Category: Urban Transit

Missed Connections

I’ve discussed before the topic of missed connections on subway systems, both here and on City Metric. I’ve for the most part taken it for granted that on a rapid transit network, it’s important to ensure that whenever two lines intersect, they offer a transfer. This seems like common sense. The point of this post is not to argue for this principle, but to distinguish two different kinds of missed connections: city center misses, and outlying misses. Both are bad; if I had to say which is worse I’d say it’s the city center miss, but city center misses and outlying misses are bad for distinct reasons.

A useful principle is that every pair of rapid transit lines should intersect, unless one is a shuttle, or both are circumferential. If the city is so large that it has multiple circular lines at different radii (Beijing has two, and London vaguely has two as well depending on how one counts the Overground), then they shouldn’t intersect, but rapid transit networks should be radial, and every radial line should connect to every other line, with all radial-radial transfers ideally located within the center. City center misses weaken the network by making some radials not connect, or perhaps connect at an inconvenient spot. Outlying misses often permit more central transfers, and their problem is that they make it harder to transfer to the better or less crowded radial on the way to the center. London supplies a wealth of examples of the latter without the former.

What counts as a missed connection?

Fundamentally, the following picture is a missed subway connection:

The red and blue lines intersect without a transfer. Even if a few stations later there is a transfer, this is a miss. In contrast, the following picture is not a missed connection:

It might be faster for riders to transfer between the southern and western leg if there were a station at the exact physical intersection point, but as long as the next station on the red line has a transfer to the blue line it counts, even if the blue line has one (or more) stations in the middle. Washington supplies an example of this non-miss: it frustrates riders that there’s no connection between Farragut West and Farragut North, but at the next station south from the intersection on the Red Line, Metro Center, there is a transfer to the Blue and Orange Lines. London supplies another pair of examples: the Northern line and the Waterloo and City line appear to intersect the District line without a transfer, but their next station north from the physical intersection point, Bank, has an in-system transfer to Monument on the District.

There are still a few judgment calls in this system. One is what to do at the end of the line. In this case, I rule it a missed connection if the terminal clearly has an intersection without a transfer; if the terminal is roughly between the two stations on the through-line, it doesn’t count. Another is what to do about two lines that intersect twice in close succession, such as the Bakerloo and Hammersmith and City lines in London, and Metro Lines 4 and 10 in Paris. In such cases, I rule that, if there’s just one station on the wrong side (Paddington on Bakerloo, Mabillon on M10) then I rule it a single intersection and allow transfers at the next station over, by which standard London has a missed connection (Edgware Road has no Bakerloo/H&C transfer) and Paris doesn’t (Odeon has an M4/M10 transfer).

How many missed connections are there?

In Paris, there are three missed connections on the Metro: M9/M12, M5/M14, M9/M14. As I discuss on City Metric, it’s no coincidence that two of these misses involve Line 14, which has wide stop spacing. Narrow stop spacing makes it easier to connect within line-dense city centers, and Paris famously has the densest stop spacing of any major metro system. M9/M12 and M9/M14 morally should connect at Saint-Augustin and Saint-Lazare, but in fact there is no in-system transfer. M5/M14 should connect at Gare de Lyon, but when M5 was built it was not possible to get the line to the station underground and then have it cross the Seine above-ground, so instead it meets M1 at Bastille, while M14 doesn’t serve since it expresses from Gare de Lyon to Chatelet. A fourth missed connection is under construction: the extension of M14 to the north misses M2 at Rome, prioritizing long stop spacing over the connection to the M2/M6 circumferential.

In Tokyo, there are many misses. I am not sure why this is, but judging by line layout, Tokyo Metro and Toei try to stick to major roads whenever possible, to avoid tunneling under private property, and this constrains the ability of newer lines to hit station locations on older lines. If I understand this map correctly, there are 19 missed connections: Ginza/Hibiya (Toranomon and Kasumigaseki should connect), Ginza/Mita, Ginza/Yurakucho, Ginza/Shinjuku, Marunouchi/Mita (Ginza and Hibiya should connect), Marunouchi/Yurakucho, Asakusa/Yurakucho, Asakusa/Hanzomon, Hibiya/Namboku, Hibiya/Yurakucho (Tsukiji and Shintomicho should connect), Hibiya/Hanzomon, Hibiya/Shinjuku, Hibiya/Oedo, Tozai/Oedo, Tozai/Fukutoshin, Mita/Oedo, Chiyoda/Oedo twice, and Oedo/Fukutoshin. Oedo is particularly notable for being a circumferential line that misses a large number of transfers.

In New York, there are even more misses. Here the culprit is clear: the two older layers of the subway, the IRT and BMT, have just two missed connections. One, 3/L at Junius Street and Livonia Avenue, is an outlying miss. The other is central: Bowling Green on the 4-5 and Whitehall on the R-W should connect. But the newer layer, the IND, was built to drive the IRT and BMT into bankruptcy through competition rather than to complement them, and has a brutal number of misses: ABCD/2-3, ACE/1-2-3, AC-F/2-3-4-5, AC-G/2-3-4-5-BQ-DNR, BD/NQRW, BDFM/NQRW, BD/JZ, E/1, E/F, M/NW, R/7, F/BD-NQ, F/NRW, F-Q/4-5-6, F/NW, G/7, G/JMZ. Counting individual track pairs, this is 46 misses, for a total of 48 including the two IRT/BMT misses; I’m excluding local-only transfers, such as Columbus Circle and 53rd/Lex, and counting the 42nd Street Shuttle as an express version of the 7, so it doesn’t miss the BDFM transfer.

Finally, London only has eight misses. In Central London there are three: the Metropolitan or Hammersmith and City line misses the Bakerloo line as discussed above, and also the Victoria line and Charing Cross branch of the Northern line at Euston. The other five are outlying: the Central line misses the Hammersmith and City line at Wood Lane/White City, and its branches miss the Piccadilly line’s Uxbridge branch three times; the fifth miss is Metropolitan/Bakerloo. But one more miss is under construction: the Battersea extension of the Northern line is going to intersect the Victoria line without a transfer.

The difference between the two kinds of miss

Many misses are located just a few stations away from a transfer. In New York, some misses are just a station away from a transfer, including the G/7 miss in Long Island City, the E/1 miss between 50th Street and 59th Street, and several more are a few stations away, such as the various BDFM/NQRW misses. In London, these include two of the three Central London transfers: there is an H&C/Bakerloo transfer at Baker Street and an H&C-Met/Victoria transfer at King’s Cross-St. Pancras. As a result, not counting the Waterloo and City line, only two trunk lines in the system do not have any transfer: the Charing Cross branch of the Northern line and the Metropolitan/H&C line.

On a radial network, if two lines don’t have any transfer, then the network is degraded, since passengers can’t easily connect. In New York, this is a huge problem: some station pairs even within the inner networks require two transfers, or even three counting a cross-platform local/express transfer. My interest in subway networks and how they function came about when I lived in Morningside Heights on the 1 and tried socializing with bloggers in Williamsburg near the JMZ.

In Paris the three misses are also a problem. Line 4 is the only with a transfer to every other main line. Line 9 intersects every other line, and Line 14 will when its northern extension opens, but both miss connections, requiring some passengers to take three-seat rides, in a city infamous for its labyrinthine transfer stations. Fundamentally, the problem is that the Paris Metro is less radial than it should be: some lines are laid out as grid routes, including Lines 3, 5, and 10; moreover, Lines 8, 9, 12, and 13 are radial but oriented toward a different center from Lines 1, 4, 7, and 11.

In London, in contrast, there is almost no pair of stations that require a three-seat ride. The Charing Cross branch of the Northern line doesn’t make any stop that passengers from the H&C or Met line can’t get to from another line with one interchange (Goodge Street is walking distance to Warren Street). A bigger problem is the lack of interchange to the Central line on the west, which makes the connections between the H&C stations on the west and some Central line stations awkward, but it’s still only a small number of stations on each line. So the problem in London is not network robustness.

Rather, the problem in London is severe capacity limitations on some lines. Without good outlying interchanges, passengers who want to get between two lines need to ride all the way to the center. Most likely, passengers between the Piccadilly and Central line branches to the west end up driving, as car ownership in West London is relatively high. Passengers without a car have to instead overload the Central line trunk.

The same problem applies to misses that are strictly speaking not missed connections because the two lines do not actually intersect. In Paris, this occurs on Line 7, which swings by the Opera but doesn’t go far enough west to meet Lines 12 and 13. In London, the best example is Hammersmith station: the H&C and District lines have separate stations without an interchange, but they do not intersect since it’s the terminus of the H&C line and therefore I don’t count it as a miss. But morally it’s an outlying miss, preventing District line riders from changing to the H&C line to reach key destinations like Euston, King’s Cross, and Moorgate without overloading the Victoria or Northern line.

In New York this problem is much less acute. The only outlying misses are the 3/L and the ABCD/2-3; the 3/L connects two very low-ridership tails, so the only serious miss is on the Upper West Side. There, passengers originating in Harlem can walk to either line, since the two trunks are two long blocks apart, and passengers originating in Washington Heights can transfer from the A-C to the 1 at 168th Street; at the other end, passengers bound for Midtown can transfer at Columbus Circle, using the underfull 1 rather than the overcrowded 2 and 3.

The role of circumferential lines

Outlying transfers are useful in distributing passengers better to avoid capacity crunches, but they are incidental. They occur when formerly competing suburban lines get shoehorned into the same subway network, or when two straight roads intersect, as in Queens. But the task of distributing passengers between radial lines remains important and requires good connections between as many pairs of radials as possible.

The usual solution to this is a circumferential line. In Moscow, there are several missed connections in the center (Lines 3/6, 3/7, 6/9) and one more planned (8/9), but the Circle Line helps tie in nearly all the radii together, with just one missed connection (to Line 10 to the north) and one more under construction (to Line 8 to the west). The point of the Circle Line is to allow riders to connect between two outlying legs without congesting the center. This is especially important in the context of Moscow, where there are only a handful of interchange stations in the center, most of which connect more than two lines.

In London, the Overground is supposed to play this role. However, the connections between the Underground and Overground are weak. From Highbury and Islington clockwise, the Overground misses connections to the Central line, the Victoria line, the main line of the District line, the Piccadilly line, the Hammersmith and City line, both branches of the Northern line, and the Piccadilly line (it also misses the Metropolitan line, but that’s on a four-track stretch where it is express and local service is provided by the Jubilee line, with which there is a transfer). Much of this is an unforced error, since the Underground lines are often above-ground this far out, and stations could be moved to be better located for transfers.

In New York, the only circumferential line is the G train, which has uniquely bad transfers, legacy of the IND’s unwillingness to build a system working together with the older subways. Triboro RX (in the original version, not the more recent version) would play this role better: with very little tunneling, it could connect to every subway line going counterclockwise from the R in Bay Ridge to the B-D and 4 at Yankee Stadium. On the way, it would connect to some major intermediate centers, including Brooklyn College and Jackson Heights, but the point is not just to connect to these destinations in the circumferential direction but also to facilitate transfers between different lines.

Going forward, cities with large metro network should aim to construct transfers where feasible. In New York there are perennial proposals to connect the 3 and L trains; these should be implemented. In London, the missed outlying transfers involve above-ground stations, which can be moved. The most important miss, White City/Wood Lane, is already indicated as an interchange on the map, but does not to my understanding have an in-system transfer; this should be fixed.

Moreover, it is especially important to have transfers from the radial lines to the circumferential ones. These improve network connectivity by allowing passengers to change direction (from radial to circumferential, e.g. from east-west to north-south within Queens), but also help passengers avoid congested city centers like outlying radial-radial transfers. Where circumferential lines don’t exist, they should be constructed, including Triboro in New York and Line 15 in Paris; where they do, it’s important to ensure they don’t miss connections the way the Overground does.

The MTA Genius Challenge is as Bad as Expected

A year ago, Governor Andrew Cuomo declared a competitive $2.5 million grant, to be disbursed by what he dubbed the Genius Challenge. I wrote about it at the time, expressing skepticism that it would lead to anything useful. The panel of eight judges had only one person with background in the transportation industry, a former FRA administrator. The word “genius” itself is a tech mainstay that to me mostly means “I don’t know any Fields Medalists.” And the topics within the scope of the grant seemed more about what the tech industry thought were the most pressing issues and not what the lowest-hanging actually were. I had very low expectations, and the announcement of the winning entries met them.

The grant has three topics: signaling, rolling stock (interpreted broadly), and underground mobile or wireless service. The last three is by far the least important; it also got only half a million dollars, whereas each of the other two got a full million. Each of the two main ideas shows how weak the very concept of the genius grant is, but they do so in dramatically different ways.

The rolling stock winners included a vendor asking for a grant for New York to use its rolling stock (CRRC); the problems with that idea are more akin to those of the signaling section, so I will cover them there. A second rolling stock winner was a proposal to use better data collection to facilitate preventive maintenance; this idea may or may not work, it’s hard to tell from layers of obfuscating business language. It’s the third idea that deserves the most attention, and the most scorn: lengthening trains but not platforms.

The crank Idea: lengthening trains

The genius competition gave a $330,000 grant for the idea of lengthening trains from 10 to 14 cars without lengthening the platforms. Trains would alternate between only berthing the first 10 cars and only berthing the last 10. Transit Twitter has already dumped on this idea, and for good reason: the proposal reads like a crank paper purporting to prove the Riemann hypothesis or another famous result, starting with a lot of trivial observations and then making a leap of logic buried somewhere in the middle.

The basic problem with running trains that are longer than the platforms is that passengers need to be able to move to the correct car, which takes time. The report says that this is done on the London Underground, which is true, but only at outlying stations – as is the case on the subway in New York. The conductor announcement “only the first five cars will open” is familiar to anyone riding the 3 train and was familiar to anyone riding the 1 train before the new South Ferry station opened. This is fine as long as the station in question is low-volume enough that the extra dwell time does not interfere with operations. Lengthening trains beyond the platforms at busier stations than Harlem-148th Street 145th Street or South Ferry would result in a shuffle forcing passengers to scramble within the train (if moving between cars is possible) or on the platform (if it isn’t). The dwell times would be brutal and would almost certainly reduce capacity measured in passengers per hour.

The proposal handwaves this critical flaw by saying that dwell times would decrease because crowding would decrease. This assumes that dwell times are a function exclusively of on-train crowding, rather than of the number of passengers getting on or off the train. The same number of passengers would have the same platform space, but would actually only be able to use a fraction of it: many would only be able to use the 6 cars that go to their chosen destination, and at those cars, the volume of passengers per unit of platform length would rise.

The second handwave is unlimited stations, with longer platforms. Acknowledging that the busiest stations should have all doors open, the proposal says,

[P. 20] Third, 18.5% of rides occur through just 10 stations in Manhattan. In the medium term, the platforms can, and should, be extended at these 10 stations to enable customers that embark and disembark at them to use any car at both ends of their trip. Accordingly, 9.25% of the customers that presently need to use the middle cars could instead use the end cars.

This is the equivalent of the logical leap from trivial to wrong in a crank paper. First, the number of central stations that would need to be lengthened is much more than 10, including some key origins (86th/Lex, Jamaica Center, etc.) and transfer points (West 4th, Canal, 96th/Broadway). And second and more importantly, the busiest stations are multilevel complexes, where just adding more pedestrian circulation is hard; London is spending a considerable amount of money on that at Bank. Lengthening platforms at these stations is prohibitively expensive. This problem is discussed in cities with constrained underground platforms in the CBD, such as Vancouver, where nearly all Expo Line stations are above-ground (thus, relatively easy to lengthen), but the most crowded in Downtown Vancouver are in a tunnel, where platform reconstruction costs too much to be economic.

The bigger question is why the judges did not catch the error. The proposal brings up London as an example, which serves to bring the magic of the foreign to people who are unfamiliar with best industry practices. Saying that New York does the same is equally true, and in a way more relevant to the proposal (since New York doesn’t let people move between cars, making this more challenging than in London), but would raise questions like “can the dwell times of relatively light stations like South Ferry or Harlem-145th be replicated at the top 40 stations?”. London is Anglophone and some reformist New Yorkers have used it as a source of foreign ideas the way they wouldn’t use non-Anglophone cities. But the judges didn’t do the basic due diligence of checking whether London really implements the idea as widely as the proposal implies, and if not, then why not.

The rent-seeking idea: CBTC by another name

New York State awarded four applicants $250,000 each for ideas about signaling. All four ideas boiled down to the same thing: introducing new technology for communication between trains permitting the functional equivalent of moving-block signaling, at a lower cost than preexisting communication-based train control (CBTC) installations.

The grantees all have experience in the transportation industry. Rail signaling vendors Thales and Ansaldo propose to use cameras to read automated signals; train sensor provider Metrom Rail and veteran rail manager and consultant Robert James propose ultra-wide broadband to improve train location precision. There’s nothing obviously wrong about their proposals. Nor is there anything outlandish, which is why each of the two technologies has two independent applicants behind it. Thales and Ansaldo in particular have experience in advanced signaling – Thales supplied CBTC to the L 7 train in New York and to Metro Line 13 in Paris, and Ansaldo supplied rail automation to Copenhagen and CBTC to a number of Paris Metro lines.

Even then, questions about cost remain. Robert James’ and Metrom’s proposals leave a bad taste in my mouth for their cost estimates. James has a systemwide cost estimate somewhat less than $200 million, not much more than $500,000 per km; Metrom says its system costs “$3 million per mile” and compares itself positively with legacy CBTC systems at $20 million per mile. Actual costs of CBTC without automation in Paris on Line 13 were about 5 million euros per km according to Wikipedia, and this includes modification of the railyards and not just the signaling system. So the Metrom system’s claimed figure is still cheaper, but not by quite as much. Metrom also complains that in Boston, CBTC would not improve capacity much because it would prohibit double-berthing, an issue that is only relevant to a subway-surface system and not to a full metro.

The broader problem with this part of the grant is that if the MTA put out an RFP about CBTC on the subway, it would get bids from Ansaldo, Thales, and Metrom, and James might well bid or consult for a bidder. It would be able to judge the technical merit of each proposal in much closer detail than given in the competition. Instead, the state is paying vendors to market their technology to the public, which would influence future procurement.

While the grant asks about whether the technology is proprietary, it makes no attempt at establishing a multi-vendor standards. Such standards exist: Thales and Ansaldo are both listed as ERTMS vendors. In France there’s already a discussion in the trade press about whether using ERTMS is better than using CBTC; the discussion specifically mentions New York’s uniqueness as a network with connected rather than isolated lines, and says CBTC is designed for isolated lines whereas ERTMS is designed for shared lines, such as the RER system. European experts might well recommend that New York use ERTMS for the subway, even though it’s a system originally designed for mainline rail.

New York’s highly-branched system means it must be more conservative with new technology – there’s nowhere to test it, now that the L and 7 already have CBTC. The shuttles might be useful test cases, or the 1 and 6 trains on weekdays, but without isolated lines, the cost of a mistake in procurement or technological failure is much higher. This suggests the MTA should try to reduce the complexity of branching (which is what I would’ve proposed if it had been within the grant’s scope), and until then concentrate on imitating proven technology rather than innovating. This is especially important given the potential for rent-seeking, in which the vendors use the grant to market themselves to the state over competitors selling similar product.

The judges don’t know any better

Would a panel of judges with more familiarity with metro operations around the first world have come to better decisions? Probably. Through blogs, railfan forums, and comments, I know people with great knowledge of existing operations in a number of cities in the first world, and for the most part they think highly enough of their local systems that they’d ask of any innovation, “why hasn’t it been implemented here already?”.

I wrote in 2011 that people in the US who make technical arguments in favor of public transit tend to be skeptical of many proposals, to the point of finding existing US agencies incompetent. This is US-specific: London Reconnections is a technical blog but it tends to support Transport for London’s process, Swiss and Japanese railfans seem to trust their local rail operators, and even Transport Paris is more positive about STIF’s capital investment than New York-based blogs are about the MTA’s. Experts (and not just bloggers like me) could point out innovations their cities have that can be imported into New York, as well as shoot down bad ideas for which “why doesn’t London/Paris/Tokyo do it?” is a useful sanity check.

Note that sometimes there is a legitimate reason to do something that nobody else has tried. New York’s highly branched network makes ERTMS a better deal there than on other metro systems, and an RFI would be prudent. But because the details of implementation matter more than the idea of innovative genius, it has to go through the regular procurement process.

Cuomo attempted to inject the inventions of the American tech industry into the subway. Instead, he created space for cranks to promulgate their ideas and for vendors to have a leg up over their competitors in any future bid. In effect, his attempt to improve the economic productivity of the public sector to be more in line with that of the American tech industry is going to make the public sector less productive, through weaker institutions (namely, a less robust CBTC bid) and distraction (namely, the useless train lengthening idea).

Transit and Scale Variance Part 3: Grids

This is my third post about scale variance in transit planning; see parts 1 and 2. In part 1, I discussed how good bus networks exist at a certain scale, which can’t easily be replicated at larger scale (where the slowness of city buses makes them less useful). In part 2, I went over a subway planning feature, especially common in the communist bloc, that again works only at a specific scale, namely cities with enough population for 3-4 subway lines; it gets more complex in larger cities, and cannot be imported to bus networks with 3-4 lines. In this post, I will focus on one scale-variant feature of surface transit: the grid.

The grid works only for surface transit and not for rapid transit, and only at a specific scale, so constrained as to never be maximally useful in an entire city, only in a section of a city. This contrasts with what Jarrett Walker claims about grids. Per Jarrett, grids are the perfect form of a transit network and are for the most part scale-invariant (except in very small networks). One of the impetuses for this post is to push back against this: grids are the most useful at the scale of part of a transit city.

Grid Networks Versus Radial Networks

I’ve written a few posts exhorting subway planners to build their networks in a certain way, which, in the most perfect form, is radial. In particular, tangential subway lines, such as the G train in New York (especially when it ran to Forest Hills), Line 10 in Paris, and Lines 3 and 6 in Shanghai, are weak. When the G train was running to Forest Hills, most local passengers would switch from it to the next Manhattan-bound train, leading New York City Transit to send more Manhattan-bound local subways to Forest Hills and eventually to cut back the G to Long Island City. Based on these examples, I contend that on a subway network, every line should be either radial, serving the CBD, or circumferential, going around the CBD.

My post about New York light rail proposes a network with some lines that are neither: in the Bronx, my proposal is essentially a grid, with north-south routes (Grand Concourse, Webster, 3rd) and east-west ones (161st, Tremont, Fordham) and one that combines both (145th-Southern). Regular commenter NewtonMARunner criticized me for this on Twitter. I answered that the lines in my proposal are based on the busiest buses in the Bronx, but this simply shifts the locus of the question to the existing network: if transit lines should be radial or circumferential, then why are the tangential Bx19 bus (145th-Southern) or the Bx40/42 and Bx36 (Tremont, with a long radial eastern tail) so successful?

To answer this requires thinking more carefully about the role of circumferential routes, which by definition don’t serve the most intensely-used nodes. In Paris, Lines 2 and 6 form a ring that misses five out of six train stations and passes just outside the CBD, and yet they are both busy lines, ranking fourth and fifth in ridership per km. The reason is that they are useful for connecting to radial Metro lines and to some RER lines (namely, the RER A and the southern half of the RER B). Tangential lines miss connections much more easily: in the west, Line 10 here has a decent transfer to Line 9 and a somewhat decent one to Line 8, but to Lines 12 and 13 it’s already not very direct. The G train in New York has the same problem to the south – few connections to lines that actually do go into Manhattan.

Consider the following three possible networks:

The radial network is a typical subway network. The full grid lets you go from everywhere to everywhere with just one transfer, at the cost of having far more route length than the radial network. The partial grid no longer lets you go from everywhere to everywhere easily, and has the outer two lines in each service direction missing city center, but still has more overall route-length than the radial network. The principle here is that a grid plan is useful only if the grid can be complete.

The scale, then, is that rapid transit is so expensive that there’s no money for a complete grid, making a radial plan more appropriate. But surface transit, especially by bus, can be spread across a grid more readily. The Bronx’s size, density, and bus ridership patterns are such that a mostly complete grid is feasible within the western two-thirds of the borough, supplemented by the subway. In this environment, a tangential route is fine because it hits all the radial routes it could, and could provide useful two-seat rides to a large variety of destinations.

Are Grids Really Grids?

Chicago has a relentless bus grid. The three busiest north-south routes are the tangential 8 (Halsted), 9 (Ashland), and 49 (Western), which are 22, 29, and 26 km long respectively. None enters the Loop; Halsted, the easternmost, is at the closest approach 800 meters from the Loop, across a freeway. The two busiest east-west routes, the 77 (Belmont) and 79 (79th), are also far from the Loop.

However, I contend that these routes don’t really form a grid, at least not in the sense that passengers ride between two arbitrary points in Chicago by riding a north-south bus and connecting to an east-west bus. Instead, their outer ends form tails, which people ride to the L, while their inner ends are standard circumferentials, linking two L branches. The L in turn is purely radial and doesn’t follow the Chicago grid, with the Blue Line’s O’Hare Branch, the Orange Line, and the Brown Line all running diagonally.

Vancouver is similar. The north-south routes are radial, veering to enter Downtown. The east-west ones are more circumferential than tangential: they connect the Expo and Canada Lines, and most also connect to UBC. The Broadway buses (9 and 99) pass so close to Downtown Vancouver they’re more tangential, but they also offer the shortest path between the Expo and Canada Lines (making them a strong circumferential) while at the same time serving high job density on Central Broadway (giving them some characteristics of a radial).

In the absence of a radial rail network to connect to, long grid routes are less useful. Cities have a center and a periphery, and the center will always get more ridership, especially transit ridership. The outermost grid routes are often so weak that they should be pruned, but then they weaken the lines they connect to, making it necessary to prune even more lines until the grid is broken.

The Optimal Scale for a Grid

A strong transit grid will not form in a city too small for it. There needs to be a large enough center with enough demand for transit ridership to justify more than a purely radial bus network with a timed transfer. At the same time, the city cannot be too big, or else the arterial buses are too slow to be useful for ordinary work and leisure trips, as in Los Angeles.

What’s more, there is no Goldilocks zone, just right for a grid. Chicago is already too big for a bus grid without the radial rail layer. It’s also too big for what Jarrett calls grid accelerators – that is, rapid transit routes that replace bus grid lines: the Red Line is plausibly a grid accelerator, but the other lines in Chicago are not, and if there were L lines only at grid points, then the Red Line and the one east-west route would get overcrowded heading toward the Loop. Even Vancouver, a compact metro area hemmed by mountains and the ocean, relies on the diagonal Expo Line to serve Downtown and doesn’t really have a grid beyond city limits. A less dense city in the same land area could have a grid, but without much traffic or a strong CBD, cars would always beat transit on time and only the poor would ride the bus.

The scale in which grid networks work more or less on their own seems to be that of Vancouver proper, or that of the Bronx. Vancouver is 115 km^2 and the Bronx is 110 km^2; Vancouver’s bus grid spills over to Metrotown and the Bronx’s to Upper Manhattan, but in both cases these are small increases in the relevant land area.

Tellingly, Vancouver still relies on the bus network to feed SkyTrain; the Canada Line is a grid accelerator, but the Expo Line is not. The Bronx is the more interesting case, because it is not a city or even the center of the city, but rather a dense outlying portion of the city with an internal arterial grid. In both cases, the grid is supplementary to the radial rail core, even if the routes that use it have a lot of independent utility (Metro Vancouver has higher bus ridership than rail ridership, and the Bronx buses combined have slightly more ridership than the combined number of boardings on the Bronx subway stations).

Geographical constraints matter as well. The Bronx and Upper Manhattan are hemmed by water and by the administrative border of the city (which also includes a sharp density gradient), and Vancouver is hemmed by water and by a density gradient in the east. This makes it easier to equip both with grids that are close enough to the complete grid in the middle image above rather than the incomplete one in the third image. The Bronx’s lower-density eastern tails happen to meet up with those of Queens, forming circumferential routes, and also have enough north-south subway lines to feed that they remain useful.

In a transit city, the grid cannot come first. Even if there is a street grid, the spine of the network has to be radial as soon as there is demand for more than two rapid transit lines. The role of surface transit remains feeding rapid transit. Grids look attractive, but the optimal scale for them is awkward: large-scale surface transit grids are too slow, forcing the city to have a rapid transit backbone, and if the city is too small for that then the arterial grid provides too good auto access for public transit to be useful.

The RPA Construction Cost Report

A much-awaited Regional Plan Association report about construction costs in New York has come out, as a supplement to the Fourth Regional Plan, and I’m unimpressed. I thought that I would either enjoy reading the RPA’s analysis, or else be disappointed by it. Instead, I’ve found myself feeling tepid toward most of the analysis; my objections to the report are that its numbers have serious mistakes, that the recommendations at the end conflict with the analysis, and that it seems to overvalue other English-speaking countries, even when their construction costs are the highest in the world outside the US.

The big contrast is with Brian Rosenthal’s expose in the New York Times. The main comparison city to New York there is Paris, where the extension of Metro Line 14 resembles New York’s subway extensions; for the article, Brian talked to construction managers here, and either visited the site himself or talked to people who did, to compare the situation with that of New York. As a result, I learned things from Brian’s article that I did not know before (namely, that the excavation per station for the Line 14 extension wasn’t less voluminous than for Second Avenue Subway). The RPA report gives a few details I wasn’t familiar with, such as escalators’ share of construction costs, but nothing that seems big.

I feel like I slag on the RPA a lot nowadays – it started with their report from three years ago about Outer Borough transit and continued with their wrong approach to Triboro, but more recently I didn’t think much of their take on suburban TOD, or the Gateway project, or the Fourth Regional Plan in general. This isn’t out of malice or jealousy; when I talked to Tom Wright six months ago I sympathized with the political constraints he was operating under. The problem is that sometimes these constraints lead either to unforced errors, or to errors that, while I understand where they come from, are big enough that the organization should have pushed and made sure to avoid them. In the case of the construction cost report, the errors start small, but compound to produce recommendations that are at times counterproductive; agency officials reading this would have no way of reducing costs.

Mistakes in the Numbers

The RPA is comparing New York’s costs unfavorably with those of other cities around the world, as well as one American city (Los Angeles). However, at several points, the numbers appear different from the ones I have seen in the news media. Three places come to mind – the first is a nitpick, the second is more serious but still doesn’t change the conclusions, the third is the most egregious in its implications.

The first place is right at the beginning of the report. In the executive summary, on page 2, the RPA gives its first example of high New York costs:

The Second Avenue Subway (SAS), for example, has the distinction of being the world’s most expensive subway extension at a cost of $807 million per track mile for construction costs alone. This is over 650% more per mile than London’s Northern Line extension to Battersea — estimated at $124 million per track mile.

Both sets of numbers are incorrect – in fact, contradicted by the rest of the document. SAS is $1.7 billion per route-km, which is $850 million per track-km. The Northern line extension to Battersea is also much more expensive. I can’t tell whether these figures are missing something, such as stations or overheads, but as headline numbers, they’re both lowballed.

The second place is when the report discusses station construction costs. Not having seen any advance copy, I wrote about this issue two weeks ago, just before the report came out: the three new SAS stations cost $821, $649, and $802 million, according to the Capital Program Dashboard. In contrast, on pp. 16-17, the RPA gives lower figures for these stations: just $386 million, $244 million, and $322 million. The RPA’s source is “Capital Construction Committee reports,” but my post on station costs looked at some of those and found costs that are not much lower than those reported in the Dashboard. The RPA figures for the last two stations, 86th and 72nd, seem close to the costs of finishes alone, and it’s possible that the organization made a mistake and confused the cost of just finishes (or perhaps just excavation) with the total cost, combining both excavation and finishes.

With the correct costs, the difference from what Paris spends on a station (about $110 million on average) seems so stark that the recommendations must center station construction specifically, and yet they don’t.

The third and most problematic mistake is table 10 on page 50, which lists a number of subway projects and their costs. The list is pretty short, with just 11 items, of which 3 are in New York, another is in Los Angeles, one is in Toronto, and 2 are in London. The Toronto project, the Spadina subway extension to Vaughan, and one of the London projects, the Northern line extension, are both lowballed. The RPA says that the Northern line extension’s cost is $1.065 billion, but the most recent number I’ve seen is £1.2 billion, which in PPP terms is $1.7 billion. And the Vaughan extension, listed as $1.961 billion in the report, is now up to C$3.2 billion, about $2.55 billion in PPP terms. Perhaps the RPA used old numbers, before cost escalations, but in such a crucial report it’s important to update cost estimates even late in the process.

But most worryingly, the costs on table 10 also include mistakes in the other direction, in Paris and Tokyo. The cost estimate listed for Line 14 South in Grand Paris Express is $4.39 billion. But the Cour des Comptes’ report attacking Grand Paris Express’s cost overruns lists the line’s cost as only €2.678 billion, or about $3.3 billion; this is in 2012 euros, but French inflation rates are very low, well below 1% a year, and at any rate, even applying American inflation rates wouldn’t get the cost anywhere near $4 billion. In Tokyo, the RPA similarly inflates the cost of the Fukutoshin Line: it gives it as $3.578 billion, but a media report after opening says the cost was ¥250 billion, or about $2.5 billion in today’s PPP conversion, with even less inflation than in France.

I can understand why there would be downward mistakes. Reports like this take a long time to produce, and then they take even longer to revise even after they are supposedly closed to further edits; I am working on a regional rail report for TransitMatters that has been in this situation for three months, with last-minute changes, reviews by stakeholders, and printing delays. However, the upward mistakes in Paris and Tokyo are puzzling. It’s hard to explain why, since the RPA’s numbers are unsourced; it’s possible they heard them from experts, but didn’t bother to write down who those experts were or to check their numbers.

The Synthesis Doesn’t Follow the Analysis

Manuel Melis Maynar’s writeup in Tunnelbuilder about how as CEO of Madrid Metro he delivered subway construction for, in today’s money, around $60 million per km, includes a number of recommendations. The RPA report cites his writeup on several occasions, as well as his appearance at the Irish Parliament. It also cites secondary sources about Madrid’s low construction costs, which appear to rely on Melis’s analysis or at least come to the same conclusions independently. However, the RPA’s set of recommendations seems to ignore Melis’s advice entirely.

The most glaring example of this is design-build. Melis is adamant that transit agencies separate design from construction. His explanation is psychological: there are always some changes that need to be made during construction (one New York-based construction manager, cited on p. 38 of the RPA study, says “there is no 100% design”), and contractors that were involved in the design are more likely to be wedded to their original plans and less flexible about making little changes. This recommendation of Melis’s is absent from the report, and on the contrary, the list of final recommendations includes expansion of design-build, a popular technique among reformers in New York and in a number of English-speaking cities.

Another example is procurement. I have heard the same explanation for high New York costs several times since I first brought up the issue in comments on Second Avenue Sagas: the bidding process in New York picks the lowest-cost proposal regardless of technical merit (Madrid, in contrasts, scores proposals 50% on technical merit, 30% on cost, and 20% on speed), and to avoid being screwed by dishonest contractors, the state writes byzantine, overexacting specs. As a result, nobody wants to do business with public works in New York, which means that in practice very few companies bid, leading to one-bid contracts. Brian’s article in the New York Times goes into how contractors have an MTA premium since doing business with the MTA is so difficult, and there’s also less competition, so they charge monopoly rates.

The RPA report’s analysis mentions this (pp. 3-4):

In addition, the MTA’s practice of selecting the lowest qualified bidder, even though they are permitted to issue Requests-for-Proposals, has resulted in excessive rebidding and the selection of teams that cannot deliver, resulting in millions of dollars in emergency repairs.

However, the list of recommendations at the end does not include any change to procurement practices to consider technical merit. The recommendations include post-project review for future construction, faster environmental review, reforms to labor rules, and value capture, but nothing about reforming the procurement process to consider technical merit.

Finally, the report talks about the problem of change orders repeatedly, on pp. 3, 15-16, and 38-39, blaming the proliferation of change orders for part of the cost escalation on SAS. Melis addresses this question in his writeup, saying that contracts should not be awarded for a lump sum but rather be itemized, so that change orders come with pre-agreed costs per item. None of this made it to the final recommendations.

There’s a World Outside the Anglosphere

If the report’s recommendations are not based on its own analysis, or on correct construction cost figures, then what are they based on? It seems that, like all failed reform ideas around the US, the RPA is shopping for ideas from other American cities or at least English-speaking ones that look good. Its recommendations include “adopt London’s project delivery model” and “expand project insurance and liability models,” the latter of which is sourced to the UK and Australia. Only one recommendation so much as mentions a non-English-speaking city: “develop lessons learned and best-practice guidance as part of a post-project review” mentions Madrid in passing, but focuses on Denver and Los Angeles.

This relates to the pattern of mistakes in the cost figures. Were the numbers on table 10 right, the implication would be that London, Paris, and Tokyo all have similar construction costs, at $330, $350, and $400 million per km, and Toronto is cheaper, at $230 million per km. In this situation, London would offer valuable lessons. Unfortunately, the RPA’s numbers are wrong. Using correct numbers, London’s costs rise to $550 million per km, while those of Paris and Tokyo fall to $260 and $280 million. Toronto’s costs rise to $300 million per km, which would be reasonable for an infill subway in a dense area (like the Fukutoshin Line and to some extent the Metro Line 14 extension), but are an outrage for a suburban extension to partly-undeveloped areas.

Using correct numbers, the RPA should have known to talk to people in countries that don’t speak English. Many of the planners and engineers in those countries speak English well as a second language. Many don’t, but New York is a large cosmopolitan city with immigrants with the required language skills, especially Spanish.

Nonetheless, the RPA report, which I am told cost $250,000 to produce, does not talk to experts in non-English-speaking countries. The citations of Melis are the same two English-language ones I have been citing for years now; there is no engagement with his writings on the subject in Spanish or his more recent English-language work (there’s a paper he coauthored in 2015 that I can’t manage to get past the paywall update: kind souls with academic access sent me a copy and it’s not as useful as I’d hoped from the abstract), nor does the RPA seem to have talked to managers in Madrid (or Barcelona) today. Across more than 200 footnotes, 30-something are sourced to “expert interviews,” and of those all but a handful are interviews with New York-based experts and the rest are interviews with London-based ones.

As a result, while the report is equipped to explain New York’s internal problems, it fails as a comparative piece. The recommendations themselves are primarily internal, based on things Americans have been discussing among themselves for years: streamlining environmental review, simplifying labor rules, expanding design-build.

The labor reforms mentioned include exactly one specific case of excessive staffing, reported in the New York Times (and, beforehand, on an off-hand remark by then-MTA Capital Construction chief Michael Horodniceanu), about the number of workers it takes to staff a tunnel-boring machine. The New York Times article goes into more detail about the entire process, but the RPA report ignores that in favor of the one comparison that had been going around Transit Twitter for years. Instead of proposing specifics for reducing headcounts, the report talks about changing the way workers are paid for each day, relying on internal reforms proposed by people dissatisfied with the unions rather than on any external analysis.

The Cycle of Failure

I’ve been reading policy papers for maybe a decade – mostly American, a few Israeli or Canadian or British or French. There’s a consistent pattern in that they often treat the practices of what they view as a peer city or country as obvious examples of what to do. For example, an American policy paper on Social Security privatization might explain the Chilean system, and recommend its implementation, without much consideration of whether it’s really best industry practice. Such papers end up at best moving sideways, and at worst perpetuate the cycle of failure, by giving governments the appearance of reform while they in fact cycle between bad options, or occasionally stumble upon a good idea but then don’t understand how to implement it correctly.

If New York wants to study whether design-build is a good idea, it’s not enough to put it in the list of recommendations. It needs to do the legwork and read what the best experts say (e.g. Melis is opposed to it) and look at many cities at once to see what they do. I would feel embarrassed writing a long report like this with only 7 case studies from outside the US. I’d want to examine many more: on the cheap side, Stockholm, Milan, Seoul, Barcelona, Madrid, Athens, Naples, Helsinki; on the expensive side, London, Singapore, Hong Kong, Toronto, Melbourne, Munich, Amsterdam; in between, Paris, Tokyo, Brussels, Zurich, Copenhagen, Vienna. On anything approaching the RPA’s budget for the paper, I’d connect with as many people in these places as I could in order to do proper comparative analysis.

Instead, the RPA put out a paper that acknowledges the cost difference, but does not make a real effort to learn and improve. It has a lot of reform ideas, but most come from the same process that led to the high construction costs New York faces today, and the rest come from London, whose construction costs would astound nearly everyone in the world outside the US.

One of the things I learned working with TransitMatters is that some outside stakeholders, I haven’t been told who, react poorly to non-American comparison cases, especially non-English-speaking ones. Ignorant of the world beyond their borders, they make up excuses for why knowledge that they don’t have is less valuable. Even within the group I once had to push back against the cycle of failure when someone suggested a nifty-looking but bad idea borrowed from a low-transit-use American city. The group’s internal structure is such that it’s easy for bad ideas to get rejected, but this isn’t true of outside stakeholders, and from my conversation with Tom Wright about Gateway I believe the RPA feels much more beholden to the same stakeholders.

The cycle of failure that the RPA participates in is not the RPA’s fault, or at least not entirely. The entire United States in general and New York in particular is resistant to outside ideas. The political system in New York as well as the big nonprofits forms an ecosystem of Americans who only talk to other Americans, or to the occasional Canadian or Brit, and let bad ideas germinate while never even hearing of what best industry practices are. In this respect the RPA isn’t any worse than the average monolingual American exceptionalist, but neither is it any better.

The Subway in New York is not at Capacity

It seems to be common wisdom that the subway in New York is at capacity. Last year, the New York Times ran an article that repeated the MTA’s claims that growing delays come from overcrowding (which they don’t). A few weeks ago the NY Times quoted Riders Alliance campaign manager Rebecca Bailin saying “Our system is at capacity” and “subways are delayed when people can’t fit in them.” So far so good: some parts of the subway have serious capacity issues, which require investment in organization and electronics (but not concrete) to fix. But then some people make a stronger claim saying that the entire system is at capacity and not just parts of it, and that’s just wrong.

A few days ago there was an argument on Twitter involving the Manhattan Institute’s Nicole Gelinas and Alex Armlovich on one side and Stephen Smith on the other. Stephen made the usual YIMBY point that New York can expect more population growth in the near future. Nicole argued instead that no, there’s no room for population growth, because the subway is at capacity. Alex chimed in,

People are not going to be willing to pay market rents for places they can’t commute from. A large number of folks underestimate the self-regulation of NYC housing–it just can’t get that bad, because people can always just move to Philly

Like, if upzone Williamsburg, people who move into new housing aren’t going to try to ride the L–they’ll only come if they can walk/bike or ride in off-peak direction. Just like people are leaving in response to the shutdown. Neighborhoods and cities are in spatial equilibrium!

I responded by talking about rents, but in a way my response conceded too much, by focusing on Williamsburg. The L train has serious crowding problems, coming from lack of electrical capacity to run more than 20 trains per hour per direction (the tracks and signals can handle 26 trains, and could handle more if the L train had tail tracks at its 8th Avenue terminal). However, the L train is atypical of New York. The Hub Bound Report has data on peak crowding into the Manhattan core, on table 20 in appendix II. The three most crowded lines entering the Manhattan core, measured in passengers per floor area of train, are the 2/3, 4/5, and L. Those have 3.6-3.8 square feet per passenger, or about 3 passengers per m^2, counting both seated and standing passengers; actual crowding among standees is higher, around 4 passengers per m^2. Using a study of seating and standing capacity, we can get exact figures for average space per standee, assuming all seats are occupied:

Line Peak tph Seats Standee area Passengers Passengers/m^2
1 18 7,920 3,312 13,424 1.66
2/3 Uptown 23 9,200 4,393 28,427 4.38
A/D 17 9,792 3,980 23,246 3.38
B/C 13 6,994 2,899 12,614 1.94
4/5 Uptown 24 8,640 4,752 28,230 4.12
6 21 9,240 3,864 21,033 3.03
F Queens 13 5,967 3,560 17,816 3.33
N/Q/R 23 10,908 6,179 29,005 2.93
E/M 22 8,568 5,856 22,491 2.38
7 24 9,504 5,227 20,895 2.18
L 19 6,080 4,321 23,987 4.14
J/M/Z 19 6,384 4,363 16,657 2.68
F Brooklyn 14 6,426 3,834 14,280 2.05
B/D/N/Q (4 tracks) 38 18,612 10,008 43,550 2.49
A/C 20 10,112 4,504 21,721 2.58
2/3 Brooklyn 16 6,400 3,056 13,536 2.34
R 8 4,608 1,873 5,595 0.53
4/5 Brooklyn 20 7,200 3,960 16,504 2.35

Three additional snags are notable: crowding in 53rd Street Tunnel looks low, but it averages high crowding levels on the E with low crowding levels on the M (see review), and the 1 and 7 achieve peak crowding well outside Midtown (the 1 at 96th at the transfer to the 2/3 and the 7 at Jackson Heights at the transfer to the E/F) whereas the table above only counts crowding entering Manhattan south of 59th Street. But even with these snags in mind, there is a lot of spare capacity on the Upper West Side away from 72nd and Broadway, and in Queens in Long Island City, where passengers can take the undercrowded 7 or M. Crowding in Brooklyn is also low, except on the L. In both Brooklyn and on the West Side locals there’s also track capacity for more trains if they are needed, but New York City Transit doesn’t run more trains since peak crowding levels are well below design guidelines.

This isn’t a small deal. Williamsburg is where there is the most gentrification pressure, but the Upper West Side is hardly a slum – it’s practically a byword for a rich urban neighborhood. The trains serving Brooklyn pass through some tony areas (Park Slope) and gentrified ones (South Brooklyn), as well as more affordable middle-class areas further south. From NYCT’s perspective, developing South Brooklyn and Southern Brooklyn is especially desirable, since these areas are served by trains that run through to Queens, Uptown Manhattan, and the Bronx, and with the exception of the B are all much more crowded at the other end; in effect, lower subway demand in Brooklyn means that NYCT is dragging unused capacity because of how its through-service is set up.

Actual perceived crowding is always higher than the average. The reason is that if there is any variation in crowding, then more passengers see the crowded trains. For example, if half the trains have 120 passengers and half have 40, then the average number of passengers per train is 80 but the average perceived number is 100, since passengers are three times likelier to be on a 120-person train than on a 40-person train.

Subway in New York has high variation in crowding, probably unusually high by international standards, on account of the extensive branching among the lines. The E/M example is instructive: not only are the E trains more crowded than the M trains, but also they come more often, so instead of a perfect E-M alternation through 53rd Street, there are many instances of E-E-M, in which an E train following the M is more crowded than an E train following another E train. I criticized NYCT’s planning guidelines on this account in 2015, and believe it contributes to higher crowding levels on some lettered lines than the table shows. However, the difference cannot be huge. Evidently, in the extreme example of trains with 40 or 120 passengers, the perceived crowding is only 25% higher than the actual average, and even the maximum crowding is only 50% higher. Add 50% to the crowding level of every branching train in Brooklyn and you will still be below the 2/3 and 4/5 in Uptown Manhattan.

So on the Upper West Side and in Long Island City and most of Brooklyn, there is spare capacity. But there’s more: since the report was released, Second Avenue Subway opened, reducing crowding levels on the Upper East Side. Second Avenue Subway itself only has the Q, and could squeeze additional trains per hour by shuffling them around from other parts of the system. In addition, the 4/5 and 6 have reportedly become much more tolerable in the last year, which suggests there is spare capacity not only on the Upper East Side but also in the Bronx.

Moreover, because the local trains on Queens Boulevard aren’t crowded, additional development between Jackson Heights and Queens Plaza wouldn’t crowd the E or F trains, but the underfull M and R trains. This creates a swath of the borough, starting from Long Island City, in which new commuters would not have a reason to use the parts of the system that are near capacity. It’s especially valuable since Long Island City has a lot of new development, which could plausibly spill over to the east as the neighborhood fills; in contrast, new development on the Upper West Side runs into NIMBY problems.

Finally, the residential neighborhoods within the Manhattan core, like the Village, are extremely desirable. They also have active NIMBY groups, fighting tall buildings in the guise of preservation. But nowhere else is it guaranteed that new residential development wouldn’t crowd peak trains: inbound trains from Brooklyn except the 4/5 are at their peak crowding entering Lower Manhattan rather than Midtown, so picking up passengers in between is free, and of course inbound trains from Uptown and Queens drop off most of their peak morning load in Midtown.

It’s not just a handful of city neighborhoods where the infrastructure has room. It’s the most desirable residential parts of Manhattan and Brooklyn, and large swaths of middle-class areas in Brooklyn and parts of Queens. In those areas, the subway is not at capacity or even close to it, and there is room to accommodate new commuters at all hours of day. To the extent there isn’t new development there, the reason is, in one word, NIMBYism.

Construction Costs: Metro Stations

It is relatively easy to come up with a database of urban rail lines and their construction costs per kilometer. Construction costs are public numbers, reported in the mass media to inform citizens and taxpayers of the costs of public projects. However, the next step in understanding what makes American construction costs (and to a lesser extent common law construction costs) so high is breaking down the numbers. The New York Times published an excellent investigative piece by Brian Rosenthal looking at why Second Avenue Subway specifically is so expensive, looking at redundant labor and difficulties with contractors. But the labor examples given, while suggestive, concern several hundred workers, not enough for a multibillion dollar cost difference. More granularity is needed.

After giving examples of high US construction costs outside New York, I was asked on social media whether I have a breakdown of costs by item. This motivated me to look at station construction costs. I have long suspected that Second Avenue Subway splurged on stations, in two ways: first, the stations have full-length mezzanines, increasing the required amount of excavation; and second, the stations were mostly excavated from inside the tunnel, with only a narrow vertical access shaft, whereas most subway lines not crossing under older lines have cut-and-cover stations. The data I’m going to present seems to bear this out.

However, it is critical to note that this data is much sparser than even my original post about construction costs. I only have data for three cities: New York, London, and Paris.

In New York, Second Avenue Subway consisted of three new stations: 96th Street, 86th Street, and 72nd Street. Their costs, per MTA newsletters: 72nd Street cost $740 million, 86th Street cost $531 million, 96th Street cost $347 million for the finishes alone (which were 40% of the costs of 72nd and 86th). MTA Capital Construction also provides final numbers, all somewhat higher: 72nd Street cost $793 million, 86th Street cost $644 million, 96th Street cost $812 million. The 96th Street cost includes the launch box for the tunnel-boring machine, but the other stations are just station construction. The actual tunneling from 96th to 63rd Street, a little less than 3 km, cost $415 million, and systems cost another $332 million. Not counting design, engineering, and management costs, stations were about 75% of the cost of this project.

In Paris, Metro stations are almost a full order of magnitude cheaper. PDF-p. 10 of a report about Grand Paris Express gives three examples, all from the Metro rather than GPX or the RER, and says that costs range from €80 million to €120 million per station. Moreover, the total amount of excavation, 120,000 m^3, is comparable to that involved in the construction of 72nd Street, around 130,000 m^3, and not much less than that of 86th Street, around 160,000 m^3 (both New York figures are from an article published in the Gothamist).

A factsheet about the extension of Metro Line 1 to the east breaks down construction costs as 40% tunneling, 30% stations, 15% systems, and 15% overheads. With three stations and a total cost of €910 million over 5 km, this is within the range given by the report for GPX. The tunneling itself is according to this breakdown €364 million. An extension of Line 12 to the north points toward similar numbers: it has two stations and costs €175 million, with all tunneling having already been built in a previous extension. Piecing everything together, we get the following New York premiums over Paris:

Tunneling: about $150 million per km vs. $90 million, a factor of 1.7
Stations: about $750 million per station vs. $110 million, a factor of 6.5
Systems: about $110 million per km vs. $35 million, a factor of 3.2
Overheads and design: 27% of total cost vs. 15%, which works out to a factor of about 11 per km or a factor of 7 per station

Rosenthal’s article documents immense featherbedding in staffing the TBMs in New York, explaining much more than a factor of 1.7 cost difference. This is not by itself surprising: Parisian construction costs are far from Europe’s lowest, and there is considerable featherbedding in operations (for example, train driver productivity is even lower than in New York). It suggests that Paris, too, could reduce headcounts to make tunnel construction cheaper, to counteract the rising construction costs of Grand Paris Express.

But the situation with the stations is not just featherbedding: the construction technique New York chose is more expensive. The intent was to reduce street disruption by avoiding surface construction. Having lived on East 72nd Street for a year during construction, I can give an eyewitness account of what reducing disruption meant: there was a giant shaft covering about half the width of Second Avenue, reducing sidewalk width to 7 feet, between 72nd and 73rd Streets. This lasted for years after I’d moved away, since this method is so expensive and time-consuming. Under cut-and-cover, this disruption would cover several blocks, over the entire length of the station, but it would be finished quickly: the extension of Line 12 is currently in the station digging phase, estimated to take 18 months.

London provides a useful sanity check. Crossrail stations are not cut-and-cover, since the line goes underneath the entirety of the Underground network in Central London. Canary Wharf is built underwater, with 200,000 m^3 of excavation and 100,000 m^3 of water pumped; it’s technically cut from the top, but is nothing like terrestrial cut-and-cover techniques. The cost is £500 million. It’s a more complex project than the comparably expensive stations of Second Avenue Subway, but helps showcase what it takes to build stations in areas where cut-and-cover is not possible.

Another useful sanity check comes from comparing subway lines that could use cut-and-cover stations and subway lines that could not. Crossrail is one example of the latter. The RER A’s central segment, from Nation to Auber, is another: Gare de Lyon and Chatelet-Les Halles were built cut-and-cover, but in the case of Les Halles this meant demolishing the old Les Halles food market, excavating a massive station, and moving the Metro Line 4 tunnel to be closer to the newly-built station. The total excavated volume for Les Halles was about 560,000 m^3, and photos show the massive disruption, contributing to the line’s cost of about $750 million per km in today’s money, three times what Paris spends on Metro extensions. In London, all costs are higher than in Paris, but without such difficult construction, the extension of the Underground to Battersea is much cheaper than Crossrail, around $550 million per km after cost overruns and mid-project redesigns.

The good news is that future subway extensions in the United States can be built for maybe $500-600 million per km rather than $1.5-2 billion if stations are dug cut-and-cover. This is especially useful for Second Avenue Subway’s phase 2, where the segments between the station boxes already exist thanks to the aborted attempt to build the line in the 1970s, and thus cut-and-cover stations could simply connect to already-dug tunnels. It could also work for phases 3 and 4, which cross over rather than under the east-west lines connecting Manhattan with Queens and Brooklyn. The same technique could be used to build outer extensions under Utica and Nostrand in Brooklyn. Among the top priorities for New York, only a crosstown subway under 125th Street, crossing under the north-south line, would need the more expensive station construction technique; for this line, a large-diameter TBM would be ideal, since there would be plenty of space for vertical circulation away from the crossing subway lines.

There would still be a large construction cost premium. Changing the construction method is not enough to give New York what most non-English-speaking first-world cities have: getting down to $200 million per kilometer would require changes to procurement and labor arrangements, to encourage competition between the contractors and more efficient use of workers. Evidently, overheads are a larger share of Second Avenue Subway cost than of Parisian costs. But saving money on stations could easily halve construction costs, and aspirationally reduce them by a factor of three or four.

Grand Paris Express Cost Overruns: Organization Before Electronics Before Concrete

Paris is building a suburban Metro expansion, consisting of 200 km of which 160 are underground, carrying automated trains. This program, dubbed Grand Paris Express, is intended to provide circumferential service in the inner suburbs (on future Metro Lines 15, 16, and 18) and some additional radial service from the suburbs into Central Paris (on future Line 17 and extensions of Lines 11 and 14). The estimated cost was about €25 billion in 2012 prices – about average for a European subway. But now a bombshell has dropped: the cost estimate should be revised upward by 40%, to 35 billion for the 200 km GPX scheme and €38 billion for GPX plus related projects (such as GPX contribution to the RER E extension). You can read it in English-language media on Metro Report, but more detail is available in French-language media, such as Le Monde, and in the original report by the Cour des Comptes, the administrative court charged with auditing government finances. The goal of this post is to suggest how Ile-de-France should react to the cost overruns, using best industry practices from neighboring countries.

First, it’s worthwhile to look at the problems the Cour des Comptes report identifies. It includes a moderate amount of scope creep, on page 40, which helped raise the budget by €3.5 billion between 2013 and 2017:

  • €592 million for separate maintenance facilities at Aulnay for M15 and the other lines (M14, M16, M17).
  • €198 million for interoperability between two segments of M15 in the south and east; the original plan made M15 not a perfect circle but a pinch, without through-service between south and east, and building connections to permit through-running at the southeast costs extra.
  • €167 million for a second railyard for storing trains on M15 East.

On page 47, there is a breakdown of the larger cost overrun accumulated in 2017, by segment. The bulk of the overrun comes from new risk assessments: whereas the budget in early 2017 was €22.4 billion plus €2.8 billion for contingency, the new cost estimate is €27.7 billion plus €7 billion for contingency. This is a combination of geological risk and management risk: the report criticizes the project for lacking enough management to oversee such a large endeavor, and recommends target costs for each segment as well as better cost control to reduce risk.

Reducing the scope of GPX to limit its cost is thankfully easy. For a while now I have puzzled over the inclusion of M18 and M17 (which the report calls M17 North, since M17 South is shared with M16 and M14 in an awkward branch). Whereas M15 is a circular line just outside city limits, serving La Defense and many other major inner-suburban nodes, and M16 is another (semi-)circular alignment to the northeast of M15, M18 is a southwestern circumferential far from any major nodes, connecting Versailles, Massy-Palaiseau, and Orly on a circuitous alignment. Between the major nodes there is very little, and much of what it does connect to is already parallel to the RER B and to one branch of the RER C, which is being replaced with an orbital tram. The suburbs served are high-income and have high car ownership, and transit dependence is unlikely, making M18 an especially weak line.

M17 North is weak as well. It is a weird line, an underground radial connecting to Charles-de-Gaulle, already served by the RER B and by the under-construction CDG Express money waste. The route is supposed to be faster than the RER B, but it is no more direct, and makes more stops – the RER B runs a nonstop train between Gare du Nord and the airport every 15 minutes off-peak. It serves hotels near Saint-Lazare better using the connection to M14, but the RER B serves these hotels, as well as the hotels near Etoile, using a wrong-way transfer at Chatelet-Les Halles with the RER A.

The Cour des Comptes report itself does not recommend pruning these two lines, but its cost-benefit calculations per line on page 29 suggest that they should be deleted. On page 30 it says outright that the cost-benefit calculation for M18 is unfavorable. But on page 29 we see that the benefit-cost ratio of M18, not counting contingency costs, is barely higher than 1, and that of M17 North is a risky 1.3. In contrast, M15 South, the section already under construction, has a benefit-cost ratio of 1.7. M15 West has a ratio of 2.3, M15 East 1.5, M14 South 2.1, and M16 about 2. The M11 eastern extension is not included on the list.

Blog supporter Diego Beghin brought up on social media that M17 and M18 are already most at risk, and local elected officials are seeking assurances from the state that these lines will not be canceled. However, given their low potential ridership, the state should cancel them over local objections. Their combined cost is €4.9 billion, or €6.3 billion with contingency, about the same as the total cost overrun since early 2017.

Instead of pouring concrete on tunnels through lightly-developed high-income southwestern suburbs and on a third express route to the airport, the region should learn from what Germany and Switzerland have had to do. Germany has higher construction costs than France, which has forced it to prioritize projects better. Swiss construction costs seem average or below average, but the entire country has only two-thirds the population of Ile-de-France, and the public’s willingness to subsidize transit as a social service is much smaller than that of the French public. Hence the Swiss slogan, electronics before concrete (and its German extension, organization before electronics before concrete).

The M18 route already has a mainline rail route paralleling it – one of the branches of the RER C. This is an awkward branch, allowing trains from Versailles to enter the core trunk from either east or west, and ridership is so low that SNCF is downgrading it to an orbital tram-train. Thus, there is no need for a new Versailles-Massy connection. Two more destinations of note, Orsay and Orly, are also not necessary. Orsay is notable for its university, but there is already a connection from the university to Massy-Palaiseau and the city via the RER B with a little bit of walking to the station, and the connection to Versailles isn’t important enough to justify building a new line. Orly is a major airport, with about 90,000 travelers per day, but most of the traffic demand there is to the city (which it will connect to via the M14 South extension), and not to Versailles. While many tourists visit Versailles, this is just one stop on their journey, and their hotels are in the city or perhaps near Eurodisney in Marne-la-Vallee.

The M17 route is a more complex situation. The only new stops are Le Mesnil-Amelot, beyond the airport, with little development; Le Bourget-Aeroport, on the wrong side of a freeway; and Triangle de Gonesse, which is farmland. All three are development sites rather than places with existing demand, and development can be built anywhere in the region. However, the new airport-city connection is interesting, as relief for the RER B.

That said, there are better ways to relieve the RER B. The RER B has trains running nonstop between the airport and the city, but only off-peak. At the peak, trains run local every six minutes, with another branch (to Mitry-Claye) also getting a train every six minutes. The trains are very crowded, with obstructed corridors and not enough standing space in the vestibules, and with 20 trains per hour on the RER B and another 12 on the RER D, delays are common. Fixing this requires some improvement in organization, and some in concrete.

The concrete (and electronics) improvement is easier to explain: the shared RER B and D tunnel is a bottleneck and should be quadrupled. With four tracks rather than two, there would be space for more RER B as well as RER D trains; 24 trains per hour on each would be easy to run, and 30 would be possible with moving-block signaling of the same kind used on the RER A. This would provide more capacity not just to the northeast, around Aulnay-sous-Bois, but also north and northwest, since the RER D could take over more branches currently used by Transilien H.

The cost of quadrupling the tunnel is hard to estimate. Local rail advocate group ADUTEC explains the problem. In 2003 a proposal was estimated to cost €700 million, but construction would disrupt service, and in 2013 a study proposed new stations platforms at Les Halles for the RER D, raising the project’s cost to €2-4 billion. ADUTEC instead proposes building one track at a time to avoid disruption without building new platforms, saying this option should be studied more seriously; the cost estimate has to be higher than €700 million (if only because of inflation), but should still not be multiple billions.

But this project, while solving the capacity problems on the RER to the north and south in the medium term, doesn’t help connect passengers to the airport. On the contrary: more RER B traffic would make it harder to fit express trains between the local trains. Already there is little speed difference between local and express trains, about four minutes with nine skipped stations. This isn’t because the trains accelerate so quickly (they don’t) or because the maximum line speed is so low (the maximum speed on the line is 110-120 km/h). Rather, it’s because otherwise the express trains would catch up with local trains, on the airport branch or on the Mitry branch.

Fortunately, the route between the approach to Gare du Nord and Aulnay-sous-Bois, where the two RER B branches diverge, has four tracks. Right now, two are used by the RER, and two by other trains, including Transilien K but also the odd intercity train. The organizational fix is then clear: the four tracks should be reassigned so that all local trains get two tracks and all express trains (including intercities and Transilien but also airport express trains) get the other two. There is very little intercity traffic on the route, which carries no TGVs, and Transilien K has only a handful of peak trains and can be folded into the RER B.

With four tracks between Gare du Nord and Aulnay, express trains could go at full speed, saving about a minute for each skipped stop. But they shouldn’t go nonstop to the airport. They should serve Aulnay, giving it fast trains to the center. Passenger boardings by time of day are available for the SNCF-owned portion of the RER and Transilien here; Aulnay is the busiest station on the RER B north of Gare du Nord, with about 20% more weekday boardings than the second busiest (Stade de France) and 25% more morning peak boardings than the second busiest (La Courneuve). If express trains stop there, then it will free more space on local trains for the stations closer in, which would permit a service plan with half local trains and half express trains, each coming every 4-5 minutes. Today the inner stations get a local train every 3 minutes, so this is a service cut, but letting express trains handle demand from Aulnay out, on the airport branch as well as the Mitry and Transilien K branch, would mean passengers wouldn’t clog the local trains as much.

Potentially this could also reduce the demand for M16, whose northern segment, currently planned to be interlined with M14 and M17, is radial rather than circumferential. The entire M16 has a high benefit-cost ratio, but this could change in the presence of more RER B and D capacity. It may even be prudent to consider canceling M15 East and rerouting the remainder of M16 to complete the circle, a Line 15 consisting of the segments planned as M15 South, M15 West, and M16.

The study shows there is demand for two circumferentials in the east and northeast (M15 East and M16), but if RER B improvements rob M16 of its usefulness as a radial then this may change. If RER B improvements reduce the benefit-cost ratio of M16 below 1.5, then it should be canceled as well; with a budget of €4.4 billion plus another €1.2 billion in contingency, M16 could fund radial improvements that are more useful elsewhere. M15 East is a more coherent circumferential, with connections to Metro lines, whereas M16 is too far out.

But despite lack of coherence, M16 serves key destinations on the RER B. By default, the plan for GPX should be canceling M17 North and M18, and instead quadrupling the RER B and D tunnel and running more north-south RER service. Further cost overruns should be limited by the mechanisms the Cour des Comptes proposes, including tighter oversight of the project; without M17, there also may be room for removing ancillary scope, such as the Aulnay railyard.

Transit and Scale Variance Part 2: Soviet Triangles

Continuing with my series on scale-variance (see part 1), I want to talk about a feature of transit networks that only exists at a specific scale: the Soviet triangle. This is a way of building subway networks consisting of three lines, meeting in a triangle:

The features of the Soviet triangle are that there are three lines, all running roughly straight through city center, meeting at three distinct points forming a little downtown triangle, with no further meets between lines. This layout allows for interchanges between any pair of lines, without clogging one central transfer point, unlike on systems with three lines meeting at one central station (such as the Stockholm Metro).

The name Soviet comes from the fact that this form of network is common in Soviet and Soviet-influenced metro systems. Ironically, it is absent from the prototype of Soviet metro design, the Moscow Metro: the first three lines of the Moscow Metro all meet at one point (in addition to a transfer point one station away on Lines 1 and 3). But the first three lines of the Saint Petersburg Metro meet in a triangle, as do the first three lines of the Kiev Metro. The Prague Metro is a perfect Soviet triangle; Lines 2-4 in Budapest, designed in the communist era (Line 1 opened in 1896), meet in a triangle. The first three lines of the Shanghai Metro have the typology of a triangle, but the Line 2/3 interchange is well to the west of the center, and then Line 4 opened as a circle line sharing half its route with Line 3.

Examples outside the former communist bloc are rarer, but include the first three lines in Mexico City, and Lines 1-3 in Tehran (which were not the first three to open – Line 4 opened before Line 3). In many places subway lines meet an even number of times, rather than forming perfect diameters; this is especially bad in Spain and Japan, where subway lines have a tendency to miss connections, or to meet an even number of times, going for example northwest-center-southwest and northeast-center-southeast rather than simply crossing as northwest-southeast and northeast-southwest.

But this post is not purely about the Soviet triangle. It’s about how it fits into a specific scale of transit. Pure examples have to be big enough to have three subway lines, but they can’t be big enough to have many more. Moscow and Saint Petersburg have more radial lines (and Moscow’s Line 5 is a circle), but they have many missed connections, due to poor decisions about stop spacing. Mexico City is the largest subway network in the world in which every two intersecting lines have a transfer station, but most of its lines are not radial, instead connecting chords around city center.

Larger metro networks without missed connections are possible, but only with many three- and four-way transfers that create crowding in corridors between platforms; in Moscow, this crowding at the connection between the first three lines led to the construction of the Line 5 circle. In many cases, it’s also just difficult to find a good high-demand corridor that intersects older subway lines coherently and is easy to construct under so much older infrastructure.

The result is that the Soviet triangle is difficult to scale up from the size class of Prague or Budapest (not coincidentally, two of the world’s top cities in rail ridership per capita). It just gets too cumbersome for the largest cities; Paris has a mixture of radial and grid lines, and the Metro still undersupplies circumferential transportation to the point that a circumferential tramway that averages 18 km/h has the same ridership per km as the New York City Subway.

It’s also difficult to scale down, by adapting it to bus networks. I don’t know of any bus networks that look like this: a handful of radial lines meeting in the core, almost never at the same station, possibly with a circular line providing crosstown service. It doesn’t work like this, because a small-city bus network isn’t the same as a medium-size city subway network except polluting and on the surface. It’s scaled for minimal ridership, a last-resort mode of transportation for the poorest few percent of workers. The frequency is a fraction of the minimum required to get even semi-reasonable ridership.

Instead, such networks work better when they meet at one city center station, often with timed transfers every half hour or hour. A crosstown line in this situation is useless – it cannot be timed to meet more than one radial, and untimed transfers on buses that come every half hour might as well not even exist. A source who works in planning in Springfield, Massachusetts, a metro area of 600,000, explained to me how the Pioneer Valley Transit Authority (PVTA) bus system works, and nearly all routes are radial around Downtown Springfield or else connect to the universities in the area. There are two circumferential routes within Springfield, both with horrifically little ridership. Providence, too, has little to no circumferential bus service – almost every RIPTA bus goes through Kennedy Plaza, except some outlying routes that stay within a particular suburb or secondary city.

The principle here is that the value of an untimed transfer depends on the frequency of service and to some extent on the quality of station facilities (e.g. shelter). Trains in Prague come every 2-3 minutes at rush hour and every 4-10 minutes off-peak. When the frequency is as low as every 15 minutes, transferring is already questionable; at the typical frequency of buses in a city with a bus-based transportation network, passengers are extremely unlikely to do it.

This raises the question, what about denser bus networks? A city with enough budget for 16 buses running at once is probably going to run 8 radii (four diameters) every half hour, with a city-center timed transfer, and service coverage extending about 24 minutes out of the center in each direction. But what happens if there’s enough budget for 60 buses? What if there’s enough budget for 200 (about comparable to RIPTA)?

Buses are flexible. The cost of inaugurating a new route is low, and this means that there are compelling reasons to add more routes rather than just beef up frequency on every route. It becomes useful to run buses on a grid or mesh once frequency rises to the point that a downtown timed transfer is less valuable. (In theory the value of a timed transfer is scale-invariant, but in practice, on surface buses without much traffic priority, schedules are only accurate to within a few minutes, and holding buses if one of their connections is late slows passengers down more than not bothering with timing the transfers.)

I know of one small city that still has radial buses and a circular line: Växjö. The frequency on the main routes is a bus every 10-15 minutes. But even there, the circular line (bus lines 2 and 6) is a Yamanote-style circle and not a proper circumferential; all of the buses meet in the center of the city. And this is in a geography with a hard limit to the built-up area, about 5-6 km from the center, which reduces the need to run many routes in many different directions over longer distances (the ends of the routes are 15-20 minutes from the center).

There’s also a separate issue, different from scale but intimately bundled with it: mode share. A city with three metro lines is capable of having high transit mode share, and this means that development will follow the lines if it is given the opportunity to. As the three lines intersect in the center, the place for commercial development is then the center. In the communist states that perfected the Soviet triangle, buildings were built where the state wanted them to be built, but the state hardly tried to centralize development. In Stockholm, where the subway would be a triangle but for the three lines meeting at one station, the lack of downtown skyscrapers has led to the creation of Kista, but despite Kista the region remains monocentric.

There is no chance of this happening in a bus city, let alone a bus city with just a handful of radial lines. In a first-world city where public transit consists of buses, the actual main form of transportation is the car. In Stockholm, academics are carless and shop at urban supermarkets; in Växjö, they own cars and shop at big box stores. And that’s Sweden. In the US, the extent of suburbanization and auto-centricity is legendary. Providence has some inner neighborhoods built at pedestrian scale, but even there, car ownership is high, and retail that isn’t interfacing with students (for example, supermarkets) tends to be strip mall-style.

With development happening at automobile scale in smaller cities with smaller transit networks, the center is likely to be weaker. Providence has more downtown skyscrapers than Stockholm, but it is still more polycentric, with much more suburban job sprawl. Stockholm’s development limits in the center lead to a smearing of commercial development to the surrounding neighborhoods (Spotify is headquartered two stops on the Green Line north of T-Centralen, just south of Odengatan). In Providence, there are no relevant development limits; the tallest building in the city is empty, and commercial development moves not to College Hill, but to Warwick.

With a weaker center, buses can’t just serve city center, unless the operating budget is so small there is no money for anything else. This is what forces a bus network that has money for enough buses to run something that looks like a transit network but not enough to add rail to have a complex everywhere-to-everywhere meshes – grids if possible, kludges using available arterial streets otherwise.

This is why bus and rail networks look so profoundly different. Bus grids are common; subway grids don’t exist, except if you squint your eyes in Beijing and Mexico City (and even there, it’s much easier to tell where the CBD is than by looking at the bus map of Chicago or Toronto). But by the same token, the Soviet triangle and near-triangle networks, with a number of important examples among subway network, does not exist on bus networks. The triangle works for cities of a particular size and transit usage intensity, and only in rapid transit, not in surface transit.

Quick Note: U-Shaped Lines

Most subway lines are more or less straight, in the sense of going north-south, east-west, or something in between. However, some deviate from this ideal: for example, circular lines. Circular lines play their own special role in the subway network, and the rest of this post will concern itself only with radial lines. Among the radials, lines are even more common, but some lines are kinked, shaped like an L or a U. Here’s a diagram of a subway system with a prominently U-shaped line:

Alert readers will note the similarity between this diagram and my post from two days ago about the Washington Metro; the reason I’m writing this is that Alex Block proposed what is in effect the above diagram, with the Yellow Line going toward Union Station and then east along H Street.

This is a bad idea, for two reasons. The first is that people travel in lines, not in Us. Passengers going from the west end to the east end will almost certainly just take the blue line, whereas passengers going from the northwest to the northeast will probably drive rather than taking the red line. What the U-shaped layout does it put a one-seat ride on an origin-and-destination pair on which the subway is unlikely to be competitive no matter what, while the pairs on which the subway is more useful, such as northeast to southwest, require a transfer.

The second reason is that if there are U- and L-shaped lines, it’s easy to miss transfers if subsequent lines are built:

The purple line has no connection to the yellow line in this situation. Were the yellow and red line switched at their meeting point, this would not happen: the purple line would intersect each other subway line exactly once. But with a U-shaped red line and a yellow line that’s not especially straight, passengers between the purple and yellow lines have a three-seat ride. Since those lines are parallel, origin-and-destination pairs between the west end of the purple line and east end of the yellow line or vice versa require traveling straight through the CBD, a situation in which the subway is likely to be useful, if service quality is high. This would be perfect for a one-seat or two-seat ride, but unfortunately, the network makes this a three-seat ride.

The depicted purple line is not contrived. Washington-based readers should imagine the depicted purple line as combining the Columbia Pike with some northeast-pointing route under Rhode Island Avenue, maybe with an additional detour through Georgetown not shown on the diagram. This is if anything worse than what I’m showing, because the purple/red/blue transfer point is then Farragut, the most crowded station in the city, with already long walks between the two existing lines (there isn’t even an in-system transfer between them.). Thus the only direct connection between the western end of the purple line (i.e. Columbia Pike) and what would be the eastern end of the yellow line (i.e. H Street going east to Largo) requires transferring at the most crowded point, whereas usually planners should aim to encourage transfers away from the single busiest station.

When I created my Patreon page, I drew an image of a subway network with six radial lines and one circle as my avatar. You don’t need to be a contributor to see the picture: of note, each of the two radials intersects exactly once, and no two lines are tangent. If the twelve ends of six lines are thought of as the twelve hours on a clock, then the connections are 12-6, 1-7, 2-8, 3-9, 4-10, and 5-11. As far as possible, this is what subway networks should aspire to; everything else is a compromise. Whenever there is an opportunity to build a straight line instead of a U- or L-shaped lines, planners should take it, and the same applies to opportunities to convert U- or L-shaped lines to straight ones by switching lines at intersection points.

What Washington Metro Should Build

I’ve been thinking intermittently about how to relieve the capacity crunch on the Washington Metro. The worst peak crowding is on the Orange Line heading eastbound from Arlington to Downtown Washington, and this led to proposals to build a parallel tunnel for the Blue Line. Already a year ago, I had an alternative proposal, borrowing liberally from the ideas of alert reader Devin Bunten, who proposed a separate Yellow Line tunnel instead. Matt Yglesias’s last post about it, using my ideas, made this a bigger topic of discussion, and I’d like to explain my reasoning here.

Here is the map of what I think Metro needs to do:

Existing stations have gray fill, new ones have white fill. The Yellow Line gets its own route to Union Station, either parallel to the Orange Line and then north via the Capitol (which is easier to build) or parallel to the Green Line (which passes closer to the CBD), and then takes over the route to Glenmont. The rump Red Line then gets a tunnel under H Street, hosting the busiest bus in the city, and then takes over the current Blue Line to Largo, with an infill station in Mayfair for a transfer to the Orange Line and another at Minnesota Avenue for bus connections.

The Blue Line no longer presents a reverse-branch. It is reduced to a shuttle between the Pentagon and Rosslyn. Matt mistakenly claims that reducing the Blue Line to a shuttle is cost-free; in fact, it would need dedicated tracks at Rosslyn (if only a single track, based on projected frequency), an expensive retrofit that has also been discussed as part of the separate Blue Line tunnel project. At the Pentagon, initially shared tracks would be okay, since the Yellow Line is still a branch combined with the Green Line today; but the separate Yellow Line tracks would then force dedicated turnback tracks for the Blue Line at the Pentagon as well. Frequency should be high all day, and at times of low frequency (worse than about a train every 6 minutes), the lines in Virginia should be scheduled to permit fast transfers between both the Yellow and Orange Lines and the Blue Line.

The reverse branch today limits train frequency at the peak, because delays on one line propagate to the others. Peak capacity on Metro today is 26 trains per hour. I don’t know of anywhere with reverse-branching and much higher capacity: the London Underground lines that reverse-branch, such as the Northern line, have similar peak traffic, whereas ones that only conventionally branch (Central) or don’t branch at all (Victoria) are capable of 35-36 peak trains per hour. This means that my (and Devin’s, and Matt’s) proposed system allows more capacity even in the tunnel from Rosslyn to Foggy Bottom, which gets no additional connections the way 14th Street Bridge gets to feed a new Yellow Line trunk.

The big drawback of the plan is that the job center of Washington is Farragut, well to the west of the Yellow and Green Lines. WMATA makes origin-and-destination data publicly available, broken down by period. In the morning peak, the top destination station for each of the shared Blue and Yellow Line stations in Virginia is either the Pentagon or Farragut; L’Enfant Plaza is also high, and some stations have strong links to Gallery Place-Chinatown. Metro Center is actually faster to reach by Yellow + Red Line than by taking the Blue Line the long way, but Farragut is not, especially when one factors in transfer time at Gallery Place. The saving grace is that eliminating reverse-branching, turning Metro into four core lines of which no two share tracks, allows running trains more frequently and reliably, so travel time including wait time may not increase much, if at all.

This is why I am proposing the second alternative for the route between L’Enfant Plaza and Union Station. Devin proposed roughly following the legacy rail line. In the 1970s, it would have been better for the region to electrify commuter rail and add infill stops and just run trains on the route, and today a parallel route is appealing; Matt even proposed using the actual rail tunnel, but, even handwaving FRA regulations, that would introduce schedule dependency with intercity trains, making both kinds of trains less reliable. This route, the southeastern option among the two depicted in dashed lines, is easier to build, in that there are multiple possible streets to dig under, including C and E Streets, and giant parking lots and parks where the tracks would turn north toward the Capitol and Union Station. It also offers members of Congress and their staffers a train right to the officeUnfortunately, it forces Farragut-bound riders to transfer to the Orange Line at L’Enfant Plaza, slowing them down even further.

The second alternative means the Yellow Line stays roughly where it is. Four-tracking the shared Yellow and Green Line trunk under 7th Street is possible, but likely expensive. Tunneling under 8th Street is cheaper, but still requires passing under the Smithsonian Art Museum and tunneling under private property (namely, a church) to turn toward H Street. Tunneling under 6th Street instead is much easier, but this is farther from 7th Street than 8th Street is, and is also on the wrong side for walking to Metro Center and points west; the turn to H Street also requires tunneling under a bigger building. By default, the best route within this alternative is most likely 8th Street, then.

A variant on this second alternative would keep the Red Line as is, and connect the Yellow Line to the subway under H Street and to Largo. This is easier to construct than what I depict on my map: the Yellow Line would just go under H Street, with a Union Station stop under the track and new access points from the tracks to a concourse at H Street. This would avoid constructing the turns from the Red Line to H Street next to active track. Unfortunately, the resulting service map would look like a mess, with a U-shaped Red Line and an L-shaped Yellow Line. People travel north-south and east-west, not north-north or south-east.

Under either alternative, H Street would provide subway service to most of the remaining rapid transit-deprived parts of the District west of the Anacostia River. Some remaining areas near the Penn and Camden Lines could benefit from infill on commuter rail, and do not need Metro service. The big gaps in coverage in the District would be east of the river, and Georgetown.

Georgetown is the main impetus for the Blue Line separation idea; unfortunately, there’s no real service need to the east, along K Street, so the separate Blue Line tunnel would be redundant. In the 1970s it would have been prudent to build a Georgetown station between Foggy Bottom and Rosslyn, but this wasn’t done, and fixing it now is too much money for too little extra ridership; Bostonian readers may notice that a similar situation arises at the Seaport and BCEC, which should be on the Red Line if it were built from scratch today, but are unserved since the Red Line did not go there in the 1900s and 10s, and attempting to fix it by giving them their own subway line is a waste of money.

East of the river, the Minnesota Avenue corridor would make a nice circumferential rapid bus. But there are no strong radial routes to be built through it; the strongest bus corridor, Pennsylvania Avenue, serves a small node at the intersection with Minnesota and thereafter peters out into low-frequency branches.

This means that if the Yellow Line separation I’m proposing is built, all parts of the District that could reasonably be served by Metro will be. If this happens, Metro will have trunk lines with frequent service, two not branching at all and two having two branches on one side each; with passengers from Alexandria riding the Yellow Line, the Orange crush will end. The main issue for Metro will then be encouraging TOD to promote more ridership, and upgrading systems incrementally to allow each trunk line to carry more trains, going from 26 peak trains per hour to 30 and thence 36. Washington could have a solid rapid transit skeleton, which it doesn’t today, and then work on shaping its systems and urban layout to maximize its use.