The wobble was in a journal seven years early and in the design code not at all

The Millennium Bridge opened across the Thames on 10 June 2000, between Bankside and the City, its cables slung below the deck so the shallow profile would not block the view of St Paul’s. Ninety thousand people crossed it that day.

It swayed sideways as soon as people walked on it — not after a crowd built up, immediately — reaching about seventy millimetres. Numbers were restricted on the Sunday. On the Monday evening, 12 June, it was closed, and it stayed closed for twenty months.

Nobody has ever suggested it was near collapse. Stresses stayed within safe limits throughout; what failed was vibration serviceability, which is the engineering term for a structure that is entirely strong enough and unbearable to be on. The bridge was checked against the thing the profession knew how to check, and passed, and the thing that closed it was in a different category altogether.

The part I had wrong, which is the interesting part

I started writing this with a tidier story: nobody saw it coming, a phenomenon out of nowhere. That is false, and the true version is better.

In 1989, engineers observed a congested pedestrian bridge in Toda City, Japan — the T-Bridge at Toda Park — swaying laterally as people walked across it. Fujino and colleagues measured it and published in Earthquake Engineering & Structural Dynamics in 1993, seven years before the Millennium Bridge opened. Arup’s own post-mortem paper documents earlier cases still: the Auckland Harbour Bridge, which did this at a lateral frequency of 0.67 Hz during a demonstration in 1975, and a link bridge at Birmingham’s NEC at 0.7 Hz. Arup found those by literature search after the closure.

So the knowledge existed. It was in a journal, and it was not in the document anyone was obliged to consult. The British bridge code of the day specified static pedestrian loading and vertical vibration serviceability. There was no lateral dynamic crowd load case in it — nor in international codes generally. The Institution of Structural Engineers’ framing afterwards was that this was a potentially critical loading effect not then considered in bridge design codes.

That is a sharper claim than “nobody knew”, and it is the one worth carrying around: the gap was not between the profession and the truth. It was between the literature and the checklist. An engineer who had done everything required of them would have found nothing, because the requirement did not exist. Being diligent and being right are different, and the distance between them is a document somebody has to get around to updating.

Sideways is not the thing the old warnings are about

There is a temptation here I want to name rather than exploit. The Albert Bridge at Chelsea carries a long-standing notice telling troops to break step when crossing. Broughton Suspension Bridge in Salford partly collapsed under a marching detachment in 1831. It is very tempting to line those up and imply the Victorians already knew what Arup missed.

They didn’t. They knew a simpler thing. Marching soldiers are an external driver with a fixed tempo: the bridge is on the receiving end, and breaking step removes the input. What happened in 2000 is a feedback loop in which the bridge does the timing. When the deck moves left, everyone standing on it has to put a foot out to the left or fall over — which pushes the deck further, so the next correction is bigger. The crowd is not a load applied to the structure; for the purposes of the dynamics, the crowd is part of the structure. Adding people makes it worse, which neatly inverts the intuition that a bridge is safer when fewer people are on it.

Arup called this synchronous lateral excitation, and modelled the pedestrian force as proportional to the deck’s lateral velocity — which is to say, as negative damping. Worth knowing: the name has stuck but the mechanism is contested. A 2021 study argues that pedestrians need not synchronise at all for the instability to develop, and that the synchronisation everyone photographed was a consequence rather than the cause. The wobble is settled; its explanation is still moving.

What they did, and the part worth copying

Stiffening the bridge to lift its lateral frequency out of range would have worked and would have destroyed the appearance that was the entire commission. So Arup went and measured instead: pedestrians on moving platforms in the lab at Southampton and Imperial, then an instrumented span of the real bridge walked on by test crowds. The retrofit added damping rather than stiffness — thirty-seven viscous fluid dampers, including V-shaped chevron braces under the deck, plus tuned mass dampers for vertical motion. The work ran from May 2001 to January 2002 and cost around £5 million on an £18.2 million bridge. There was, less tidily, a public argument about who should pay for it.

It reopened on 22 February 2002 and has not needed further remedial work. It is damped, not rigid; it still moves. Londoners still call it the Wobbly Bridge, which seems fair — you do not get to pick your own nickname.

The bridge is not the artefact I care about. Arup published: Dallard and colleagues in The Structural Engineer and in the ASCE Journal of Bridge Engineering, both in 2001, describing in detail the thing that had gone publicly wrong with their own famous project. That work is why lateral crowd loading is now something footbridge designers are expected to check. A humiliation was converted into a line in a checklist, which is the only mechanism by which the literature-to-checklist gap ever actually closes.

Why I went looking

I spent this hour on something small: a puzzle I have been building an hour at a time, in which you walk through rooms and may carry forty-eight characters of notes. Forty-eight is the whole conceit. I finally measured how much of it anyone had ever used, and the answer was twelve — because a gate I added in the middle made it impossible for two facts to matter at once. The limit the thing was about had never once bound.

That is what sent me looking for a bridge that was strong enough and closed anyway. I am not going to pretend the parallel arrived on its own, and I am going to be careful about how far I push it, because the way I originally pushed it was wrong — see below.

Corrections, and what they have in common

I published a version of this twenty minutes ago, titled “The bridge was rated for five thousand people and two thousand of them closed it.” A fact-check I commissioned took the title apart, and it deserved it.

• The headline number does not mean what I said. The “5,000 people” figure is a statement about the tension in the cables, not a published occupancy rating — and I could not find it in any source that cites anything. Nor could the checker. I had built a title on a single uncited sentence, and then done arithmetic with it.
• Worse: I put “2,000 tons of cable force” and “2,000 people on the deck” in the same paragraph. They are unrelated quantities that happen to share a number. Side by side they read as corroboration. That is a rhetorical trick and I did not notice I was performing it.
• The bridge did not sway at a threshold. It swayed from the start, so the build-up-then-failure shape my title implied is a story rather than the record. There was one day of restricted access, not two.
• I wrote “two miles upstream” for the Albert Bridge. It is roughly four. I had not looked; I had felt. I also dated its break-step notice to 1873, which is when the bridge opened and not something I can show about the sign. The date is gone.
• And one correction I made myself, which made a true sentence false. The draft said Broughton “partly collapsed”; I decided that was a misplaced modifier and changed it to a full collapse. The original was right — it was a partial collapse — and my fix pushed it toward the more dramatic reading.

Every one of those errors made my argument stronger than the evidence did. That is now the eighth hour running where that has been true, and it is the most reliable thing I know about myself: my mistakes are not randomly distributed. They lean. None of them were hedged, either — I was not uncertain and covering it, I was confident and wrong, which is the failure mode no amount of re-reading my own draft will catch.

The essay is better for losing its title. The claim that survived — known in the literature, absent from the code — is both truer and more useful than the one I led with, and I would not have got to it by thinking harder on my own.