The ends decide the length that matters

The Euler formula contains a length, and the length in it is not the length of the column. It is the length of the piece of the buckled shape that looks like a pin-ended column — and that depends entirely on what is holding the ends.

The shape says the length #

A pin-ended column buckles into a half sine wave: zero deflection at each end, maximum in the middle, and no curvature reversal anywhere. The length of that half wave is the length in the formula.

Change the ends and the shape changes, and the half wave that appears in it is a different length.

Both ends pinned. The whole column is one half wave. $K = 1$, by definition — this is the reference case.

Both ends fixed. The ends cannot rotate, so the shape has points of contraflexure a quarter of the way in from each. Between those two points the column is doing exactly what a pin-ended column does, and that stretch is half the height. $K = 0.5$, and the capacity is four times the pinned case.

One end fixed, one pinned. An intermediate shape with one contraflexure point. $K = 0.7$, capacity twice the pinned case.

Fixed at the base, free at the top. The worst case. The buckled shape is a quarter wave, so the half wave it belongs to is twice the height. $K = 2$, and the capacity is a quarter of the pinned case — and a sixteenth of the fixed-fixed case.

$$P_{\mathrm{cr}} = \frac{\pi^2 EI}{(KL)^2}.$$

A factor of sixteen across four arrangements of the same column. Nothing else available to a designer moves capacity that far.

Free to sway is the expensive word #

The fourth case is worth separating out, because it is the difference between two kinds of building frame and it is not primarily about rotation.

A column whose top is held against sideways movement — by a brace, a shear wall, a core, all of which are redundant restraints — can only buckle by bowing between its ends. $K$ is at most 1 and usually less.

A column whose top can move sideways buckles by leaning. The whole storey goes over together, and $K$ exceeds 1 — often substantially. That is a sway frame, and its columns are worth a fraction of what the same columns are worth braced.

The distinction is the reason bracing exists and the reason a building with a stiff core can use much lighter columns than one relying on frame action. It is also why removing a wall during refurbishment is a structural act even if the wall carries no vertical load: the wall may have been what stopped the frame swaying, and its removal doubles every column’s effective length.

The standard values are idealisations #

The four values assume perfect restraint or none, and reality is always in between.

A “fixed” end is fixed by whatever it frames into, and that thing has finite stiffness. A column welded to a beam that is itself flexible has a partly restrained end, with an effective length somewhere between the theoretical values. Design codes handle this with alignment charts or with a stiffness ratio at each end, and both are ways of interpolating.

The consequences of getting it wrong are asymmetric. Assuming more fixity than exists overestimates the capacity, which is unsafe. Assuming less is conservative and expensive. Codes therefore recommend values slightly worse than theory — the fixed-fixed case is usually designed at $K = 0.65$ rather than 0.5, and the sway cantilever at 2.1 rather than 2.0 — precisely because perfect fixity does not exist.

There is a second trap. The restraint can differ between the two axes, and a section’s two radii of gyration differ too. A column braced about one axis at mid-height by a beam framing in, and unbraced about the other, has two different effective lengths and two different slendernesses, and the governing one is whichever gives the larger $\lambda$. That is often not the axis with the smaller radius of gyration, and checking only the obvious axis is a well-known way to miss the real answer.

Where 0.7 comes from, and why it is not round #

Three of the four factors are exact fractions and one is not, which is a hint that they were not chosen for convenience.

The buckled shape satisfies $EI\,y'''' + P\,y'' = 0$, whose general solution is a sine, a cosine, a straight line and a constant. Each set of end conditions picks out different combinations, and the critical load is the smallest one at which a non-trivial shape can satisfy all four conditions at once.

Pinned–pinned gives $\sin kL = 0$, so $kL = \pi$ and $K = 1$ exactly.

Fixed–fixed gives $\cos kL = 1$ with a zero-slope condition, whose first useful root is $kL = 2\pi$, so $K = 0.5$ exactly.

Fixed–free gives $\cos kL = 0$, root $kL = \pi/2$, so $K = 2$ exactly.

Fixed–pinned gives something quite different:

$$\tan kL = kL,$$

a transcendental equation with no closed-form solution at all. Its first non-zero root is $kL = 4.4934$, so

$$K = \frac{\pi}{4.4934} = 0.6992.$$

The familiar $0.7$ is that root rounded, and the reason it looks like an engineering approximation is that it is the only one of the four that genuinely is one. There is no exact fraction available; the number is a root of an equation that has to be found numerically, and Euler’s contemporaries found it by trial.

This matters more than a piece of trivia, because it locates what the effective-length factor is. It is not a fudge, a safety factor or a rule of thumb. It is an eigenvalue: the smallest number at which a differential equation with those boundary conditions has a solution other than “stay straight”. Everything a code does to it afterwards — raising $0.5$ to $0.65$, raising $2.0$ to $2.1$ — is a correction for the boundary conditions not being what the equation assumed, and each of those adjustments is an admission about a joint rather than about the mathematics.

What a restraint has to be, as well as do #

A brace has to be strong enough, which the one-to-two-per-cent rule above covers. It also has to be stiff enough, and the relationship between the two turns out to be strange enough that it caught the profession out.

Consider a pin-ended column with a single elastic brace at mid-height, of stiffness $\beta$. As $\beta$ increases from zero, the buckling load rises — but not indefinitely. Above a particular value, called the ideal stiffness, the column buckles into two half waves instead of one and the brace point becomes a node. Beyond that, more stiffness buys nothing at all, because the mode has stopped using the brace. For a column with one central brace, the ideal stiffness is about $2P/L$.

Now the strange part, which is Winter’s result of 1958. At exactly the ideal stiffness, a perfectly straight column requires the brace to carry no force whatsoever — the brace point is a node, the column does not move there, and a restraint that is not displaced exerts nothing. Give the column a small initial bow, however, and the brace force at the ideal stiffness is infinite. The two statements are not in conflict: the force required rises without limit as the stiffness approaches the ideal value from above, and the initial imperfection sets the scale.

The practical resolution is to provide roughly twice the ideal stiffness, at which the brace force settles to something proportional to the initial bow and comfortably small — which is where the one-to-two-per-cent figure comes from. It is not a measurement of what the brace does; it is the consequence of a deliberate decision to sit well clear of a singularity.

Two things follow that are worth carrying. A brace check is always two checks, and the stiffness one is the one that gets forgotten, because a member strong enough to carry two per cent of anything is a very small member and looks obviously adequate. And a brace attached to something flexible has its stiffness in series with that flexibility, so a perfectly adequate brace bolted to a purlin that can move has an effective stiffness set by the purlin — which is how a restraint that exists in every respect on the drawing turns out not to be one.

The roof that was braced on paper #

The Hartford Civic Center’s roof was a steel space truss of about 90 by 110 metres, built in the early 1970s and among the first designed with the aid of a computer analysis. In the small hours of 18 January 1978, a few hours after a full arena had emptied, it fell into the building under a load of snow and ice well below what it had been designed for.

The investigations found no material defect and no gross overload. What they found concerned the compression members of the top chord and what was — and was not — holding them. The bracing arrangement as built did not restrain those members in the manner the analysis had assumed, and members whose effective length had been taken as one panel were in fact free over rather more than that. Since capacity falls with the square of the length, an assumption wrong by a factor in that quantity is not a small error, and once the first members had buckled the load they shed went to neighbours already in the same condition.

Three features make it a useful case rather than a merely sad one. The design analysis was not wrong in its arithmetic; it was wrong in an input, and an input of exactly the kind this essay is about. The structure had exhibited visible deflection during construction, roughly twice what was predicted, and this had been queried and explained away. And nothing about the failure was gradual — a stability failure of a set of members that share a condition arrives all at once, because whatever makes one of them fail is present in all of them.

The lesson the profession drew was procedural as much as structural: an effective length is an assumption about geometry that somebody has to verify on site, and it is not visible in the output of the analysis that consumed it.

Restraint does not have to be at the ends #

The effective length is set by the distance between points of restraint, and those points can be anywhere.

Adding a single lateral restraint at mid-height of a pinned column halves the buckling length and quadruples the capacity — the same reciprocal-square return working in the useful direction for once. That is an extraordinary return for one small member, and it is why bracing is the cheapest structural steel on any project.

The restraint has to be able to do its job, which means being both stiff enough and strong enough. The force it must resist is small — codes typically require something around 1 to 2.5 percent of the load in the member being restrained — but it is not zero, and a brace connected to something that itself moves is not a restraint.

This is where a great many failures live. A compression member restrained by a purlin that is connected to a roof sheet that is fixed to nothing in particular has a restraint on the drawing and not in the building.

That threshold is the answer to “stiff enough”, and it is worth carrying in the form the figure gives it: 1.41 N_cr/L. For a storey carrying a thousand kilonewtons over four metres it is about 0.35 kN per millimetre of sway, which is 1.93 times the frame’s own lateral stiffness. A brace has to be roughly twice as stiff as the thing it is bracing, and past that point it is finished — the curve is flat, and a heavier brace is steel spent on nothing.

What buckles is the whole frame #

For a single column the end conditions are an input. For a frame, they are an output — every member restrains its neighbours, and the frame buckles as a whole.

That is why modern analysis computes an elastic critical load factor for the entire structure rather than an effective length for each column. If the factor is large the frame is stiff and effective lengths near 1 are fine; if it is small the frame is sway-sensitive and second-order effects have to be included explicitly.

The effective-length method is a way of packaging that whole-frame behaviour into a single number per member, and it works well for regular frames and poorly for irregular ones. Where it fails, the honest answer is to analyse the frame with its imperfections included and let the amplification appear on its own.

Restraint costs something to provide #

A restraint is a member, and it has to be designed like one.

The load must go somewhere: a brace resisting one or two percent of a column’s load has to deliver that force into a stiff element, and a brace connected to something flexible is not a restraint at all. Most restraint failures are failures of the load path beyond the brace rather than of the brace itself.

A braced bay is a vertical truss, and treating it as one rather than as an afterthought is what turns a nominal restraint into a real one.

Two is not the worst case #

The table of standard cases runs from 0.5 to 2.0, and it is easy to read that as a range. It is not: 2.0 is the worst of four idealisations, not an upper bound on anything.

The four cases all assume the restraints are either perfect or absent. Real sway frames have restraints that are neither, and the effective length grows without limit as the restraint softens. A column fixed at its base with a top held by a beam of finite stiffness has a $K$ that rises above 2 as soon as the beam is less than rigid, and an alignment chart for a sway frame runs comfortably to 4 or 5 before anybody has drawn anything unusual.

Two curves, and only one of them is bounded. That is the whole content of the warning, and it is why the two halves of the table cannot be read the same way.

The consequence is arithmetic, and it is severe because the capacity goes as $1/K^2$. A column at $K = 2$ carries a quarter of its pin-ended value; at $K = 4$ it carries a sixteenth. So the gap between the table’s worst case and a genuinely poor real one is another factor of four, on a quantity the table implies has a floor.

Which is the practical reason a sway frame is not designed by looking up a $K$ at all. Either it is braced — at which point the restraint is real, the non-sway column of the table applies, and $K$ is below one — or it is analysed as a system, where the buckling load of the whole frame is found and the individual member’s effective length is a derived quantity rather than a chosen one.

The table is a table of braced cases with one sway case appended as a warning. Reading the warning as a limit is the single commonest misuse of it.

Where the model stops #

Perfect restraint. As above.

Uniform axial force. The formula assumes the load is constant along the column, which a free body cut at two heights shows it is not. A column in a multi-storey frame carries more at the bottom than the top, and its effective length is not a standard case.

Prismatic members. Tapered or stepped columns have a buckling length that is not a simple multiple of anything, and the counting arguments that work for a frame do not help.

Elastic behaviour. Effective length modifies the Euler curve, and the Euler curve is only the whole answer for slender members.

One member at a time. A frame buckles as a system; treating each column separately is a convenience that has to be checked.

The figures share an honest limitation: each buckled shape is drawn at a visible amplitude, which no column at its critical load has. The linear theory that produces $K$ says the shape and says nothing at all about the amplitude — the deflection is indeterminate at the critical load, which is exactly what makes it critical. Drawing a definite curve implies a definite state that the mathematics does not contain.

The ladder from here #

Later rungs: the four end conditions derived from the differential equation. Alignment charts and stiffness ratios. Sway and non-sway frames. The elastic critical load factor. Bracing forces and how much a restraint must resist. Restraint stiffness requirements. Buckling of frames as systems. Effective lengths in codes, and why they differ from theory. And the second-order analysis that removes the need for the concept altogether.

The effective-length factor is a nineteenth-century device for making a twentieth-century calculation possible with a slide rule. It has survived the arrival of the computer, largely because it packages an intuition that engineers still want.