Transit-Oriented Development and Rail Capacity

Hayden Clarkin, inspired by the ongoing YIMBYTown conference in New Haven, asks me about rail capacity on transit-oriented development, in a way that reminds me of Donald Shoup’s critique of trip generation tables from the 2000s, before he became an urbanist superstar. The prompt was,

Is it possible to measure or estimate the train capacity of a transit line? Ie: How do I find the capacity of the New Haven line based on daily train trips, etc? Trying to see how much housing can be built on existing rail lines without the need for adding more trains

To be clear, Hayden was not talking about the capacity of the line but about that of trains. So adding peak service beyond what exists and is programmed (with projects like Penn Station Access) is not part of the prompt. The answer is that,

  1. There isn’t really a single number (this is a trip generation question).
  2. Moreover, under the assumption of status quo service on commuter rail, development near stations would not be transit-oriented.

Trip generation refers to the formula connecting the expected car trips generated by new development. It, and its sibling parking generation, is used in transportation planning and zoning throughout the United States, to limit development based on what existing and planned highway capacity can carry. Shoup’s paper explains how the trip and parking generation formulas are fictional, fitting a linear curve between the size of new development and the induced number of car trips and parked cars out of extremely low correlations, sometimes with an R^2 of less than 0.1, in one case with a negative correlation between trip generation and development size.

I encourage urbanists and transportation advocates and analysts to read Shoup’s original paper. It’s this insight that led him to examine parking requirements in zoning codes more carefully, leading to his book The High Cost of Free Parking and then many years of advocacy for looser parking requirements.

I bring all of this up because Hayden is essentially asking a trip generation question but on trains, and the answer there cannot be any more definitive than for cars. It’s not really possible to control what proportion of residents of new housing in a suburb near a New York commuter rail stop will be taking the train. Under current commuter rail service, we should expect the overwhelming majority of new residents who work in Manhattan to take the train, and the overwhelming majority of new residents who work anywhere else to drive (essentially the only exception is short trips on commuter rail, for example people taking the train from suburbs past Stamford to Stamford; those are free from the point of view of train capacity). This is comparable mode choice to that in the trip and parking generation tables, driven by an assumption of no alternative to driving, which is correct in nearly all of the United States. However, figuring out the proportion of new residents who would be commuting to Manhattan and thus taking the train is a hard exercise, for all of the following reasons:

  • The great majority of suburbanites do not work in the city. For example, in the Western Connecticut and Greater Bridgeport Planning Regions, more or less coterminous with Fairfield County, 59.5% of residents work within one of these two regions, and only 7.4% work in Manhattan as of 2022 (and far fewer work in the Outer Boroughs – the highest number, in Queens, is 0.7%). This means that every new housing unit in the suburbs, even if it is guaranteed the occupant works in Manhattan, generates demand for more destinations within the suburb, such as retail and schools.
  • The decision of a city commuter to move to the suburbs is not driven by high city housing prices. The suburbs of New York are collectively more expensive to live in than the city, and usually the ones with good commuter rail service are more expensive than other suburbs. Rather, the decision is driven by preference for the suburbs. This means that it’s hard to control where the occupant of new suburban housing will work purely through TOD design characteristics such as proximity to the station, streets with sidewalks, or multifamily housing.
  • Among public transportation users, what time of day they go to work isn’t controllable. Most likely they’d commute at rush hour, because commuter rail is marginally usable off-peak, but it’s not guaranteed, and just figuring the proportion of new users who’d be working in Manhattan at rush hour is another complication.

All of the above factors also conspire to ensure that, under the status quo commuter rail service assumption, TOD in the suburbs is impossible except perhaps ones adjacent to the city. In a suburb like Westport, everyone is rich enough to afford one car per adult, and adding more housing near the station won’t lower prices by enough to change that. The quality of service for any trip other than a rush hour trip to Manhattan ranges from low to unusable, and so the new residents would be driving everywhere except their Manhattan job, even if they got housing in a multifamily building within walking distance of the train station.

This is a frustrating answer, so perhaps it’s better to ask what could be modified to ensure that TOD in the suburbs of New York became possible. For this, I believe two changes are required:

  • Improvements in commuter rail scheduling to appeal to the growing majority of off-peak commuters as well as to non-commute trips. I’ve written about this repeatedly as part of ETA but also the high-speed rail project for the Transit Costs Project.
  • Town center development near the train station to colocate local service functions there, including retail, a doctor’s office and similar services, a library, and a school, with the residential TOD located behind these functions.

The point of commercial and local service TOD is to concentrate destinations near the train station. This permits trip chaining by transit, where today it is only viable by car in those suburbs. This also encourages running more connecting bus service to the train station, initially on the strength of low-income retail workers who can’t afford a car, but then as bus-rail connections improve also for bus-rail commuters. The average income of a bus rider would remain well below that of a driver, but better service with timed connections to the train would mean the ridership would comprise a broader section of the working class rather than just the poor. Similarly, people who don’t drive on ideological or personal disability grounds could live in a certain degree of comfort in the residential TOD and walk, and this would improve service quality so that others who can drive but sometimes choose not to could live a similar lifestyle.

But even in this scenario of stronger TOD, it’s not really possible to control train capacity through zoning. We should expect this scenario to lead to much higher ridership without straining capacity, since capacity is determined by the peak and the above outline leads to a community with much higher off-peak rail usage for work and non-work trips, with a much lower share of its ridership occurring at rush hour (New York commuter rail is 67-69%, the SNCF part of the RER and Transilien are about 46%, due to frequency and TOD quality). But we still have no good way of controlling the modal choice, which is driven by personal decisions depending on local conditions of the suburb, and by office growth in the city versus in the suburbs.

High-Speed Rail Ridership Estimator Applet

Thijs Niks made a web applet for calculating high-speed rail network ridership estimates. This is based on the gravity model that I’ve used to construct estimates. The applet lets one add graph nodes representing metro areas and edges representing connections between them. It estimates ridership based on the model, construction costs based on a given choice of national construction costs, and overall profitability after interest. It can also automate the exact distances and populations, using estimates of population within a radius of 30 km from a point, and estimates of line length based on great circle length. The documentation can be found here and I encourage people to read it.

This is a very good way of visualizing certain things both about high-speed rail networks and the implications of a pure gravity model. For one, Metcalfe’s law is in full swing, to the point that adding to a network improves its finances through adding more city pairs than just the new edges. The German network overall is deemed to have insufficient financial rate of return due to the high costs of construction (and due to a limitation in the applet, which is that it assumes all links cost like high-speed rail, even upgraded classical lines like Berlin-Hamburg). But if the network is augmented with international connections to Austria, Czechia, Poland, Belgium, the Netherlands, France, and Switzerland, then it moves into the black.

To be clear, this is not a conclusion of the applet. Rather, the applet is a good visualization that this is a conclusion of the model. The model, with the following formula,

\mbox{Ridership} = c\cdot\mbox{Population}_{A}^{0.8}\cdot\mbox{Population}_{B}^{0.8}/(\max\{\mbox{distance}, d\})^{2}

is open to critique. The minimum distance d can be empirically derived from ridership along a line with intermediate stops; I use 500 km, or around a trip time of 2:15. The constant c is different in different geographies, and I don’t always have a good explanation for it. The TGV has a much higher constant than the Shinkansen (by a factor of 1.5), which can be explained by its much lower fares (a factor of about 1.7). But Taiwan HSR has a much higher constant than either, with no such obvious explanation. This is perilous, because Taiwan is a much smaller country than the others for which I’ve tested the model (Japan, South Korea, France, Germany, Spain, Italy). There may be reason to believe that at large scale, c should be lower for higher-population geographies, like the entirety of Europe; the reason is that if c is truly independent of population size, then the model implies that the propensity to travel per individual is not constant, but rather is larger in larger geographies, with an exponent of 0.6. This could to some extent be resolved if we have robust Chinese data – but China has other special elements that make a straight comparison uncertain, namely much lower incomes (reducing travel) and much higher average speeds (increasing travel).

The other issue is that the value of c used in the applet is much higher than the one I use. I use 75,000 for Shinkansen and 112,500 for Europe, with the populations of the metro areas stated in millions of people, the distance given in kilometers, and the ridership given in millions of riders per year. The applet uses 200,000, because its definition of metro area is not taken from national lists but from a flat applet giving the population in a 30 km radius from a point, which reduces Paris from 13 million people to 10.3 million people; it also omits many secondary cities in France that get direct TGVs to the capital, most notably Saint-Etienne and Valence, collectively dropping 12% of the modeled Paris-PACA ridership and 37% of the Paris-Rhône-Alpes ridership. (Conversely, the same method overestimates the size of metro Lille.)

Potentially, if the definition of a metro area is the population within a fixed radius, then the 0.8 exponent may need to be replaced with 1, since the fixed radius already drops many of the suburbs of the largest cities. The reason the gravity model has an exponent of 0.8 and not 1 is that larger metro areas have diseconomies of scale, as the distance from the average residence to the train station grows. Empirically, splitting combined statistical areas in the US into smaller metro areas and metropolitan divisions fits an exponent of 0.8 rather than 1, as some of those divisions (for example, Long Island) don’t have intercity train stations and have a longer trip time; it is fortunate that training the same model on Tokyo-to-secondary city Shinkansen ridership results in the same 0.8 exponent. However, if the definition of the metropolitan area is atypically unfair to New York and other megacities then the exponent is likely better converted to the theoretically simpler 1.

Timetable Padding Practices

Two weeks ago, the Wall Street Journal wrote this piece about our Northeast Corridor report. Much of it was based on a series of interviews William Boston did with me, explaining what the main needs on the corridor are. One element stands out since the MTA responded to what I was saying about schedule padding – I talk about how Amtrak and Metro-North both pad the timetables on the Northeast Corridor by about 25%, turning a technical travel time of an hour into 1:15 (best practices are 7%), and in response, the MTA said that they pad their schedules 10% and not 7%. This is an incorrect understanding of timetable padding, which speaks poorly to the competence of the schedule planners and managers at Metro-North.

The article says,

Aaron Donovan, a spokesman for the Metropolitan Transportation Authority, says the extra time built into Metro-North schedules generally averages 10%, depending on destination and length of trips, and takes into account routine track maintenance and capital work that can increase runtime. Metro-North continually reviews models, signal timing, equipment, and other elements of operation to improve travel times and reliability for customers, he says.

This is, to be clear, incorrect. Metro-North routinely recovers longer delays than 10%; delay recovery on the New Haven Line can reach well over 20 minutes out of a nominally two-hour trip, around 25% of the unpadded trip length. The reason this is incorrect isn’t that Donovan is dishonest or incompetent (he is neither of these two things), but almost certainly that the planners he spoke with genuinely believe they only pad 10%, because they, like all American railroaders, do not know how modern rail scheduling is done.

Modern rail scheduling practices in the higher-reliability parts of Europe and Japan start with the technical timetable, based on the actual speed zones and trains’ performance characteristics. This includes temporary speed restrictions. The ideal maintenance regime does not use them, instead relying on regular nighttime maintenance windows during which all tracks are out of service. However, temporary restrictions may exist if a line is taken out of service and trains are rerouted along a slower route, which is regrettably common in Germany. Modern signaling systems are capable of incorporating temporary speed restrictions – this is in fact a core requirement for American positive train control (PTC), since American maintenance practices rely on extensive temporary restrictions for work zones and one-off slowdowns. If the signal system knows the exact speed zones on each section of track, then so can the schedule planners.

The schedule contingency figure is computed relative to the best technical schedule. It is not computed relative to any assumption of additional delays due to dispatch holds or train congestion. The 7% figure used in Switzerland, Sweden, and the Netherlands takes care of the high levels of congestion on key urban segments.

The core urban networks in these countries stack favorably with Metro-North in track utilization. The Hirschengraben Tunnel in Zurich runs 18 S-Bahn trains per hour in each direction most of the day and 20 at rush hour with some extra S20 runs, and the Weinberg Tunnel runs 8 S-Bahn trains per hour and if I understand the network graphic right 7.5 additional intercities per hour. I urge people to go look at the graphic and try tracking down the lines just to see how extensively branched and reverse-branched they are; this is not a simple network, and delays would propagate. The reason the Swiss rail network is so punctual is that, unlike American rail planning, it integrates infrastructure and timetable development. This means many things, but what is relevant here is that it analyzes where delays originate and how they propagate, and focuses investments on these sections, grade-separating problematic flat junctions if possible and adding pocket tracks if not.

Were I to only take timetable padding into account relative to an already more tolerant schedule incorporating congestion and signaling limitations, I would cite much lower figures for timetable padding. Switzerland speaks of a uniform 7% pad, but in Sweden the figures include two components, a percentage (taking care of, among other things, suboptimal driver behavior) and a fixed number of minutes per 100 km, which at current intercity speeds resolve to 7% as in Switzerland. But relative to the technical trip time, the pad factors based on both observed timetable recovery and actual calculations on current speed zones are in the 20-30% range, and not 10%.

Of course, at no point do I suggest that Metro-North and Amtrak could achieve 7% right now, through just writing more aggressive timetables. To achieve Swiss, Dutch, and Swedish results, they would need Swiss, Dutch, or Swedish planning quality, which is sorely lacking at both railroads. They would need to write better timetables – not just more aggressive ones but also simpler ones: Metro-North’s 13 different stopping patterns on New Haven Line trains out of 16 main line peak trains per hour should be consolidated to 2. This is key to the plan – the only way Northern Europe makes anything work is with fairly rigid clockface timetables, so that one hour or half-hour is repeated all day, and conflicts can be localized to be at the same place every time.

Then they would need to invest based on reliability. Right now, the investment plans do not incorporate the timetable, and one generally forward-thinking planner found it odd that the NEC report included both high-level infrastructure proposals and proposed timetables to the minute. In the United States, that’s not the normal practice – high-level plans only discuss high-level issues, and scheduling is considered a low-level issue to be done only after the concrete is completed. In Northern European countries with competently-run railways and also in Germany, the integration of the timetable and infrastructure is so complete that draft network graphics indicating complete timetables of every train to the minute are included in the proposal phase, before funding is committed. In Switzerland, such a timetable is available before the associated infrastructure investments go to referendum.

Under current American planning, the priorities for Metro-North are in situ bridge replacements in Connecticut because their maintenance costs are high even by Metro-North’s already very expensive standards. But under good planning, the priority must be grade-separating Shell Interlocking (CP 216) just south of New Rochelle, currently a flat junction between trains bound for Grand Central and ones bound for Penn Station. The flat junctions to the branches in Connecticut need to be evaluated for grade-separation as well, and I believe the innermost, to the New Canaan Branch, needs to be grade-separated due to its high traffic while the ones to the two farther out branches can be kept flat.

None of this is free, but all of this is cheap by the standards of what the MTA is already spending on Penn Station Access for Metro-North. The rewards are substantial: 1:17 trip times from New Haven to Grand Central making off-peak express stops, down from 2 hours today. The big ask isn’t money – the entire point of the report is to figure out how to build high-speed rail on a tight budget. Rather, the big ask is changing the entire planning paradigm of intercity and commuter rail in the United States from reactive to proactive, from incremental to comfortable with groun-up redesigns, from stuck in the 1950s to ready for the transportation needs of the 21st century.

Reasons and Explanations

David Schleicher has a proposal for how Congress can speed up infrastructure construction and reduce costs for megaprojects. Writing about what further research needs to be done, he distinguishes reasons from explanations.

I have argued that many of the stories we tell about infrastructure costs involve explanations but not reasons. There are plenty of explanations for why projects cost so much, from too-deep train stations to out-of-control contractors, but they don’t help us understand why politicians often seem not to care about increasing costs. For that, we need to understand why there is insufficient political pressure to encourage politicians to do better.

I hope in this post to go over this distinction in more detail and suggest reasons. The key here is to look not just at costs per kilometer, but also costs per rider, or benefit-cost ratios in general. The American rail projects that are built tend to have very high benefits, to the point that at normal costs, their benefit-cost ratios would be so high that they’d raise the question of why it didn’t happen generations ago. (If New York’s construction costs had stayed the same as those of London and Paris in the 1930s, then Second Avenue Subway would have opened in the 1950s from Harlem to Lower Manhattan.) The upshot is that such projects have decent benefit-cost ratios even at very high costs, which leads to the opposite political pressure.

Those high benefit-cost ratios can be seen in low costs per rider, despite very high costs per kilometer. Second Avenue Subway Phase 1 cost $6 billion in today’s money and was projected to get 200,000 daily riders, which figure it came close to before the pandemic led to reductions in ridership. $30,000/rider is perfectly affordable in a developed country; Grand Paris Express, in 2024 prices, is estimated to cost 45 billion € and get 2 million daily riders, which at PPP conversion is if anything a little higher than for Second Avenue Subway. And the United States is wealthier than France.

I spoke to Michael Schabas in 2017 or 2018 about the Toronto rail electrification project, asking about its costs. He pointed out to me that when he was involved in the early 2010s studies for it, the costs were only mildly above European norms, but the benefits were so high that the benefit-cost ratio was estimated at 8. Such a project could only exist because Canada is even more of a laggard on passenger rail electrification than the United States – in Australia, Europe, Japan, or Latin America a system like GO Transit would have been electrified generations earlier, when the benefit-cost ratio would have been solid but not 8. The ratio of 8 seemed unbelievable, so Metrolinx included 100% contingency right from the start, and added scope instead of fighting it – the project was going to happen at a ratio of 2 or 8, and the extra costs bringing it down to 2 are someone else’s revenue.

The effect can look, on the surface, as one of inexperience: the US and Canada are inexperienced with projects like passenger rail electrification, and so they screw them up and costs go up, and surely they’ll go down with experience. But that’s not quite what’s happening. Costs are very high even for elements that are within the American (or Canadian) experience, such as subway and light rail lines, often built continuously in Canadian and Western US cities. Rather, what’s going on is that if a feature has been for any reason underrated (in this case, mainline rail electrification, due to technological conservatism), then by the time anyone bothers building it, its benefit-cost ratio at normal costs will be very high, creating pressure to add more costs to mollify interest groups that know they can make demands.

This effect even happens outside the English-speaking world, occasionally. Parisian construction costs for metro and RER tunnels are more or less the world median. Costs for light rail are high by French standards and low by Anglosphere ones. However, wheelchair accessibility is extremely expensive: Valérie Pécresse’s plan to retrofit the entire Métro with elevators, which are currently only installed on Line 14, is said to cost 15 to 20 billion euros. There are 300 stations excluding Line 14, so the cost per station, at 50-67 million € is even higher than in New York. In Madrid, a station is retrofit with four elevators for about 10 million €; in Berlin, they range between 2 and 6 million (with just one to two elevators needed; in Paris, three are needed); in London, a tranche of step-free access upgrades beginning in 2018 cost £200 million for 13 stations. This is not because France is somehow inexperienced in this – such projects happen in secondary cities at far lower costs. Moreover, when France is experimenting with cutting-edge technology, like automation of the Métro starting with Line 1, the costs are not at all high. Rather, what’s going on with accessibility costs is that Paris is so tardy with upgrading its system to be accessible that the benefits are enormous and there’s political pressure to spend a lot of money on it and not try saving much, not when only one line is accessible.

In theory, this reason should mean that once the projects with the highest benefit-cost ratios are built, the rest will have more cost control pressure. However, one shouldn’t be so optimistic. When a country or city starts out building expensive infrastructure, it gets used to building in a certain way, and costs stay high. Taiwanese MRT construction costs began high in the 1990s, and the result since then has not been cost control pressure as more marginal lines are built, but fewer lines built, and rather weak transit systems in the secondary cities.

Major reductions happen only in an environment of extreme political pressure. In Italy, the problem in the 1980s was extensive corruption, which was solved through mani pulite, a process that put half of parliament under indictment and destroyed all extant political parties, and reforms passed in its wake that increased transparency and professionalized project delivery. High costs by themselves do not guarantee such pressure – there is none in Taiwan or the United Kingdom. In the United States there is some pressure, in the sense that the thinktanks are aware of the problem and trying to solve it and there’s a decent degree of consensus across ideologies about how. But I don’t think there’s extreme political pressure – if anything the tendency for local activist groups is to work toward the same failed leadership that kept supervising higher costs, whereas mani pulite was a search-and-destroy operation.

Without such extreme pressure, what happens is that a very strong project like Caltrain or GO Transit electrification, the MBTA Green Line Extension, the Wilshire subway, or Second Avenue Subway is built, and then few to no similar things can be, because people got used to doing things a certain way. The project managers who made all the wrong decisions that let costs explode are hailed as heroes for finally completing the project and surmounting all of its problems, never mind that the problems were caused either by their own incompetence or that of predecessors who weren’t too different from them. The regulations are only tweaked or if anything tightened if a local political power broker feels not listened to. Countries and cities build to a certain benefit-cost ratio frontier, and accept the cost of doing business up to it; the result is just that fewer things are built in high cost per kilometer environments.

How One Bad Project Can Poison the Entire Mode

There are a few examples of rail projects that fail in a way that poisons the entire idea among decisionmakers. The failures can be total, to the point that the project isn’t built and nobody tries it again. Or the outcome can be a mixed blessing: an open project with some ridership, but not enough compared with the cost or hassle, with decisionmakers still choosing not to do this again. The primary cases I have in mind are Eurostar and Caltrain electrification, both mixed blessings, which poisoned international high-speed rail in Europe and rail electrification in the United States respectively. The frustrating thing about both projects is that their failures are not inherent to the mode, but rather come from bad project management and delivery, which nonetheless is taken as typical by subsequent planners, who benchmark proposals to those failed projects.

Eurostar: Flight Level Zero airline

The infrastructure built for Eurostar is not at all bad: the Channel Tunnel, and the extensions of the LGV Nord thereto and to Brussels. The UK-side high-speed line, High Speed 1, had very high construction costs (about $160 million/km in today’s prices), but it’s short enough that those costs don’t matter too much. The concept of connecting London and Paris by high-speed rail is solid, and those trains get a strong mode share, as do trains from both cities to Brussels.

Unfortunately, the operations are a mess. There’s security and border control theater, which is then used as an excuse to corral passengers into airline-style holding areas with only one or two boarding queues for a train of nearly 1,000 passengers. The extra time involved, 30 minutes at best and an hour at worst, creates a serious malus to ridership – the elasticity of ridership with respect to travel time in the literature I’ve seen ranges from -1 to -2, and at least in the studies I’ve read about local transit, time spent out of vehicle usually counts worse than time spent on a moving train (usually a factor of 2). It also holds up tracks, which is then used as an excuse not to run more service.

The excusemaking about service is then used to throttle the service offer, and raise prices. As I explain in this post, the average fare on domestic TGVs is 0.093€/passenger-km, whereas that on international TGV services (including Eurostar) is 0.17€/p-km, with the Eurostar services costing more than Lyria and TGV services into Germany. This includes both Eurostar to London and the services between Paris and Brussels, which used to be called Thalys, which have none of the security and border theater of London and yet charge very high fares, with low resulting ridership.

The origin of this is that Eurostar was conceived as a partnership between British and French elites, in management as well as the respective states. They thought of the Chunnel as a flashy project, fit for high-end service, designed for business travelers. SNCF management itself believes in airline-style services, with fares that profiteer off of riders; it can’t do it domestically due to public pressure to keep the TGV affordable to the broad public, but whenever it is freed from this pressure, it builds or recommends that others build what it thinks trains should be like, and the results are not good.

What rail advocates have learned from this saga is that cross-border rail should decenter high-speed rail. Their first association of cross-border high-speed rail is Eurostar, which is unreasonably expensive and low-ridership even without British border and security theater. Thus, the community has retreated from thinking in terms of infrastructure, and is trying to solve Eurostar’s problem (not enough service) even on lines where they need competitive trip times before anything else. Why fight for cross-border high-speed rail if the only extant examples are such underperformers?

This dovetails with the mentality that private companies do it better than the state, which is dominant at the EU level, as the eurocrats prefer not to have any visible EU state. This leads to ridiculous press releases by startups that lie to the public or to themselves that they’re about to launch new services, and consultant slop that treats rail services as if they are airlines with airline cost structures. Europe itself gave up on cross-border rail infrastructure – the EU is in panic mode on all issues, the states that would be building this infrastructure (like Belgium on Brussels-Antwerp) don’t care, and even bilateral government agreements don’t touch the issue, for example France and Germany are indifferent.

Caltrain: electrification at extreme costs

In the 2010s, Caltrain electrified its core route from San Francisco to Tamien just south of San Jose Diridon Station, a total length of 80 km, opening in 2024. This is the only significant electrification of a diesel service in the United States since Amtrak electrified the Northeast Corridor from New Haven to Boston in the late 1990s. The idea is excellent: a dense corridor like this with many stations would benefit greatly from all of the usual advantages of electrification, including less pollution, faster acceleration, and higher reliability.

Unfortunately, the costs of the project have been disproportionate to any other completed electrification program that I am aware of. The entire Caltrain Modernization Project cost $2.4 billion, comprising electrification, resignaling (cf. around $2 million/km in Denmark for ETCS Level 2), rolling stock, and some grade crossing work. Netting out the elements that are not direct electrification infrastructure, this is till well into the teens of millions per kilometer. Some British experiments have come close, but the RIA Electrification Challenge overall says that the cost on double track is in the $3.8-5.7 million/km range in today’s prices, and typical Continental European costs are somewhat lower.

The upshot is that Americans, never particularly curious about the world outside their border, have come to benchmark all electrification projects to Caltrain’s costs. Occasionally they glance at Canada, seeing Toronto’s expensive electrification project and confirming their belief that it is far too expensive. They barely look at British electrification projects, and never look at ones outside the English-speaking world. Thus, they take these costs as a given, rather than as a failure mode, due to poor design standards, poor project management, a one-off signaling system that had very high costs by American standards, and inflexible response to small changes.

And unfortunately, there was no pot of gold at the end of the Caltrain rainbow. Ridership is noticeably up since electric service opened, but is far below pre-corona levels, as the riders were largely tech workers and the tech industry went to work-from-home early and has still not quite returned to the office, especially not in the Bay Area. This one failure, partly due to unforeseen circumstances, partly due to poor management, has led to the poisoning of overhead wire electrification throughout the United States.

Did the Netherlands Ever Need 300 km/h Trains?

Dutch high-speed rail is the original case of premature commitment and lock-in. A decision was made in 1991 that the Netherlands needed 300 km/h high-speed rail, imitating the TGV, which at that point was a decade old, old enough that it was a clear success and new enough that people all over Europe wanted to imitate it to take advantage of this new technology. This decision then led to complications that caused costs to run over, reaching levels that still hold a European record, matched only by the almost entirely tunneled Florence-Bologna line. But setting aside the lock-in issue, is it good for such a line to run at 300 km/h?

The question can be asked in two ways. The first is, given current constraints on international rail travel, did it make sense to build HSL Zuid at 300 km/h?. It has an easy answer in the negative, due to the country’s small size, complications in Belgium, and high fares on international trains. The second and more interesting is, assuming that Belgium completes its high-speed rail network and that rail fares drop to those of domestic TGVs and ICEs, did it make sense to build HSL Zuid at 300 km/h?. The answer there is still negative, but the reasons are specific to the urban geography of Holland, and don’t generalize as well.

How HSL Zuid is used today

This is a schematic of Dutch lines branded as intercity. The color denotes speed and the thickness denotes frequency.

The most important observation about this system is that HSL Zuid is not the most frequent in the country. Frequency between Amsterdam and Utrecht is higher than between Amsterdam and Rotterdam; the frequency stays high well southeast of Utrecht, as far as Den Bosch and Eindhoven. The Dutch rail network is an everywhere-to-everywhere mesh with Utrecht as its central connection point, acting as the main interface between Holland in the west and the rest of the country in the east. HSL Zuid is in effect a bypass around Utrecht, faster but less busy than the routes that do serve Utrecht.

This needs to be compared with the other small Northern European country with a very strong legacy rail network, Switzerland. The map, with the same thickness and color scheme but not the same length scale, is here:

The orange segments are Alpine base tunnels, with extensive freight rail. The main high-cost investment in passenger-dedicated rail in Switzerland is not visible on the map, because it was built for 200 km/h to reduce costs: Olten-Bern, with extensive tunneling as well as state-of-the-art ETCS Level 2 signaling permitting 110 second headways. The Swiss rail network is centered not on one central point but on a Y between Zurich, Basel, and Bern, and the line was built as one of the three legs of the Y, designed to speed up Zurich-Bern and Basel-Bern trains to be just less than an hour each, to permit on-the-hour connections at all three stations.

By this comparison, HSL Zuid should not have been built this way. It is not useful for a domestic network designed around regular clockface frequencies and timed connections, in the Netherlands as much as in Switzerland. There is little interest in further 300 km/h domestic lines – any further proposals for increasing speeds on domestic trains are at 200 km/h, and as it is the domestic trains only go 160. In a country this size, so much time is spent on station approaches that the overall benefit of high top speed is reduced. Indeed, from Antwerp to Amsterdam, Eurostar trains take 1:20 over a distance of 182 km, an average speed comparable to the fastest classical lines in Europe such as the East Coast Main Line or the Southern Main Line in Sweden. The heavily-upgraded but still legacy Berlin-Hamburg line averages about 170 km/h when the trains are on time, and if German trains ran with Dutch punctuality and Dutch padding it would average 190 km/h.

What about international services?

The fastest trains using HSL Zuid are international, formerly branded Thalys, now branded Eurostar. They are also unfathomably expensive. Where NS’s website will sell me Amsterdam-Antwerp tickets for around 20€ if I’m willing to chain trips on slow regional trains, or 27€ on intercities doing the trip in 1:37 with one transfers, Eurostar charges 79-99€ for this trip when I look up available trains in mid-August on a weekday. For reference, the average domestic TGV trip over this distance is around 18€ and the average domestic ICE trip is around 22€. It goes without saying that the line is underused by international travelers – the fares are prohibitive.

Beyond Antwerp, the other problem is that Belgium has built HSL 1 from the French border to Brussels, HSL 2 and 3 from Leuven to the German border, and HSL 4 from Antwerp to the Dutch border, but has not bothered building a fast line between Brussels and Antwerp (or Leuven). The 46 km line between Brussels South and Antwerp Central takes 46 minutes by the fastest train. A 200 km/h high-speed line would do the trip in about 20, skipping Brussels Central and North as the Eurostars do.

But what if the fares were more reasonable and if Belgian trains weren’t this slow? Then, we would expect to see a massive increase in ridership, since the line would be connecting Amsterdam with Brussels in 1:40 and Paris in slightly more than three hours, with fares set at rates that would get the same ridership seen on domestic trains. The line would get much higher ridership.

And yet the trip time benefits of 300 km/h on HSL Zuid over 250 km/h are only 3 minutes. While much of the line is engineered for 300, the line is really two segments, one south of Rotterdam and one north of it, totaling 95 km of 300 km/h running plus extensive 200 and 250 km/h connections, and the total benefit to the higher top speed net of acceleration and deceleration time is only about 3 minutes. The total benefit of 300 km/h relative to 200 km/h is only about 8 minutes.

Two things are notable about this geography. The first is that the short spacing between must-serve stations – Amsterdam, Schiphol, Rotterdam, Antwerp – means that trains never get the chance to run fast. This is partly an artifact of Dutch density, but not entirely. England is as dense as the Netherlands and Belgium, but the plan for HS2 is to run nonstop trains between London and Birmingham, because between them there is nothing comparable in size or importance to Birmingham. North Rhine-Westphalia is about equally dense, and yet trains run nonstop between Cologne and Frankfurt, averaging around 170 km/h despite extensive German timetable padding and a slow approach to Cologne on the Hohenzollernbrücke.

The second is that the Netherlands is not a country of big central cities. Randstad is a very large metropolitan area, but it is really an agglomeration of the separate metro areas for Amsterdam, Rotterdam, the Hague, and Utrecht, each with its own set of destinations. The rail network needs to serve this geography with either direct trains or convenient transfers to all of the major centers. HSL Zuid does not do that – it has an easy transfer to the Hague at Rotterdam, but connecting to Utrecht (thus with the half of the Netherlands that isn’t Holland) is harder.

In retrospect, the Netherlands should have built more 200 and 250 km/h lines instead of building HSL Zuid. It could have kept the higher speed to Rotterdam but then built direct Rotterdam-Amsterdam and Rotterdam-Utrecht lines topping at 200 km/h, using the lower top speed to have more right-of-way flexibility to avoid tunnels. Separate trains would be running from points south to either Amsterdam or Utrecht, and fares in line with those of domestic trains would keep demand high enough that the frequency to both destinations would be acceptable.

In contrast, 300 km/h lines, if there are no slow segments like Brussels-Antwerp in their midst and if fares are reasonable, can be successful, if the single-core cities served are larger and the distances between stations are longer. The distances do not need to be as long as on some LGVs, with 400 km of nonstop running between Paris and Lyon – on a 100 km nonstop line, such as the plans for Hanover-Bielefeld including approaches or just the greenfield segment on Hamburg-Hanover, the difference between 200 and 300 km/h is 9 minutes, so about twice as much as on HSL Zuid and HSL 4 relative to their total length. This works, because while western Germany is dense much like the Netherlands, it is mostly a place of larger city cores separated by greater distances. The analogy to HSL Zuid elsewhere in Europe is as if Germany decided to build a line to 300 km/h standards internally to the Rhine-Ruhr region, or if the UK decided to build one between Liverpool and Manchester.

Can Ridership Surges Disrupt Small, Frequent Driverless Metros?

A recent discussion about the Nuremberg U-Bahn got me thinking about the issue of transfers from infrequent to frequent vehicles and how they can disrupt service. The issue is that driverless metros like Nuremberg’s rely on very high frequency on relatively small vehicles in order to maintain adequate capacity; Nuremberg has the lowest U-Bahn construction costs in Germany, and Italian cities with even smaller vehicles use the combination of short stations and very high frequency to reduce costs even further. However, all of this assumes that passengers arrive at the station evenly; an uneven surge could in theory overwhelm the system. The topic of the forum discussion was precisely this, but it left me unconvinced that such surges could be real on a driverless urban metro (as opposed to a landside airport people mover). The upshot is that there should not be obstacles to pushing the Nuremberg U-Bahn and other driverless metros to their limit on frequency and capacity, which at this point means 85-second headways as on the driverless Parisian lines.

What is the issue with infrequent-to-frequent transfers?

Whenever there is a transfer from a large, infrequent vehicle to a small, frequent one, passengers overwhelm systems that are designed around a continuous arrival rate rather than surges. Real-world examples include all of the following:

  • Transfers from the New Jersey Transit commuter trains at the Newark Airport station to the AirTrain.
  • Transfers from OuiGo TGVs at Marne-la-Vallée to the RER.
  • In 2009, transfers from intercity CR trains at Shanghai Station to the metro.

In the last two cases, the system that is being overwhelmed is not the trains themselves, which are very long. Rather, what’s being overwhelmed is the ticket vending machines: in Shanghai the TVMs frequently broke, and with only one of three machines at the station entrance in operation, there was a 20-minute queue. A similar queue was observed at Marne-la-Vallée. Locals have reusable farecards, but non-locals would not, overwhelming the TVM.

In the first case, I think the vehicles themselves are somewhat overwhelmed on the first train that the commuter train connects to, but that is not the primary system capacity issue either. Rather, the queues at the faregates between the two systems can get long (a few minutes, never 20 minutes).

In contrast, I have never seen the transfer from the TGV to the Métro break the system at Gare de Lyon. The TGV may be unloading 1,000 passengers at once, but it takes longer for all of them to disembark than the headway between Métro trains; I’ve observed the last stragglers take 10 minutes to clear a TGV Duplex in Paris, and between that, long walking paths from the train to the Métro platforms, and multiple entrances, the TGV cannot meaningfully be a surge. Nor have I seen an airplane overwhelm a frequent train, for essentially the same reason.

What about school trips?

The forum discussion brings up two surges that limit the capacity of the Nuremberg U-Bahn: the airport, and school trips. The airport can be directly dispensed with – individual planes don’t do this at airside people movers, and don’t even do this at low-capacity landside people movers like the JFK and Newark AirTrains. But school trips are a more intriguing possibility.

What is true is that school trips routinely overwhelm buses. Students quickly learn to take the last bus that lets them make school on time: this is the morning and they don’t want to be there, so they optimize for how to stay in bed for just a little longer. Large directional commuter volumes can therefore lead to surges on buses: in Vancouver, UBC-bound buses routinely have passups in the morning rush hour, because classes start at coordinated times and everyone times themselves to the last bus that reaches campus on time.

However, the UBC passups come from a combination of factors, none of which is relevant to Nuremberg:

  • They’re on buses. SkyTrain handles surges just fine.
  • UBC is a large university campus tucked at the edge of the built-up area.
  • UBC has modular courses, as is common at American universities, and coordinated class start and end times (on the hour three days a week, every 1.5 hours two days a week).

It is notable that Vancouver does not have any serious surges coming from school trips, even with trainsets that are shorter than those of Nuremberg (40 meters on the Canada Line and 68-80 meters on the Expo and Millennium lines, compared with 76 meters). Schools are usually sited to draw students from multiple directions, and are usually not large enough to drive much train crowding on their own. A list on Wikipedia has the number of students per Gymnasium, and they’re typically high three figures with one at 1,167, none of which is enough to overwhelm a driverless 76 meter long train. Notably, school trips do not overwhelm the New York City Subway; New York City Subway rolling stock ranges from 150 to 180 meters long rather than 76 as in Nuremberg, but then the specialized high schools go as far up as 5,800 students, and one has 3,000 and is awkwardly located in the North Bronx.

Indeed, neither Vancouver nor New York schedules its trains based on whether school is in session. Both run additional buses on school days to avoid school surges, but SkyTrain and the subway do not run additional vehicles, and in both formal planning and informal railfan lore about crowding, school trips are not considered important. So school surges are absolutely real on buses, and university surges are real everywhere, but not enough to overwhelm trains. Nuremberg should not consider itself special on this regard, and can plan its U-Bahn systems as if it does not have special surges and passengers do arrive continuously at stations.

The Hamburg-Hanover High-Speed Line

A new high-speed line (NBS) between Hamburg and Hanover has received the approval of the government, and will go up for a Bundestag vote shortly. The line has been proposed and planned in various forms since the 1990s, the older Y-Trasse plan including a branch to Bremen in a Y formation, but the current project omits Bremen. The idea of building this line is good and long overdue, but unfortunately everything about it, including the cost, the desired speed, and the main public concerns, betray incompetence, of the kind that gave up on building any infrastructure and is entirely reactive, much like in the United States.

The route, in some of the flattest land in Germany, is a largely straight new high-speed rail line. Going north from Hanover to Hamburg, it departs somewhat south of Celle, and rejoins the line just outside Hamburg’s city limits in Meckelfeld. The route appears to be 107 km of new mainline route, not including other connections adding a few kilometers, chiefly from Celle to the north. An interactive map can be found here; the map below is static, from Wikipedia, and the selected route is the pink one.

For about 110 km in easy topography, the projected cost is 6.7 billion € per a presentation from two weeks ago, which is about twice as high as the average cost of tunnel-free German NBSes so far. It is nearly as high as the cost of the Stuttgart-Ulm NBS, which is 51% in tunnel.

And despite the very high cost, the standards are rather low. The top speed is intended to be 250 km/h, not 300 km/h. The travel time savings is only 20 minutes: trip times are to be reduced from 79 minutes today to 59 minutes. Using a top speed of 250 km/h, the current capabilities of ICE 3/Velaro trains, and the existing top speeds of the approaches to Hamburg and Hanover, I’ve found that the nonstop trip time should be 46 minutes, which means the planned timetable padding is 28%. Timetable padding in Germany is so extensive that trains today could do Hamburg-Hanover in 63 minutes.

As a result, the project isn’t really sold as a Hamburg-Hanover high-speed line. Instead, the presentation above speaks of great trip time benefits to the intermediate towns with local stops, Soltau (population: 22,000) and Bergen (population: 17,000). More importantly, it talks about capacity, as the Hamburg-Hanover line is one of the busiest in Germany.

As a capacity reliever, a high-speed line is a sound decision, but then why is it scheduled with such lax timetabling? It’s not about fitting into a Takt with hourly trip times, first of all because if the top speed were 300 km/h and the padding were the 7% of Switzerland, the Netherlands, and Sweden then the trip time would be 45 minutes, and second of all because Hamburg is at the extremity of the country and therefore it’s not meaningfully an intercity knot that must be reached on the hour.

Worse, the line is built with the possibility of freight service. Normal service is designed to be passenger-only, but in case of disruptions on the classical line, the line is designed to be freight-ready. This is stupid: it’s much cheaper to invest in reliability than to build a dual-use high-speed passenger and freight line, and the one country in the world with both a solid high-speed rail network and high freight rail usage, China, doesn’t do this. (Italy builds its high-speed lines with freight-friendly standards and has high construction costs, even though its construction costs in general, e.g. for metro lines or electrification, are rather low.)

Cross-Border Rail and the EU’s Learned Helplessness

I’m sitting on a EuroCity train from Copenhagen back to Germany. It’s timetabled to take 4:45 to do 520 km, an average speed of 110 km/h, and the train departed 25 minutes late because the crew needed to arrive on another train and that train was late. One of the cars on this train is closed due to an air conditioner malfunction; Cid and I rode this same line to Copenhagen two years ago and this also happened in one direction then.

This is a line that touches, at both ends, two of the fastest conventional lines in Europe, Stockholm-Malmö taking 4:30 to do 614 km (136 km/h) and Berlin-Hamburg normally taking 1:45 to do 287 km (164 km/h) when it is on time. This contrast between good lines within European member states, despite real problems with the German and Swedish rail networks, and much worse ones between them, got me thinking about cross-border rail more. Now, this line in particular is getting upgraded – the route is about to be cut off when the Fehmarn Belt Line opens in four years, reducing the trip time to 2:30. But more in general, cross-border and near-border lines that slow down travel that’s otherwise decent on the core within-state city pairs are common, and so far there’s no EU action on this. Instead, EU action on cross-border rail shows learned helplessness of avoiding the only solution for rail construction: top-down state-directed infrastructure building.

The upshot is that there is good cross-border rail advocacy here, most notably by Jon Worth, but because EU integration on this matter is unthinkable, this advocacy is forced to treat the railroads as if they are private oligopolies rather than state-owned public services. Jon successfully pushed for the incoming EU Commission of last year to include passenger rights in its agenda, to deal with friction between different national railroads. The issue is that SNCF and DB have internal ways of handling passenger rights in case of delays, due to domestic pressure on the state not to let the state railroad exploit its users, and they are not compatible across borders: SNCF is on time enough not to strand passengers, DB has enough frequency and extreme late-night timetable padding (my connecting train to Berlin is padded from 1:45 to 2:30, getting me home well past midnight) not to strand passengers; but when passengers cross from Germany to France, these two internal methods both fail.

At no point in this discussion was any top-down EU-level coordination even on the table. The mentality is that construction of new lines doesn’t matter – it’s a megaproject and these only generate headaches and cost overruns, not results, so instead everything boils down to private companies competing on the same lines. That the companies are state-owned is immaterial at the EU level – SNCF has no social mission outside the borders of France and therefore in its international service it usually behaves as a predatory monopoly profiteering off of a deliberately throttled Eurostar/Thalys market.

If there’s no EU state action, then the relationship between the operator, which is not part of the state, and the passenger, is necessarily adversarial. This is where the preference for regulations that assume this relationship must be adversarial and aim to empower the individual consumer comes from. It’s logical, if one assumes that there will never be an EU-wide high-speed intercity rail network, just a bunch of national networks with one-off cross-border megaprojects compromised to the point of not running particularly quickly or frequently.

And that is, frankly, learned helplessness on the part of the EU institutions. They take it for granted that state-led development is impossible at higher level than member states, and try to cope by optimizing for a union of member states whose infrastructure systems don’t quite cohere. Meanwhile, at the other end of the Eurasian continent, a continental-scale state has built a high-speed rail network that at this point has higher ridership per capita than most European states and is not far behind France or Germany, designed around a single state-owned network optimized for very high average speeds.

This occurs at a time when support for the EU is high in the remaining member states. There’s broad understanding that scale is a core benefit of the union, hence the regulatory harmonization ensuring that products can be shipped union-wide without cross-border friction. But for personal travel by train, these principles go away, and friction is assumed to only comprise the least important elements, because the EU institutions have decided that solving the most important ones, that is speed and frequency, is unthinkable.

Meme Weeding: Embodied Carbon

The greenhouse gases emitted by the production of concrete, called embodied carbon, are occasionally used as a green-NIMBY argument against building new things. A Berlin Green spokesperson coauthored a study opposing U-Bahn construction on the grounds that the concrete used in construction would raise emissions. More recently, I’ve seen American opponents of transit-oriented development in Manhattan, of all places, talk about the high embodied carbon of new high-rise buildings. Katja Diehl calls for a moratorium on new buildings on anti-concrete grounds, and a petition for the EU to shift regulations to be against new buildings and in favor of reuse on embodied carbon grounds got written up favorably by Kate Wagner in the Nation. Against all of this, I’ve found some numbers on the actual emissions involved in concrete production for new buildings, and they are so low as to be insignificant, 1.5 orders of magnitude less than transportation emissions. A decarbonization strategy should largely ignore embodied carbon concerns and embrace pro-growth sentiments: big buildings, big subway systems, big cities.

What is embodied carbon?

Embodied (sometimes called embedded) carbon is the carbon content emitted by the production of materials. The production of concrete emits greenhouses gases, mainly through two mechanisms: the chemical process used to produce cement emits CO2 by itself, and the energy used for production adds to the emissions of the electric grid.

What are the embodied carbon emissions of new buildings?

The embodied carbon content of concrete depends heavily on the local electricity grid as well as on the required strength of the material, with stronger requirements leading to higher emissions. The Climate Group commissioned a report on this in the British context, finding a wide range, but the average is around 250 kg of CO2-equivalent per m^3 of concrete, the 75th percentile is about 300, and the upper bound is 450. This is a cradle-to-gate figure, taking into account the existing conditions of the carbon intensivity of where concrete is produced and of the logistics system for getting it to the construction site. This is already with some reductions from a previous baseline (EC100; the UK average is around EC60), and further reductions are possible, through decarbonizing the logistics and production; the goal of the report is not to bury the concept of embodied carbon as I do but to propose ways to reduce construction industry emissions.

The question is now how to convert cubic meters of concrete into square meters of built-up area. I have not seen European figures for this, but I did find a 2012 report by the Building and Construction Authority. In Singapore, the sustainability index used is the concrete usage index (CUI), measured in meters (cubic meters per square meter). The example projects given in the study, all around 15 years old, have a CUI of 0.4-0.5 m, and it was pointed out to me on social media that in Toronto the average is 0.55 m.

250 kg/m^3 times 0.4 m equals 0.1 t-CO2 per m^2 of built-up area. A 100 m^2 apartment thus has an embodied carbon content of around 10 t-CO2. This is relative to a baseline in which there is already some concern for reducing construction emissions, both the CUI and the carbon content of concrete per m^3, but this is largely without techniques like mass timber or infra-lightweight concrete (ILC). In Singapore the techniques highlighted in the BCA report are fully compatible with the city’s high-rise character, and the example building with gold but not platinum certification has 25 stories.

Should we worry about construction emissions?

No.

An aggressively YIMBY construction schedule, say with 10 dwellings built annually per 1,000 people, say averaging 100 m^2, emits around 0.1 t/capita annually: 0.1 t/m^2 * 100 m^2/unit * 0.01 unit/capita. All figures have ranges (and if anything, 100 is high for the places that build this much urban infill housing), but factor-of-1.5 ranges don’t erase an order of magnitude analysis. The emissions produced by construction, even if it were raised to some of the highest per capita rates found in the developed world – in fact higher rates than any national average I know of – would be about two orders of magnitude lower than present-day first-world emissions. They’d be 1.5 orders of magnitude lower than transportation emissions; in Germany, transport is 22% of national emissions and rising, as all other sources are in decline whereas transport is flat.

There’s a lot of confusion about this because some studies talk about buildings in general providing a high share of emissions. The Bloomberg-era PlaNYC spoke of buildings as the top source of emissions in New York, and likewise the Nation cites WeForum saying buildings are 37% of global emissions, citing a UN report that includes buildings’ operating emissions (its topline figure is 10 Gt in operating emissions, which is 27% of global emissions in 2022). But the construction emissions are insignificantly low. This means that aggressive replacement of older buildings by newer, more energy-efficient ones is an unmixed blessing, exactly the opposite of the conclusion of the green movement.

Instead of worrying about a source of emissions measured per capita in the tens of kilograms per year rather than in the tons, environmental advocates need to prioritize the most important source of greenhouse gases. The largest in developed countries is transportation, with electricity production usually coming second, always falling over the years while transport remains flat. In cold countries, heating is a significant source of emissions as well, to be reduced through building large, energy-efficient apartment buildings and through heat pump installation.

Regulations on new construction’s embodied carbon are likely a net negative for the environment. The most significant social policy concerning housing as far as environmental impact is concerned is to encourage people to live in urban apartment buildings near train stations. Any regulation that makes this harder – for example, making demolitions of small buildings to make room for big ones harder, or demanding that new buildings meet embodied carbon standards – makes this goal harder. This can be understandable occasionally if the goal of the regulation is not environmental, for example labor regulations for construction workers. It is not understandable if the goal is environmental, as the concern over embodied carbon is. People are entitled to their opinion that small is beautiful as a matter of aesthetic judgment, but they are not entitled to alternative facts that small is environmentally friendly.