 39), China does not build HSR especially cheaply. Our findings are not too different from the World Bank’s for lines built up to the publication of the report measured in yuan per km, but we adjust for PPP and therefore the cost in dollars per km is higher, and, moreover, the more recent lines appear to be more expensive. In fact, Chinese costs are higher than European ones. The reason is that China builds its HSR almost entirely on viaduct, whereas in Europe, viaducts are rare, and segments that are not in tunnel are built at-grade or on earthworks.
  • There is positive correlation between a country’s HSR costs per km, net of tunnels, and its subway construction costs. This is not perfect correlation, but one can see Britain, the Netherlands, and Taiwan perform poorly in both areas. France and Germany are in the middle. Spain is very cheap. The exceptions are notable: Italy has cheap subways and expensive HSR, which Paolo Beria, author of one of our source papers, attributes to overbuilding and overdesign, with extensive tunnels and freight-friendly grades.

We only include under-construction or open lines

This contrasts with lines that are only in early design and may not yet have a cost – for example, Frankfurt-Mannheim will only publish its cost estimate next year, in a parliamentary budget setting in order to decide whether to proceed (for which the answer is certainly yes, as the benefits to the network are intensive). This also contrasts with canceled and indefinitely postponed lines, such as California High-Speed Rail and the Portuguese lines killed during the Great Recession’s austerity. Canceled lines are upward-biased: the state is likelier to cancel or choose not to build a line if it is more expensive than the average, as we can readily see with California, and therefore we do not wish to compare built with unbuilt lines.

The above analysis is equally true of our subway construction costs database – if a line is canceled, it is purged, even if design or even physical construction began. Gateway for example is under active design and engineering and is therefore included, even if they are still seeking funding, but if it is canceled it will be purged (but if it is rebooted, as I hope, then the sunk cost will be included, as with the Green Line Extension in Boston).

The difference is that our HSR cost database is more historic. It is close to complete for France, Germany, Italy, Spain, Belgium, and Korea, and complete for single-line Taiwan and the Netherlands and for the UK. This is because it’s just easier to find historic data for HSR than for subways, where I wish I could get a complete historic series for big cities with big systems like Paris, Madrid, and Berlin, but can’t even find 1970s-80s costs for any of them. Conversely, ongoing projects make it surprisingly difficult at times to find tunnel and viaduct percentages, and the escape path of going on Google Earth and OpenStreetMaps and measuring is not available.

What is included?

As far as possible, costs are for civil infrastructure, systems, stations, and overheads, but not rolling stock or financing charges. Austria’s Koralmbahn has two sets of numbers, differing by a factor of 2, with one source claiming that it is about whether financing is included. It is my belief that, owing to the high profitability of HSR if cost of capital is ignored, it is best to think in terms of returns on investment and not try to incorporate debt or finance charges into the actual cost.

The importance of avoiding viaducts and tunnels

The Asian tendency to build on viaduct where the line is not in tunnel leads to high costs. Likewise, the use of shallow grades and low superelevation for mixed lines or even for some dedicated lines (the Shinkansen, without any track sharing, hews to 1.5% grads) raises construction costs.

Netting out tunnels is still useful when trying to figure out itemized costs and cost control that is not about what to build, for example about labor or procurement. It is also useful when comparing lines in the mountainous terrain of Austria, Japan, Korea, and Switzerland to the easier North European Plain. But at some point, it is necessary to treat the tunnel percentage as endogenous to the planning system. The viaduct percentage, moreover, is absolutely endogenous.

France in this context does well by keeping lines at grade as much as possible. The only country with less tunneling than France is Morocco, which builds its urban and high-speed trains as if it were France, and, thanks to France’s extensive presence in the Maghreb, French contractors are intimately familiar with the local situation and build cheaply. France and Germany have similar unit costs, but Germany tunnels a lot more, less because of the terrain and more because of either politics (that is, the Erfurt detour for Berlin-Munich, forcing the line to go through thicker mountains) or a misguided attempt at building mixed lines in the 1980s and 90s.

The United States’ high projected budgets for proposed lines that never go anywhere thanks to their extreme costs come from overbuilding more than high unit prices. For example, in Baltimore, a two-track tunnel project designed for exclusive electric passenger train usage turned into a four-track tunnel with enough room for double-stacked freight with mechanical ventilation for diesel locomotives. The scope creep raised the projected budget from $750 million in the late 2000s to $4 billion in the mid-2010s.

Randal O’Toole Gets High-Speed Rail Wrong

Now that there’s decent chance of US investment in rail, Randal O’Toole is resurrecting his takes from the early Obama era, warning that high-speed rail is a multi-trillion dollar money sink. It’s not a good analysis, and in particular it gets the reality of European and Asian high-speed rail systems wrong. It displays lack of familiarity with rail practice and rail politics, to the point that most nontrivial assertions about rail in Europe and Asia are incorrect.

More broadly, the way O’Toole gets rail investment here wrong comes from making unexamined American assumptions and substituting them for a European or Japanese reality regarding rail as well as rail politics. If the US can’t do it, he thinks other countries can’t. Unfortunately, he’s even unfamiliar with recent work done on American costs, when he compares the Interstate system positively with recent high-speed rail lines.

High-Speed Rail Profitability: France

I’m currently working on building a database similar to our urban rail costs for high-speed rail. Between this and previous iterations of analyzing the TGV, I’ve been reading a lot of internal French reports about its system. Thankfully, France makes available very good public information about the costs and technical specifications of its system. It helps that I read French, but the gap between what’s available for France and Belgium (see for example line schemas) is vast. This provides crucial background that O’Toole is missing.

The most important thing to understand is that the TGV network is profitable. The Spinetta report on the fiscal losses of SNCF makes it clear, starting on p. 60, that the TGV network is profitable, and recommends favoring its development over the money-losing legacy networks, especially the branch lines. The report even calls for closing weak branch lines with only a few trains a day, which I called the Spinetta Axe at the time, in analogy with the Beeching Axe. Due to public outcry the state rejected the cuts and only implemented the organizational changes promoted by the report.

Moreover, all lines are very profitable excluding the cost of fixed capital. The Spinetta report’s TGV section says that operating costs average €0.06/seat-km, which is around 0.085€/p-km, despite overstaffing of conductors (8 per conventional 400-car TGV) and extensive travel on legacy track at low speed and higher per-km labor costs. Average TGV fare revenue per an ARAFER report from 2016 is 0.10€/p-km – compare p-km on p. 15 and revenue on p. 26. This is typical for Europe – RENFE and DB charge similar fares, and the nominal fares seem to have been flat over the last decade.

What’s dicier is cost of capital. In all other European countries for which I’m aware of the process, all of which are Northern rather than Southern, this is done with benefit-cost analysis with a fixed behind-the-scenes discount rate. France, in my view wisely, rates lines by their financial and social rates of return instead. A 2014 report about the Bordeaux-Toulouse LGV, recently given the go-ahead for 7.5 billion €, warns that the profitability of LGVs decreases as the system is built out: the LGV Sud-Est returned 15% to SNCF’s finances and 30% to French society (including rider consumer surplus), but subsequent lines only returned 4-7% to SNCF’s finances, and Bordeaux-Toulouse is likely to return less, 6% including social benefits per the study and at this point slightly less since the study assumed it would cost slightly less than the current budget.

The general theme in the French discourse on trains is that the TGV network is an obvious success. There absolutely is criticism, which focuses on the following issues:

  • Regional rail, that is not intercity rail, is underdeveloped in France outside Paris. The ridership of TER networks is pitiful in comparison with German-speaking and Nordic metropolitan areas of comparable size. For example, sourced to a dead link, Wikipedia claims 64,300 TER PACA trips per day, comprising the metropolitan areas of Marseille (1.8 million), Nice (1), Toulon (0.6), and Avignon (0.5); in Helsinki (1.5) alone, there are 200,000 daily commuter rail trips. But this isn’t really about high-speed rail, since TER planning and subsidies are devolved to regional governments, and not to SNCF.
  • SNCF has contentious labor relations. In the early 2010s, the unions went on a wave of strikes and got wage concessions that led to the evaporation of SNCF’s 600 million €/year primary surplus. The railway unions in France (“cheminots”) are unpopular, and Macron has been able to pass reforms to SNCF’s governance over their strikes and objections.
  • Future LGVs are not as strong as past ones. Real costs in France are rising, and the network already links Paris with all major secondary cities in airplane-competitive time save Nice. Interprovincial links on the network are weak, despite the construction of the LGV Rhin-Rhône, and nothing like the Deutschlandtakt is on the horizon enabling everywhere-to-everywhere travel.
  • SNCF thinks like an airline and not like a railroad. It separates passengers into different buckets as airlines do, has many executives with airline background (and Spinetta is ex-Air France), thinks passengers do not ride trains for longer than 3 hours even though at 4 hours the modal split with air is still better than 50-50, and has poor integration between the TGV and legacy rail.
  • SNCF still has a lot of accumulated debt from past operating losses, some predating the TGV and the start of regional subsidies for regional rail. It was hoped that TGV profits could cover them, but they can’t. This mirrors the controversy in Japan in the 1980s, where, in the breakup of JNR into the JRs and their privatization, debt from past operating losses was wiped but not debt from Shinkansen construction (see Privatization Best Practices, PDF-p. 106).

However, saying that the existing network is a failure is the domain of cranks and populists. It is unrecognizable from the discussion of transportation investments in France.

What O’Toole says about high-speed rail

O’Toole’s understanding of internal French (or Spanish, or Japanese) issues is weak. This isn’t surprising – Americans to a good approximation never have good insights on the internal issues of any other country, even when it speaks English. The American political sphere, which includes political thinktanks like Cato, is remarkably ignorant globally, and rather incurious. As a result, what he says about the TGV is based on an Americanized understanding. To wit:

Bus-rail competition

The Northeastern United States has a weak rail network: Amtrak averages vintage 1960s speeds and charges 2-4 times the per-km fare of the TGV. As a result, an ecosystem of private intercity buses has developed, starting with unregulated ones like Fung Wah and, as they were shut down, corporate systems like Megabus and Bolt. O’Toole is fond of these buses, with their lower fares and road-like lack of integration between infrastructure and operations.

And thus, he claims, falsely, that European high-speed rail cannibalized profitable buses. This is unrecognizable from within Europe, where intercity buses were underdeveloped until recently. In France, US-style intercity buses are called Macron buses, because the deregulation that brought them into existence passed in the mid-2010s, when Macron was the economy minister. They complement high-speed rail but do not replace it, because trains get me from Paris to the German border in 1:45 and buses don’t.

To be fair, TGV ridership has been stagnant in the last few years. But this stagnation goes back to the financial crisis, and if anything ridership picked up starting 2017 with the opening of the LGV Sud-Europe-Atlantique. So the buses are not even outcompeting the trains – they thrive in the gaps between them, just as historically they did on international routes, where rail fares are considerably higher and ridership lower.

High-speed rail construction costs

O’Toole looks at the most expensive few lines possible:

Britain’s 345‐​mile London–Scotland HS2 high‐​speed rail line was originally projected to cost £32.7 billion (about $123 million per mile) and is currently expected to cost £106 billion ($400 million per mile).

International comparisons of high-speed rail costs exist, and Britain’s costs are by far the worst. For example, a 2013 Australian comparison looking at the prospects for such a system in Australia finds that High-Speed 1/CTRL, the line linking the Channel Tunnel with London, cost A$134 million/km, and the second costliest line in the dataset was thee 94% tunneled Bologna-Florence line, at A$95 million/km.

French costs up until the LGV Bordeaux-Toulouse stood around $25-30 million per km in 2021 dollars, net of tunnels. German costs are similar, but German lines have far heavier tunneling than France, a range of 26-51% in tunnel compared with 0-6% in France. One reason is topography. But another is that Germany prefers mixed-use passenger-freight lines, which forces higher construction costs as freight requires gentler grades and, since superelevation must be lower, wider curves; France, like Japan and China, builds dedicated passenger lines, and, unlike Japan or China, keeps them largely at-grade to reduce costs.

O’Toole says, without more references, that it would cost $3-4 trillion to build a US-wide high-speed rail network. But the official Obama-era crayon, at 20,000 km, would be $500 billion at tunnel-free European costs, or maybe $600 billion with 5% tunneling, mostly in difficult places like California and across the Appalachians.

Freeway costs

O’Toole proposes more freeways, and says that to build the Interstate system today would cost $530 billion so it’s better than high-speed rail. Here is where his lack of knowledge of the most recent literature on infrastructure costs is a serious drag on his analysis: Brooks-Liscow establish that there was a large real increase in Interstate cost throughout the life of the program, so a budget that’s really a mixture of cheaper early-1960s construction and more expensive construction in the 1970s is not applicable today.

The same issue affects rail costs: the LGV Sud-Est cost, in today’s money, around $8 million/km, which cost would never recur. Brooks-Liscow explain this by greater surplus extraction from citizen voice groups, which demanded detours and route compromises raising costs. This appears true not just diachronically within the US but also synchronically across countries: so far, the low-cost subways we have investigated are all in states with bureaucratic rather than adversarial legalism, while medium-cost Germany is more mixed. Politicized demands leading to more tunneling are well-documented within Germany – the Berlin-Munich line was built through a topographically harder alignment in order to serve Erfurt, at Thuringia’s behest.

So no, today costs from the 1960s are not relevant. Today, urban motorway extensions cost double-digit millions of dollars per lane-km, sometimes more. The I-5 improvement project in Los Angeles is $1.9 billion for I-5 South, a distance of 11 km, adding two lanes (one HOV, one mixed traffic) in each direction. It’s possible to go lower than this – in Madrid this budget would buy a longer 6-lane tunnel – but then in Madrid the construction costs of rail are even lower, for both metros and high-speed lines.

The discourse on profits

In contrast with the basic picture I outlined for the TGV, French media and researchers often point out threats to rail profitability. This can easily be taken to mean that the TGV is unprofitable, and if one has an American mindset, then it’s especially easy to think this. If SNCF officials say that 20% of TGVs lose money, then surely they must be hiding something and the figure is much higher, right? Likewise, if Spinetta says that the TGV network is profitable but not all trains are, then surely the situation is even worse, right?

But no. This is an Americanized interpretation of the debate. In the US, Amtrak is under constant pressure to show book profits, and its very existence is threatened, often by people who cite O’Toole and other libertarians. Thus, as a survival strategy, Amtrak pretends it is more profitable than it really is.

This has no bearing on the behavior of railroads elsewhere, though. SNCF is not so threatened. The biggest threat from the perspective of SNCF management is union demands for higher wages, and therefore, its incentive is to cry poverty. Nobody in France takes out yardsticks of farebox recovery ratios, and therefore, nobody needs to orient their communications around what would satisfy American libertarians.

Energy

Within the European high-speed rail research community, the energy efficiency of high-speed rail is well-understood, and many studies look at real-world examples, for example the metastudy of Hasegawa-Nicholson-Roberts-Schmid. In fact, it’s understood that high-speed rail has lower energy consumption than conventional rail. For example, here is García Álvarez’s paper on the subject. This is counterintuitive, because higher speeds should surely lead to higher energy consumption, as Hasegawa et al demonstrate – but high-speed lines run at a uniform speed of 200 or 250 or 300 or 350 km/h, whereas legacy rail has many cycles of acceleration and deceleration. At speeds of up to about 200 km/h, nearly all electricity consumption is in acceleration and not maintaining constant speed, and even at 300 km/h, a late-model high-speed train consumes only above one third of its maximum power maintaining speed.

Instead of this literature, O’Toole picks out the fact that all else being equal energy consumption rises in speed, which it is not equal. Garcia in fact points out that higher speeds are better for the environment due to better competition with air, in line with environmental consensus that trains are far superior on well-to-wheels emissions to cars and planes. Worse, O’Toole is citing Chester-Horvath’s lifecycle analysis, which is not favorable to California High-Speed Rail’s energy efficiency. The only problem is that this paper’s analysis relies on a unit conversion error between BTUs and kWh, pointed out by Clem Tillier. The paper was eventually corrected, and with the correct figures, high-speed rail looks healthy.

Competition with cars and planes

Where high-speed rail exists, and the distance is within a well-understood range of around 300-800 km, it dominates travel. A 2004 report by Steer Davies Gleave has some profiles of what were then the world’s main networks. For Japan, it includes a graphic from 1998 on PDF-p. 120 of modal splits by distance. In the 500-700 km bucket, a slight majority of trips all over Japan are made by rail; this is because Tokyo-Osaka is within that range, and due to those cities’ size this city pair dominates pairs where rail is weaker, especially inter-island ones. In the 300-500 km bucket more people drive, but the Shinkansen is stronger than this on the Tokyo-Nagoya pair, it’s just that 300-500 includes many more peripheral links with no high-speed rail service. It goes without saying that high-speed rail does not get any ridership where it does not exist.

In France, this was also studied for the LGV PACA. On p. 14, the presentation lists modal splits as of 2009. Paris-Toulon, a city pair where the TGV takes around 4 hours, has an outright majority for the TGV, with 54% of the market, compared with 12% for air and 34% for driving. Paris-Cannes is 34% and Paris-Nice is 30%, both figures on the high side for their 5:00-5:30 train trips. Lyon-Nice, a 3:30 trip with awful frequency thanks to SNCF’s poor interprovincial service, still has a 25% market share for the TGV.

In general, competition with cars is understudied. Competition with planes is much more prominent in the literature, with plenty of reports on air-rail modal splits by train trip length. JR East, Central (PDF-p. 4), and West all report such market shares, omitting road transport. Many European analyses appeared in the 2000s, for example by Steer Davies Gleave again in 2006, but the links have rotted and Eurostat’s link is corrupt.

O’Toole misunderstands this literature. He lumps all air and road links, even on markets where rail is weak, sometimes for geographical factors such as mountains or islands, sometimes for fixable institutional ones like European borders. In fact, at least measured in greenhouse gas emission and not ridership, all air travel growth in Europe since 1990 has been international. International high-speed rail exists in Europe but charges higher fares and the infrastructure for it is often not built, with slowdowns in border zones. This is a good argument for completing the international network in Europe and a terrible one against building any network at all.

Topography

Even at the level of basic topography, O’Toole makes elementary errors. He discusses the Tokaido Shinkansen, pointing out its factor-of-2 cost overrun. But its absolute costs were not high, which he characterizes as,

The Tokyo–Osaka high‐​speed rail line supposedly made money, but it was built across fairly flat territory

So, first of all, the “supposedly” bit is painful given how much JR Central prints money. But “fairly flat territory” is equally bad. Japan’s mountainous topography is not an obscure fact. It’s visible from satellite image. Per Japanese Wikipedia, 13% of the route is in tunnel, more than California High-Speed Rail.

The United States can and should do better

The report is on stronger grounds when criticizing specifics of Amtrak and California High-Speed Rail. American rail construction is just bad. However, this is not because rail is bad; it’s because the United States is bad.

And there’s the rub. Americans in politics can’t tell themselves that another country does something better than the US does. If it’s in other countries and the US can’t do it, it must be, as O’Toole calls rail, obsolete. This is especially endemic to libertarians, who are intellectually detached from their European right-liberal counterparts (Dutch VVD, German FDP, etc.) even more than the American center-left is from social democrats here and the right is from the mainline and extreme right here.

So here, faced with not too hard to find evidence that high-speed rail is profitable in Europe and Asia, and in fact intercity rail is profitable here in general (direct subsidies are forbidden by EU law unless the line is classified as regional), unlike in the United States, O’Toole makes up reasons why trains here are unprofitable or unsuccessful. He says things that are not so much wrong as unrecognizable, regarding topography, buses, construction costs, debt, the state of the TGV debate, or greenhouse gas emissions.

O’Toole is aware of our transit costs comparison. I imagine he’s also aware of high-speed rail cost comparisons, which exist in the literature – if he’s not, it’s because he doesn’t want to be so aware. And yet, no matter how loudly the evidence screams “the United States needs to become more like France, Germany, Japan, Spain, etc.,” American libertarians always find excuses why this is bad or unnecessary. And then, when it comes to expanding freeways, suddenly the cost concerns go out the door and they use unrealistically low cost figures.

But figuring out why the US is bad requires way deeper dives. It requires delving into the field and understanding how procurement is done differently, what is wrong with Amtrak, what is wrong with the California High-Speed Rail Authority, how engineering is done in low- and medium-cost countries, various tradeoffs for planning lead time, and so on. It requires turning into the kind of expert that libertarians have spent the last 60 years theorizing why they need not listen to (“public choice”). And it requires a lot of knowledge of internal affairs of successful examples, none of which is in an English-speaking country. So it’s easier to call this obsolete just because incurious Americans can’t do it.

Amtrak’s Continued Ignorance

There was a congressional hearing about high-speed rail. Henry Miller in comments here took notes – thanks for this, much appreciated! The overall content was lacking; the politicians seemed like they were spinning their wheels, not because they themselves were bad (Reps. Tom Malinowski, Peter DeFazio, and Seth Moulton all raised interesting issues) but because they were getting ignorant advice from the witnesses, none of whom has any experience in successful high-speed rail networks. Among those, Amtrak deserves the most demerits, and its head, William Flynn, should lose his job purely over that testimony, if the reporting of what he said is accurate.

Flynn, based on both what Henry said in comments and on reporting in Politico Pro, said that Amtrak needs a trust fund on the model of that for American highways – and said that this is “the most important lesson we can learn” from countries with high-speed rail.

The rub is that countries with high-speed rail do not in fact have such trust funds. Financing models vary by country, but do not look like the American highway trust fund. For example, French LGVs are funded line-by-line, with the decision on each specific line taken at the highest level of government, with financing coming either purely from the public sector (as with the LGV Est) or from a higher-cost PPP (as with the LGV Sud-Europe-Atlantique).

To understand why, it’s important to understand the relationship between politics and the civil service in functioning, high-capacity states. Politicians make big decisions on spending priorities, and then the civil service implements those decisions. There is little political input on routing decisions, and the exceptions where there is tend to have the worst, highest-cost programs. So the planning is done by the civil service, which then presents a preliminary design for politicians. But the elected politicians have the final word on the yes-no decision whether to fund, and can also ask for high-level modifications (“reduce the budget,” “give the unions the wage increases they demand,” etc.).

The American highway trust fund inverts this principle. Going back to Thomas MacDonald, federal highway builders had internal sources of money without having to ask elected politicians for regular appropriations. In contrast, politicians exerted considerably petty power over routing. For example, in Twentieth-Century Sprawl, Owen Gutfreund points out that in the early planning for what became the Interstate highways, the FDR administration reduced the scope of roads to be built in Vermont from four planned routes to two in retaliation for its voting Republican in 1936. In The Big Roads, Earl Swift also notes that MacDonald himself did not think the Interstates could pay for themselves through tolls, but, due to pressure by politicians to write a positive report, the resulting report’s coauthor proposed toll-free motorways instead, hence the prohibition on tolling Interstates. MacDonald himself was fired by the Eisenhower administration for expressing concern that the roads were hollowing out the American rail network and proposing cars-and-trains investment instead of cars-only.

And here we have Amtrak’s CEO not only supporting that model, but also lying that this model is how high-speed rail has been built. In reality, no such trust funds exist anywhere with high-speed rail. I don’t know why Flynn says such a thing, which not only is verifiably wrong, but also has no reason to be believed in the first place – there is no grain of truth to it, no trust fund-like model for high-speed rail megaprojects.

As with most such fraud, he is probably lying to himself and not just to the people who pay his salary. Americans, as a collective, are wantonly ignorant of the rest of the world. The only time they interact with the rest of the world, especially countries that don’t speak English, is through intermediaries in international consulting, who get the skewed sample of world projects that invite in international consultants, omitting the bulk of public works built in states with in-house design capacity. Individual Americans can be knowledgeable, but their knowledge is not respected, even by people who profess their interest in state capacity. Thus, no matter how smart individual Americans can get, collectively America remains incurious.

This is the most acute in mainline rail. I suspect that this relates to the rail industry’s highway envy. For a railroader like Flynn, steeped in a culture that is technologically and institutionally reactionary and looks back to its heyday in the first half of the 20th century, the enemy, that is the Interstate system, is the obvious model for how to build. That this model produced severe cost overruns on the highways themselves does not matter; that treating rails institutionally like roads is inappropriate does not matter; that systems that get as much ridership in two days (cf. JR East) as Amtrak gets in an entire year and deliver a profit to their shareholders doing so work differently does not matter. The future, which is not in the United States in this field and hasn’t been in 60 years, is one in which people like Flynn do not even qualify for an internship.

And if Flynn wouldn’t qualify for an internship, why is he allowed to be the CEO? He should lose his job. The people who briefed him should lose their jobs. It is likely that full replacement of Amtrak’s planning staff and possibly the line workers too would be a big win for riders. Even total liquidation could well be a net positive relative to status quo: most Amtrak routes have no social value, and the one route that does, the Northeast Corridor, could well produce a more competent institution from among the ashes.

Without liquidation, it is still advisable to sideline Amtrak until it can be put out of its delayed customers’ misery. The best way forward institutionally is to set up an agency responsible for all Northeastern passenger rail operations, to subsume and replace Amtrak and the commuter rail operators. It will be run by people who can speak to the difference between French, German, and Japanese high-speed rail operating models, and who know how to implement integrated timed transfer networks and intermodal fare integration. It will buy imported equipment if there is no domestic equivalent for a similar price, and use standard European or East Asian methods for track geometry machines, signaling (ACSES is thankfully an Americanized variant of the European standard, ETCS), safety systems, timetabling, and so on. The United States has no shortage of dedicated people who speak Spanish, and secondarily Japanese, Korean, Chinese, Italian, German, or French.

Moreover, since in many cases the knowledge does exist among Americans but isn’t valued, it is important to let American civil servants interview for such an agency. I expect that most would come from an urban transit background, where in my experience the people are more curious than in mainline rail. But American railroaders too could join if they demonstrate sufficient knowledge of advanced-world operations.

That said, under no circumstances should the organizational culture be allowed to turn into anything like present-day American railroading. Current workers who do not qualify for this agency are to be laid off, perhaps with a pro-rated pension for partial service, and told to seek private-sector work. Flynn himself has no role to play in any successful rail agency. He must go, and it’s almost certain that the rest of Amtrak’s management should as well. Every day he stays in his job is a day American railroading plans based on assumptions that can be easily verified to be fraudulent.

The United States Needs to Learn How to Learn

I just saw an announcement from November of 2020 in which the Federal Transit Administration proposes to study international best practices… in on-demand public transit.

It goes without saying that the international best practice in on-demand micromobility is “don’t.” The strongest urban public transport networks that I know of range from not making any use of it to only doing so peripherally, like Berlin. In fact, both France and Germany have rules on taxis that forbid Uber from pricing itself below the regulated rates; Japan, too, banned Uber from operating after it tried to engage in the usual adversarial games with the state that it is so familiar with from the US.

And yet, here we see an FTA program attempting to learn from other countries not how to write a rail timetable, or how to modernize regional rail, or how to design a coordinated infrastructure plan, or how to integrate fares, or how to do intermodal service planning, or how to build subways affordably. It’s perhaps not even aware of those and other concepts that make the difference between the 40% modal split of so many big and medium-size European cities and the 10-15% modal splits that non-New York American cities top at.

Instead, the FTA is asking about a peripheral technology that markets itself very aggressively to shareholders and VCs so that it can ask for more money to fund its losses.

Earlier today I saw a new announcement of congressional hearings about high-speed rail. There are 12 witnesses on the list, of whom none has any experience with actual high-speed rail. They’re American politicians plus people who either run low-speed trains (Amtrak, Brightline) or promise new vaporware technology (Hyperloop*2, Northeast Maglev). American politicians and their staffers are not that stupid, and know that there are strong HSR programs in various European and Asian countries, and yet, in the age of Zoom, they did not think to bring in executives from JR East, DB, SNCF, SBB, etc., or historians of these systems, to discuss their challenges and recommendations.

I bring up these two different examples from the FTA and Congress because the US has trouble with learning from other places. It’s not just that it barely recognizes it needs to do so; it’s that, having not done so in the past, it does not know how to. It does not know how to form an exchange program, or what questions to ask, or what implementation details to focus on. Hearing of a problem with a public agency, its first instinct is to privatize the state to a consultancy staffed by the agency’s retirees, who have the same groupthink of the current publicly-employed managers but collect a higher paycheck for worse advice.

Worse, this is a nationwide problem. Amtrak can and should fully replace its senior management with people who know how to run a modern intercity railroads, who are not Americans. But then middle management will still think it knows better and refuse to learn what a tropical algebra is or how it is significant for rail schedule planning. They do not know how to learn, and they do not recognize that it’s a problem. This percolates down to planners and line workers, and I don’t think Americans are ready for a conversation about full workforce replacement at underperforming agencies.

This will not improve as long as the United States does not reduce its level of pride to that typical of Southern Europe or Turkey. When you’re this far behind, you cannot be proud. It’s hard with American wages being this high – the useless managers even in the public sector earn more than their Northern European counterparts and therefore will not naturally find Northern Europe to have any soft power over them. Wearing sackcloth and ashes comes more naturally with Italian or Spanish wages. But it’s necessary given how far behind the US is, and bringing in people who are an American’s ideal of what a manager ought to be rather than people who know how to run a high-speed passenger railroad is a step backward.

Sanity-Checking My HSR Ridership Model

In previous posts about modeling high-speed rail ridership, I used a gravity model for the estimation. While poking around with spreadsheets, I figured out that a good way to sanity-check the model is to run it on existing high-speed rail systems with known ridership. It turns out that the model fits the data decently but not amazingly, and tends to overestimate ridership at long distances (800 km+) and underestimate it at short ones.

The model

The model I use is a gravity model, with the constant trained on Shinkansen data from JR East and JR Central (PDF-p. 4):

\mbox{Ridership } = 75000\cdot\mbox{Pop}_{A}^{0.8}\cdot\mbox{Pop}_{B}^{0.8}/\max\{\mbox{distance}, 500\}^{2}

The populations of metro areas A and B are in millions, distance is in km, and ridership is in millions per year in both directions combined.

He blocks the door and doesn’t move to let people on or off. Credit: Barroa Artworks.

The data

I’ve tested the model on two datasets: Shinkansen, and Taiwan HSR. These are island systems with a finite, controllable number of stations; Taiwan, a single-line system, is especially easy to model. The km-points are taken from line lengths; but mini-Shinkansen lines have artificially inflated lengths to account for the greater travel time, by a factor of about 2.7, to be compatible with an average express train speed of about 220 km/h. This means the model will overrate their passenger-km, but it’s not a significant source of error as they are fairly small cities – were they bigger they’d get full Shinkansen.

Metro areas are combined, and when a metro area has several stations, they are merged and only the most prominent is depicted, such as Tokyo, Shin-Osaka, and Taipei.. In Japan I use the broader category of major metropolitan area wherever possible, with the exception of Shizuoka-Hamamatsu, which are not merged as they were distinct until recently and remain two separate city cores that only share suburbs on the margins. Otherwise I use the smaller metropolitan employment area, as the MMA is only defined for the largest cities, and not for (say) Aomori or Kanazawa.

In Taiwan there’s no real definition of metro area. The secondary cities are single-tier municipalities encompassing the metro area plus some rural areas; I take what Wikipedia calls the urban part, which is nearly the same as the municipality. Taipei and New Taipei are merged – there’s a stop in New Taipei but New Taipei is really a suburb of Taipei spreading in all directions; but Taoyuan is kept separate, as it tries to develop its own core and lies only in one direction from Taipei, to its west. Outside the cities I use county populations where the stop seems to serve the center of the county, but Chiayi is expansive and I focus on the independent Chiayi City plus the suburb the station is in, and Changhua’s station is very peripheral to the county, most of which is closer to Taichung.

Both countries charge similar fares – Wikipedia has Taiwan charging, in PPP terms, $0.25/p-km, which is close to the Shinkansen average, and compares with about $0.15/p-km in Continental Europe. In addition, both have linear population distribution, Japan along the Taiheiyo Belt and Taiwan along the west coast.

Results

The model massively underrates the ridership of THSR. It believes ridership is 26 million a year, with a total of 4.465 billion p-km; the actual numbers are 67 million and 12 billion respectively as of 2019, per Wikipedia. I have not seen ridership by city pairs, only boardings per station. The numbers do not make it obvious if there is more very short-distance ridership than I expect. The average trip length I predict is 172 km; the actual average is 178. Taichung has slightly more ridership than Zuoying, where in reality Taichung and Kaohsiung have the same populations, but Zuoying is not quite at city center whereas Taichung also draws from Changhua County, whereas the Changhua station has very low ridership. Overall, to the extent the shape of the model is correct, the minimum of 500 km in the denominator cannot be too wrong – or, if it is, the minimum must be more than the Taipei-Kaohsiung distance of 339 km or not much less than it.

In Japan, the situation is less clear. Total Shinkansen ridership is 438 million as of financial year 2018-9, per Wikipedia; this is the last year before corona, as the years end on 3-31 and in March of 2020 Japanese ridership was already suppressed due to social distancing. Passenger-km on JR East, JR Central, and JR West totaled around 100 billion, with Hokkaido and Kyushu adding scant numbers, but these are railroad-km, and the Shinkansen charges based on the distance along the legacy line and not the Shinkansen, inflating p-km by somewhat less than 10%.

In contrast, my model thinks total Shinkansen ridership is 389 million and p-km sum to 170.815 billion. The 389 vs. 438 discrepancy is easy to explain – my model ignores intra-metropolitan trips, and we know that they exist because there are some Shinkansen commuters in towns like Mishima. However, 100 vs. 171 billion p-km is harder. For this, there are several explanations, all plausible, and yet none completely satisfactory:

  • About 40 billion of the p-km involve riding through Tokyo, of which 21 billion are from the Tohoku Shinkansen and 19 from the Joetsu and Hokuriku Shinkansen. There are no through-trains, and the through-trips via Joetsu and especially Hokuriku are circuitous.
  • Yamagata and Akita between them generate around 6 billion p-km per the model; this is an overestimate, as the spreadsheet does not distinguish km that are really stand-ins for trip time from km that are actually traveled.
  • A total of 6.5 billion p-km per the model are diagonal between the Tohoku, Joetsu, and Hokuriku Shinkansen; in reality, connecting at Omiya or Takasaki is so circuitous that I expect nearly everyone drives.
  • Inter-island trips are especially likely to be done by air. Tokyo-Fukuoka has a rail-air modal split of 7.4-92.6, over a distance of 5 hours, and Nagoya-Fukuoka is only 51-49, over a distance of 3:20. This is bad for rail by European standards, where 5 hours is typically 20-30% for rail and 3:20 is a clear majority, and even by intra-Honshu Japanese standards, where Tokyo-Hiroshima at 3:55 is 68-32 and Tokyo-Okayama at 3:15 is 70-30.

All trip categories above are disproportionately long, helping explain why the model underpredicts ridership while overpredicting p-km. Subtracting all of the above one gets to not much more than 100 billion.

The model does nail certain aspects of Shinkansen ridership. Tokyo-Sendai, Tokyo-Hiroshima, and Tokyo-Okayama are easy – the model was trained in part on those specific city pairs. But in adition, overall ridership out of Tokyo and Osaka is very close to total JR Central ridership in these two regions. The model slightly overpredicts Osaka but that is expected since it lumps the Keihanshin region together whereas JR Central would not count Kobe.

Nagoya is more overpredicted, and it is possible that it is uniquely auto-oriented and this slightly reduces rail ridership, by maybe 25% below modeled prediction. If that is what is happening, then the constant 500 in the denominator of the model as well as 75,000 in the numerator should be adjusted – the reason for the choice of 500 is that Tokyo-Nagoya and Tokyo-Osaka ridership levels both follow the same model if the exponent is 0.8 and distance is ignored; if in fact Nagoya has a 25% malus then to countermand it the constant in the maximum should be lowered slightly, to 430 or a little less.

It’s tempting to rewrite the model in terms of travel time and then set the constant at 2 hours (and not 2.5 hours as I did when trying to model Germany). But note that it’s far from enough to explain the model’s gross underprediction of Taiwanese HSR ridership, an underprediction that exists across all distances in Taiwan. Nor is it possible to lower the 75,000 constant in the numerator and address any of the underprediction of Taiwan.

Tilting Trains and Technological Dead-Ends

The history of tilting trains is on my mind, because it’s easy to take a technological advance and declare it a solution to a problem without first producing it at scale. I know that 10 years ago I was a big fan of tilting trains in comments and early posts, based on both academic literature on the subject and existing practices. Unfortunately, this turned into a technological dead-end because the maintenance costs were too high, disproportionate to the real speed benefits, and further work has gone in different directions. I bring this up because it’s a good example of how even a solution that has been proven to work at scale can turn out to be a dead-end.

What is tilting?

It is a way of getting trains to run at higher cant deficiency.

What is cant deficiency?

Okay. Let’s derive this from physical first principles.

The lateral acceleration on a train going on a curve is given by the formula a = v^2/r. For example, if the speed is 180 km/h, which is 50 m/s, and the curve radius is 2,000 meters, then the acceleration is 50^2/2000 = 1.25 m/s^2.

Now, on pretty much any curve, a road or railway will be banked, with the outer side elevated above the inner side. On a railway this is not called banking, but rather superelevation or cant. That way, gravity countermands some of the centrifugal force felt by the train. The formula on standard-gauge track is that 150 mm of cant equal 1 m/s^2 of lateral acceleration. The cant is free speed – if the train is perfectly canted then there is no centrifugal force felt by the passengers or the train systems, and the balance between the force on the inner and outer rail is perfect, as if there is no curve at all.

The maximum superelevation on a railway is 200 mm, but it only exists on some Shinkansen lines. More typical of high-speed rail is 160-180 mm, and on conventional rail the range is more like 130-160; moreover, if trains are expected to run at low speed, for example if the line is dominated by slow freight traffic or sometimes even if the railroad just hasn’t bothered increasing the speed limit, cant will be even lower, down to 50-80 mm on many American examples. Therefore, on passenger trains, it is always desirable to run faster, that is to combine the cant with some lateral acceleration felt by the passengers. Wikipedia has a force diagram:

The resultant force, the downward-pointing green arrow, doesn’t point directly toward the train floor, because the train goes faster than the balance speed. This is fine – some lateral acceleration is acceptable. This can be expressed in units of acceleration, that is v^2/r with the contribution of cant netted out, but in regulations it’s instead expressed in theoretical additional superelevation required to balance, that is in mm (or inches, in the US). This is called cant deficiency, unbalanced superelevation, or underbalance, and follows the same 150 mm = 1 m/s^2 formula on standard-gauge track.

Note also that it is possible to have cant excess, that is negative cant deficiency. This occurs when the cant chosen for a curve is a compromise between faster and slower trains, and the slower trains are so much slower the direction of the net force is toward the inner rail and not the outer rail. This is a common occurrence when passenger and freight trains share a line owned by a passenger rail-centric authority (a freight rail-centric one will just set the cant for freight balance). It can also occur when local and express passenger trains share a line – there are some canted curves at stations in southeastern Connecticut on the Northeast Corridor.

The maximum cant deficiency is ordinarily in the 130-160 mm range, depending on the national regulations. So ordinarily, you add up the maximum cant and cant deficiency and get a lateral acceleration of about 2 m/s^2, which is what I base all of my regional rail timetables on.

You may also note that the net force vector is not just of different direction from the vertical relative to the carbody but also of slightly greater magnitude. This is an issue I cited as a problem for Hyperloop, which intends to use far higher cant than a regular train, but at the scale of a regular train, it is not relevant. The magnitude of a vector consisting of a 9.8 m/s^2 weight force and a 2 m/s^2 centrifugal force is 10 m/s^2.

Okay, so how does tilt interact with this?

To understand tilt, first we need to understand the issue of suspension.

A good example of suspension in action is American regulations on cant deficiency. As of the early 2010s, the FRA regulations depend on train testing, but are in practice, 6″, or about 150 mm. But previously the blanket rule was 3″, with 4-5″ allowed only by exception, mocked by 2000s-era advocates as “the magic high-speed rail waiver.” This is a matter of carbody suspension, which can be readily seen in the force diagram in the above secetion, in which the train rests on springs.

The issue with suspension is that, because the carbody is sprung, it is subject to centrifugal force, and will naturally suspend to the outside of the curve. In the following diagram, the train is moving away from the viewer and turning left, so the inside rail is on the left and the the outside rail is on the right:

The cant is 150 mm, and the cant deficiency is held to be 150 mm as well, but the carbody sways a few degrees (about 3) to the outside of the curve, which adds to the perceived lateral acceleration, increasing it from 1 m/s^2 to about 1.5. This is typical of a modern passenger train; the old FRA regulations on the matter were based on an experiment from the 1950s using New Haven Railroad trains with unusually soft suspension, tilting so far to the outside of the curve that even 3″ cant deficiency was enough to produce about 1.5 m/s^2 of lateral force felt by the passengers.

By the same token, a train with theoretically perfectly rigid suspension could have 225 mm of cant deficiency and satisfy regulators, but such a train doesn’t quite exist.

Here comes tilt. Tilt is a mechanism that shifts the springs so that the carbody leans not to the outside of the curve but to its inside. The Pendolino technology is theoretically capable of 300 mm of cant deficiency, and practically of 270. This does not mean passengers feel 1.8-2 m/s^2 of lateral acceleration; the train’s bogies feel that, but are designed to be capable of running safely, while the passengers feel far less. In fact the Pendolino had to limit the tilt just to make sure passengers would feel some lateral acceleration, because it was capable of reducing the carbody centrifugal force to zero and this led to motion sickness as passengers saw the horizon rise and fall without any centrifugal force giving motion cues.

Two lower cant deficiency-technology than Pendolino-style tilt are notable, as those are not technological dead-ends, and in fact remain in production. Those are the Talgo and the Shinkansen active suspension. The Talgo has no axles, and incorporates a gravity-based pendular system in which the train is sprung not from the bottom up but from the top down; this still isn’t enough to permit 225 mm of cant deficiency, but high-speed versions like the AVRIL permit 180, which is respectable. The Shinkansen active suspension is computer-controlled, like the Pendolino, but only tilts 2 degrees, allowing up to 180 mm of cant deficiency.

What is the use case of tilting, then?

Well, the speed is higher. How much higher the speed is depends on the underlying cant. The active tilt systems developed for the Pendolino, the Advanced Passenger Train, and ICE T are fundamentally designed for mixed-traffic lines. On those lines, there is no chance of superelevating the curves 200 mm – one freight locomotive at cant excess would demolish the inner track, and the freight loads would shift unacceptably toward the inner rail. A more realistic cant if there is much slow freight traffic is 80 mm, in which case the difference between 150 and 300 mm of cant deficiency corresponds to a speed ratio of \sqrt{(80+300)/(80+150)} = 1.285.

Note that the square root in the formula, coming from the fact that acceleration formula contains a square of the speed, means that the higher the cant, the less we care about cant deficiency. Moreover, at very high speed, 300 mm of cant deficiency, already problematic at medium speed (the Pendolino had to be derated to 270), is unstable when there is significant wind. Martin Lindahl’s thesis, the first link in the introduction, runs computer simulations at 350 km/h and finds that, with safety margins incorporated, the maximum feasible cant deficiency is 250 mm. On dedicated high-speed track, the speed ratio is then \sqrt{(200+250)/(200+130)} = 1.168, a more modest ratio than on mixed track.

The result is that for very high-speed rail applications, Pendolino-level tilting was never developed. The maximum cant deficiency on a production train capable of running at 300 km/h or faster is 9″ (230 mm) on the Avelia Liberty, a bespoke train that cost about double per car what 300 km/h trains cost in Europe. To speed up legacy Shinkansen lines, JR Central and JR East have developed active suspension, stretching the 2.5 km curves of the Tokaido Shinkansen from the 1950s and 60s to allow 285 km/h with the latest N700 trains, and allowing 360 km/h on the 4 km curves of the Tohoku Shinkansen.

What happened to the Pendolino?

The Pendolino and similar trains, such as the ICE T, have faced high maintenance costs. Active tilting taxes the train’s mechanics, and it’s inherently a compromise between maintenance costs and cant deficiency – this is why the Pendolino runs at 270 mm where it was originally capable of 300 mm. The Shinkansen’s active suspension is explicitly a compromise between costs and speed, tilted toward lower cant deficiency because the trains are used on high-superelevation lines. The Talgo’s passive tilt system is much easier to maintain, but also permits a smaller tilt angle.

The Pendolino itself is a fine product, with the tilt removed. Alstom uses it as its standard 250 km/h train, at lower cost than 350 km/h trains. It runs in China as CRH5, and Poland bought a non-tilting Pendolino fleet for its high-speed rail service.

Other medium-speed tilt trains still run, but the maintenance costs are high to the point that future orders are unlikely to include tilt. Germany has a handful of tilt trains included in the Deutschlandtakt, but the market for them is small. Sweden is happy with the X2000, but its next speedup of intercity rail will not involve tilting trains on mostly legacy track as Lindahl’s thesis investigated, but conventional non-tilting high-speed trains on new 320 km/h track to be built at a cost that is low by any global standard but still high for how small and sparsely-populated Sweden is.

In contrast, trainsets with 180 mm cant deficiency are still going strong. JR Central recently increased the maximum speed on the Tokaido Shinkansen from 270 to 285 km/h, and Talgo keeps churning out equipment and exports some of it outside Spain.

High-Speed Rail and Connecting Transit

Noah Smith is skeptical about high-speed rail in the United States. He makes a bunch of different arguments against it, but I want to zoom in on the first, the issue of connecting transit, which Noah is far from the first person to bring up. It’s a genuine drawback of rail planning in the United States, but it’s very easy to overrate its importance. Connecting transit is useful, as is the related issue of city centralization, but its effect, serious as it is, is only on already marginal high-speed routes, like Atlanta-Memphis or Dallas-Kansas City. Los Angeles suffers from lacking connecting transit, but it’s also so big that nothing it connects to is marginal. Finally, high-speed rail and urban centralization are not in competition, but rather are complements, as in the history of the TGV.

Connections and centralization

Modal choice is about door-to-door trip times. This is why a large majority of people take a train that takes three hours over a plane that takes one: hardly anyone lives near the airport or has an airport as their ultimate destination. In practice, people are much likelier to be living near and traveling to a destination near a city center station.

The importance of connections then is that connecting urban transit extends the range of the train station. I didn’t live at Gare de Lyon or Gare de l’Est, but I could take the Métro there and it was a short trip, much shorter and more reliable than taking the RER to the airport, which made it easier for me to ride the TGV. With reliable connections, I showed up at Gare de l’Est four minutes before a train to Saarbrücken was due to depart, printed my ticket on-site, and walked leisurely to the platform, boarding still with two minutes to spare.

Regional rail has the same effect, at longer range. It’s not as convenient as urban rail, but it feeds the main intercity rail station and is timetabled, so if the system is punctual, passengers can time themselves to the main train station. In Switzerland the connections are even timed, enabling people who travel from smaller cities like St. Gallen to points west to transfer at Zurich Hauptbahnhof within a short window. However, this is completely absent from France: the regional trains are unreliable, and Paris has through-running on the RER but no single central station that can collect connections from secondary centers like Meaux or Versailles.

Finally, centralization is important because the reach of an urban transportation system is measured in units of time and not distance. Even racists who are afraid of taking the trains in Paris and rely exclusively on cars can take a cab from a train station to their ultimate destination and be there shortly. The average speed of the Métro is low, around 25 km/h, but Paris’s density and centralization mean that it’s enough to connect from the main TGV stations to where one lives or works.

But the US doesn’t have that, right?

What Noah gets wrong is that the US has connecting transit as in Paris in a number of big cities, and nearly every even semi-plausible high-speed line connects to at least one such city. Here’s Noah on New York:

The best thing about using the Shinkansen in Japan is that you can get to and from the high-speed rail station using a dense, convenient network of local trains. In America there is no such network. Thus, when I imagine taking the train from SF to L.A., I imagine taking a scooter or an Uber to and from the train station. In L.A., which is so spread out that I probably won’t stay in a small area, I imagine I’d rent a car. That’s a very different experience from using the Shinkansen in Japan. And in NYC, it would mean dealing with the nightmare that is Penn Station — a thoroughly stressful and inconvenient experience.

Let’s discuss New York now; Los Angeles deserves a separate section in this post. Noah lived on Long Island for years; he could connect to any intercity train by taking the LIRR to Penn Station and changing there. It’s this connection that he describes as a nightmare. But the question is, a nightmare compared to what? It’s clearly far less convenient than the timed Swiss connections, or even untimed connections between the Berlin S-Bahn and intercity trains. But the LIRR is a timetabled train, and while delays happen, they’re measured in minutes, not tens of minutes. Passengers can time themselves to arrive 10 minutes before the intercity train departs, even today.

All of this gets easier if a minimally competent agency is in charge and track numbers are scheduled in advance and printed on the ticket as they are here or in Japan. Penn Station is crowded, but it’s not a stampede crush and people who know their commuter train arrives on track 19 and the intercity train leaves on track 14, as written in the ticket, can make the connection in 3 minutes.

The secondary transit cities of the US are dicier. Their modal splits are all in the teens; San Francisco (excluding Silicon Valley) is the highest, with 17.5%. In that way, they’re comparable to Lyon, Marseille, Nice, Bordeaux, Toulouse, Strasbourg, and Lille. However, the way non-New York transit systems work in the US is, the system is usually semi-decent at ferrying people to and from city center, it’s just not strong for other destinations. In Boston, for example, people could transfer to the subway at South Station or Back Bay and cover a decent chunk of urban destinations. It’s nowhere nearly as good as the options for Paris or Berlin, but it’s not the same as not having any connecting transit.

Destination centralization

The connecting transit critique of high-speed rail in the American discourse goes back at least to the Obama era; Richard Mlynarik used it to argue against what he views as inflated California HSR ridership expectations, and everyone who commented on transit blogs in 2008-9 had to address the critique in one way or another. In 2012, I posted about the issue of destination centralization, that is, that destinations are more centralized than origins, especially at long distance. For example, at the time Manhattan had 22% of New York metro jobs, but 36% of jobs involving out-of-county commuting – and the longer the trip, the likelier one’s destination is to be in Manhattan.

The data I looked at was the distribution of five-star hotels, which are incredibly centralized. Depending on data sources, 50 out of 56 such hotels in metro New York were in Manhattan, or perhaps 36 in 37. In Boston, either all are in Downtown or Back Bay, or all but one are and the one is in Cambridge, a few Red Line stops from South Station. In Philadelphia, they’re in Center City.

In New York, there are clusters of lower-priced hotels outside Manhattan. The biggest such clusters are in strategic locations in Queens, Brooklyn, or North Jersey with maximally convenient access to Manhattan, where tourists and business travelers cluster. Some hotels serve suburban office parks, such as the various Central Jersey hotels I would go to gaming conventions at, but they’re smaller and lower-end.

In the Bay Area, Richard argued in favor of the primacy of San Francisco over San Jose by citing broader data on interregional travel. San Francisco, per his dataset, absolutely dominated. More recent data can be seen here, measuring tourism revenue rather than visitor numbers, but San Francisco with 900,000 people is about comparable to Santa Clara, Alameda, and San Mateo Counties combined with their 4.4 million people. There is also a comparison of international arrivals to San Jose and San Francisco – there are several times as many of the latter; I cannot find domestic arrival numbers for San Jose that might compare with San Francisco’s 26 million visitors in 2019.

The upshot is that high-speed rail does not need to connect two strongly-centered cities to be comparable in ridership to existing lines in Europe and East Asia. It only needs to connect one. People may need to drive to a park-and-ride or take a taxi to the train station, but if their destination is New York or any of the secondary transit cities of the US, it is likely to be fairly close to the train station, even if most employment isn’t.

The Los Angeles exception

Noah is on stronger grounds when he criticizes Los Angeles. Even Los Angeles has 1.5 subway lines connecting to Union Station, soon to be augmented with the Regional Connector, but the city is weakly-centered, and a car or taxi connection to one’s ultimate destination is likely. Moreover, the destinations within Los Angeles are not centered on Downtown; for example, high-end hotels are the most likely to be found on the Westside.

However, there are two saving graces for trains to Los Angeles. The first is that Los Angeles’s transit ridership is so low because the city’s job geography is so decentralized that the network is bad at connecting local origins with local destinations. If it is guaranteed that one of the two points connected is Union Station, the city’s network is still bad for its size, but becomes usable. The under-construction Westside subway will open later this decade, providing decent (if not good) connectivity from the train station to high-end destinations in that part of the region.

The second and more important saving grace is that Los Angeles is huge. The absence of connecting transit is a serious malus for intercity rail, but people can still take a taxi, and that may add half an hour to the trip and a cab fare, but we know what adding half an hour to a three-hour train trip does and it’s a 1.5th-order effect. A 1.5th-order effect can turn a line that is projected to get a marginal 2.5% return on investment into one with a below-cost-of-capital 1.5% return. It cannot do this to lines serving Los Angeles, none of which are economically marginal, thanks to Los Angeles’s size. On my map, the only line connecting to Los Angeles that a straight gravity model doesn’t love at first sight is Los Angeles-Las Vegas, and this is a connection we know overperforms the model because of the unique tourism draw of Las Vegas.

On the same map, the other connection that everyone (including myself until I ran the number) is skeptical of, Atlanta-Florida, has the same issue as Los Angeles-Las Vegas: it connects to a very strong tourism region, and the train station would serve the biggest tourist attractions. (This is also true in the case of Los Angeles, where Anaheim is still supposed to get a station within a short shuttle distance to Disneyland.) So my model thinks it’s only 2.5% ROI, but the strong tourism volume is such that I am confident the model remains correct even with the malus for weak job centralization in both Atlanta and the cities of Florida.

High-Speed Rail and Connecting Transit

Noah makes a broader point portraying intercity and regional public transport in opposition:

Building high-speed rail without having a usable network of local trains instinctively feels like putting the cart before the horse. If I had a choice between being able to train around San Francisco conveniently, or quickly get between SF and San Jose, I’d choose either of those over being able to take a Shinkansen-style train to L.A. or Seattle. The lack of local trains and fast commuter rail simply limits my travel options much more than the lack of high-speed rail. A local train network without HSR is great; HSR lines without local trains seem like something that’s at best slightly better than what we have now.

And yes, I realize that money earmarked for “high-speed rail” sometimes goes to create faster commuter rail, and that’s good. But that doesn’t answer the question of what these maps are for.

Noah is pooh-poohing the connection between intercity and regional transit as “the money sometimes goes to create faster commuter rail,” but he’s underestimating what this means, in two ways.

First, on the Northeast Corridor specifically, any improvement to intercity transit automatically improves commuter rail. The reason is that the most cost-effective speed treatments there are shared. By far the cheapest minutes saved on the corridor come from speeding up the station throats by installing more modern turnouts and removing speed limits that exist due to agency inertia rather than the state of the physical infrastructure. Trains can save two minutes between South Station and Back Bay alone on a high seven to low eight figures budget for rebuilding the interlocking. These improvements speed up commuter rail and intercity rail equally.

Moreover, in higher speed zones, it’s necessary to invest in organization before concrete and schedule trains with timed overtakes. But this too improves the quality of regional rail. Boston-Providence trains need to be electrified and run faster to get out of intercity trains’ way more easily; even with trains holding twice for an overtake, this speeds up Providence-Boston travel by 15 minutes even while adding station stops. New York-New Haven trains had better run faster on both short- and long-distance connections – and the difference between improving intercity rail this way and in a way that is indifferent to integration with regional rail is the difference between doing it for $15 billion and doing it for $150 billion.

And second, in cities that are not traditional transit cities, high-speed rail is a really good catalyst for expanding a central business district around the station. The best example for this is Lyon. Lyon built a dedicated central business district at Part-Dieu, the Metro, and the LGV Sud-Est simultaneously. This was not sequenced as local transit first, then high-speed rail. Rather, the selection of the site for a high-speed rail station, within the city but just outside its traditional center, was simultaneous with the construction of the new business district and of an urban rail system serving it.

This is particularly useful for cities that, by virtue of size (Dallas) or location (Cleveland) could be high-speed rail hubs but do not have strong city centers. In Cleveland, demand for housing in the city is extremely weak, to the point that houses sell for well below construction costs, and demand for city center office space is likewise weak; but a train that gets to Chicago in 2 hours and to New York in about 3:15 can make the area immediately around the station more desirable. In Dallas this is more complicated because it would be the system’s primary city, but a location with convenient rail access to Houston is likely to become more desirable for office space as well. This is not in competition with local transit – it complements it, by giving existing light rail lines and potential commuter rail lines a meatier city center to connect suburban areas with.

Public Transport and Scale

Noah asks what the proposal maps are for. The answer is, they are proposals for improvement in passenger rail. There is a real issue of scale and details, which is why those maps don’t depict literally every connection. For that, there are smaller-scale maps, in the same way there is the TransitMatters proposal for Regional Rail in the Boston area, or maps I’ve made for timed connections in New England and Upstate New York between intercity and regional trains. At lower-altitude zoom there’s also the issue of local connections to buses.

A roadmap like Google Maps or a national planning map, shown at such zoom that the entirety of a continental superstate like the United States is in the field of view, will only include the highest level of the transportation hierarchy. In the case of roads, that’s the Interstates, and the map may well omit spurs and loops. At lower altitude, more roads are visible, until eventually at city scale all streets are depicted.

The same is true of public transit – and high-speed rail is ideally planned as public transit at intercity scale. A continental-scale proposal will depict high-speed rail because it depicts all cities at once and therefore what matters at this level is how to get between regions. A state map or regional map such as for New England will depict all regional connections, and a local map will depict bus connections around each train station. At no point are these in competition for resources – good integrated planning means they all work together, so that improvements in regional rail also enable better bus connections, and improvements in intercity rail enable better regional connections.

Is all of this absolutely necessary? No. France manages to make certain connections work without it, and when I try to model this as a door-to-door trip, it’s a factor of 1.5-2 question, not an order of magnitude question. But a factor of 1.5 question is still serious, and it’s one that resolves itself with good public transit planning, rather than with not building high-speed rail at all.

Modeling High-Speed Rail for Germany

I’ve used a ridership model to construct a proposal for American high-speed rail – but what about the country I live in? There’s an election this year and one of the contested issues is climate change, and with growing passenger rail advocacy, it’s not outside the realm of possibility that there will be a large federal investment in dedicated high-speed lines (“NBS”). So I think it’s useful to model what German intercity rail will look like if there is greater investment in NBSes, culminating in a nationwide network such that ICEs will spend nearly all the time on NBSes or occasionally heavily upgraded legacy lines (“ABS”) rather than on slower lines.

If anything, I’m more optimistic about this network on the 15-year horizon than about American high-speed rail. Germany is slowly building more lines, like Stuttgart-Ulm, with Ulm-Munich, Frankfurt-Mannheim, Hanover-Bielefeld, and Frankfurt-Fulda on the horizon. People are also studying the prospects of a more expansive map as part of Deutschlandtakt additions, but unfortunately many 200 km/h ABSes are considered good enough even if they’re in easy terrain for a 300 km/h NBS, like Berlin-Halle/Leipzig.

The model

The professional way to model ridership is to split the travel zone, in this case the entire country, into very small pieces. I’m instead going to use an approximation with metropolitan areas and divisions thereof. For an illustration of my model’s level of sophistication, see below:

He definitely doesn’t wear a mask on the subway. Credit: Annette Pendlebury.

The gravity model to use is approximately,

\mbox{Ridership} = \mbox{Pop}_{A}^{0.8}\cdot\mbox{Pop}_{B}^{0.8}/\mbox{distance}^{2}

The justification for the exponent 2 in the gravity model is that the elasticity of ridership with respect to trip times appears to be close to -2. The justification for the exponent 0.8 is that it empirically appears true when considering Japanese cities’ Shinkansen ridership to Tokyo; the reason for this is that metropolitan areas comprise many different subsections, and the ones farther from city center have longer effective trip time counting connection time to the train station, and larger metropolitan areas tend to have longer distance from the center to the edge.

In the linked paper, the elasticity remains -2 even at short distances. However, we’re going to assume a minimum distance below which the elasticity vanishes, to avoid predicting infinite ridership as distance goes to zero. If distance is expressed in km, the best-fit constant is 75,000, with populations and annual ridership both in millions, and then if there’s no minimum distance, the model predicts Frankfurt (with 4 million people) to Mannheim (2.8 million, 75 km away) has 92 million annual riders just between the two regions, which is utter nonsense. In Japan, ridership looks like the floor is 500 km. In Germany, I’m going to round this to 2.5 hours, and because in practice it’s a bit more than 500 km, I’m going to round the constant 0.3/2.5^2 down to 1.8. We thus get,

\mbox{Ridership} = 1.8\cdot\mbox{Pop}_{A}^{0.8}\cdot\mbox{Pop}_{B}^{0.8}/\mbox{max}\{2.5, \mbox{time}\}^{2}

The network

This is the current draft of what I think Germany should build:

Blue = existing NBS and ABS lines, red = NBSes to be built

This isn’t too different from past maps I made. Berlin-Hanover is 60 minutes on this map and not 75 as on previous maps; a nonstop Velaro Novo can do it in 60 minutes, and the projected ridership is high enough that a half-hourly stopping train for service to Wolfsburg is viable in addition to a core express service. The branch point in the Rhine-Ruhr is moved to Dortmund, which slightly slows down service to Cologne and requires more tunnels, but improves frequency to the system massively, since Dortmund is a connection point to regional trains. Göttingen-Erfurt is dropped – all it does is connect Hanover and Hamburg with Erfurt, which is very small, and speed up travel to Nuremberg and Munich by 30 minutes, which is interesting but not enough to justify 100 km of high-speed rail.

Frankfurt still has an awkward-looking loop, whose purpose is to permit trains from Mannheim to enter the central tunnel to be constructed from the east and then run through to Cologne. However, this may not be necessary – trains from Cologne to Mannheim could just as well skip Frankfurt Hbf, serving Frankfurt at the airport or at a new station to be constructed at Frankfurt Süd, analogous to Cologne-Deutz for north-south through-trains. The expected traffic level is so high that the hit to Cologne-Frankfurt frequency is not awful, and the network complexity added by the skip isn’t higher than that added by having Frankfurt-Mannheim trains enter the tunnel from both directions depending on onward destination.

The network trip times are expressed in multiples of 15 minutes, with some places where timed connections are desirable, such as Fulda between Berlin-Frankfurt and Hamburg-Munich trains. However, overall, the traffic density predicted by the model is so high that on the stronger lines, like Cologne-Frankfurt, the timetable would not look like an integrated timed transfer system but rather the more continuous rapid transit-style model seen in Japan.

The power of polycentricity

The 0.8 exponent in the formula for ridership means that if we get to divide a single metropolitan area into subregions, then its ridership will increase. This is only justifiable if trains serve all such subregions; if the trains only serve some subregions, then we have to subtract them out. When we analyze New York or Tokyo, we can’t just add up each part of the metropolitan area separately – if we do so we must remove unserved sections like Long Island or Chiba, and the effect turns out to be similar to just lumping the metro area together.

However, in the Rhine-Ruhr, trains do serve nearly all sections of the region. The shape of the network there is such that intercity trains will continue stopping at Dortmund, Bochum, Essen, Duisburg, Wuppertal, Dusseldorf, and Cologne, at a minimum. The only recognizable centers without stops are Bonn and Mönchengladbach, and Bonn is connected to Cologne by streetcar.

Dividing cities and counties that are in the Rhine-Ruhr metropolitan region into the influence zones of the seven cities with stops based on what is the closest, we get Dortmund with 1.8 million, Bochum with 0.5, Essen 2, Duisburg 1, Wuppertal 0.9, Dusseldorf 2.3 (2 if we subtract out Mönchengladbach), and Cologne 2.9. Adding them up with exponents 0.8 is equivalent to considering a monocentric metropolitan core of 18.1 million; if we subtract out Mönchengladbach, it’s 17.6 million. This is enormous – larger than Paris and London, where only one high-speed rail stop is possible per train.

This also means we need to separately consider domestic and international traffic. Randstad is polycentric as well, and at a minimum there should be stops at Utrecht (1 million), Amsterdam (2.5), and Rotterdam (3.5), which means the region acts like a monocentric region of 9 million. The upshot is that if there were a 300 km/h train connecting Utrecht with Dusseldorf and Cologne with onward connections at both ends, and fares were st at domestic ICE rates and not Thalys rates, the connection between the two conurbations alone would generate about 17 million passengers a year. Of course, the model thinks all trip times up to 2.5 hours are equivalent, and the most distant city pair, Rotterdam-Dortmund, would be perhaps 1:45, but onward connections to German cities like Mannheim, Stuttgart, and Hanover are all 2:30 or longer with a 300 km/h Dutch line, and so there are benefits to constructing such a line over running at lower speed within the Netherlands.

To the extent the Frankfurt-Mannheim region can be thought of as a polycentric megaregion, the same is true there. Frankfurt, by which I mean Hesse-Darmstadt minus Bergstrasse, is 3.7 million people; Mainz is 0.6; the Rhine-Neckar (including Bergstrasse) is 2.4 million; Karlsruhe is 1.1 million; Stuttgart is 2.5 million. The model thinks that these regions combined generate 25 million annual trips to the Rhine-Ruhr.

European Urbanism and High-Speed Rail

Europe has a number of strong national high-speed rail networks, providing much inspiration internally as well as abroad, including in the United States. With Americans looking at an infrastructure bill including high-speed rail funding, there’s a lot of discussion about what can port, hence my proposal map. That said, caution is required when doing naive comparisons with Europe. European urbanism doesn’t work the same as American urbanism, in two ways. First, European cities are more compact and transit-oriented than most American cities, which is why I somewhat discount American lines unless at least one city connected has public transit. And second, Europe has more, smaller cities than the rest of the urbanized world. This post concerns the second issue.

French and American urbanism: an example

A few months ago I poked around European and East Asian metro area lists. The upshot is that whereas in the three East Asian democracies 70% of the population lives in metropolitan areas larger than 1 million, in France only 33% does, and the median resident sorted by metro area size lives in a metro region of 350,000.

We can apply the same analysis to the United States. At the CSA level, the median American lives in Sacramento, population 2.6 million, and 68% live in metro areas of at least 1 million; at the MSA level, the median is Milwaukee, population 1.6 million, and 56% live in metro areas of at least 1 million. American metropolitan areas are unusually weakly-centered, especially at the CSA level, but otherwise they’re pretty typical of the urbanized world; it’s Europe that’s unusual in having such small cities.

The upshot is that people who are not used to this peculiarity of Europe who look at a map of European cities focus on million-plus metro areas, which are not the whole story here, especially not in France. This makes Europe look emptier than it is, which can lead people to overrate how much ridership a high-speed rail network would have at a fixed population.

France and the Midwest

Scott Hand posted a map on Twitter superimposing France on the Midwest with Chicago taking the place of Paris, arguing that they are similar in population and area:

This is a good sanity check: your Midwestern network should be of comparable magnitude to the TGV network, rather than much larger. It’s easy to say, Lyon has 2.5 million people, Detroit has 5 million people, so clearly a line to Detroit is twice as good as one to Lyon, right? But no: French urbanism supplies many more small cities, which must be accounted for as well. At the end of the day, the populations are similar, even though, in addition to Chicago, the map has three cities (Detroit, St. Louis, Cleveland) with larger metro areas than Lyon and six more larger than Marseille (Milwaukee, Indianapolis, Nashville, Cincinnati, Columbus, Pittsburgh).

The LGV Sud-Est

It’s tempting to compare Paris-Lyon to Chicago-St. Louis. Yonah Freemark did this in 2009, and Jarrett Walker already pointed out in comments that the LGV Sud-Est was always about much more than this. On hindsight, I’ll add that even that sells the LGV Sud-Est short. High-speed rail between Paris and Lyon unlocked fast service from Paris to not just Lyon but also the following metro areas, all with 2016 populations:

  • Dijon (385,000), demoted from the PLM mainline to a branch but still served
  • Grenoble (688,000)
  • Saint-Etienne (520,000)
  • Chambéry (225,000)
  • Annecy (236,000)
  • Valence (187,000)
  • Vienne (115,000)
  • Bourg-en-Bresse (128,000), not on any direct train but still close enough by regional connection or car

What’s more, TGVs would branch from Part-Dieu along legacy lines to serve these smaller cities, albeit at low frequency. Now, with the LGV extending as far south as Marseille, Valence has a through-station on an LGV just outside the built-up area. There’s also Lyria service to thee major Swiss cities; Geneva, a metro area of 1 million, lies on a low-speed extension of the LGV Sud-Est, 3:11 from Paris.

Other than Geneva, which is invisible on the map because it is farther away, the other cities listed are all very small. In the United States, people don’t usually think of metropolitan areas of such size as urban, because they are extremely dispersed and socially identify as not-urban, and because metropolitan America operates at much larger size classes. But they have recognizable urban cores and their populations must be put into any ridership model trying to train data on TGV ridership. In fact, a gravity model with exponent 0.8 predicts that the combined TGV ridership from Paris to all the above cities, excluding Lyon, is nearly twice the ridership on Paris-Lyon.

And in this context, Chicago-St. Louis simply doesn’t compare. St. Louis is somewhat larger than Lyon, yes, but within 60 km, within which radius Lyon has independent Saint-Etienne, Vienne, Bourg, and Mâcon, St. Louis only has its own exurbs. To find a proper Midwestern comparison for the LGV Sud-Est and its extensions toward Marseille, one must go east of Chicago, toward Detroit and Cleveland. Within 60 km Detroit too only has its own CSA plus Windsor, but that CSA has 5 million people, and the same line also reaches Cleveland (CSA population 3.5 million), Toledo (900,000), and Pittsburgh (2.6 million) and points east.

What this means

Having fewer, larger cities doesn’t make it harder to build high-speed rail. On the contrary – it’s easier to serve such a geography. Asia lives off of such geography; Japan and Taiwan serve nearly their entire populations on just a single line, and Korea does on one mainline with a branch. An Asianized France would be able to serve nearly its entire population on the LGV network as-is without needing low-frequency branches to Chambéry- and Valence-scale cities, and an Asianized Germany would be able to just build an all-high-speed network and connect nearly everyone and not just half the population.

There are small cities that happen to lie on convenient corridors between larger cities, the way Valence is between Lyon and Marseille, or Augsburg and Ulm are between Stuttgart and Munich. Other small cities are close enough to large cities that they’re decently-served by a large city-focused rail network, like Saint-Etienne. Those cities are compact, so a large share of the population has access to the train – this is the explanation for the 0.8 exponent in the gravity model of ridership. But overall, most cities of that scale are strewn haphazardly around the country: examples include Limoges, Amiens, and Caen in France, and Osnabrück, Chemnitz, and Rostock here.

However, this doesn’t mean that, in analyzing the impact of population on ridership, we should just pretend the small cities don’t exist. They do, and they supply extra ridership that isn’t visible if one thinks city = metro area of 1 million or more. It’s an understandable way of thinking, but Europe has a lot of ridership generated from intermediate cities and from cities that have a regional rail connection to a big city or a less frequent direct intercity train, and the models have to account for it.

So yes, that the US has so many large-by-European-standards cities means high-speed rail would work well there. However, it equally means that a naive model that just says “this looks like the LGV Sud-Est” would underperform. A better model has to account for specific city pairs. American city pairs still look okay, even with extreme levels of sprawl at the outer ends, but ultimately this means the US can have a network of approximately the same scope of the LGV network, rather than one that is much denser.