The connection is not a point, and every diagram so far says it is

Every diagram so far has been a lie of the same kind, told deliberately and told well. A truss is drawn as lines meeting at points. A portal frame is drawn as three lines with two corners. A beam is drawn as a line resting on two triangles. In each case the members have length and the joints do not, and the analysis that follows treats the joint as a coordinate: a place where forces are transferred, occupying no space and having no properties.

That idealisation is what makes structural analysis possible, and nothing here is going to retract it. What this field does is ask what was inside the point.

Where structures actually fail #

The engineering literature on structural failure is dominated by connections, and by a margin that is not close. Bridge collapses, roof collapses, the progressive failures that make textbooks — the fracture almost always starts where two pieces of steel were joined, or where a beam sat on a wall, or where a bolt passed through a plate.

It is tempting to read that as a statement about workmanship, and there is some of that. But the deeper reason is available from the theory itself, and it is the same reason the analysis had to idealise the joint in the first place.

Everything in the six fields that precede this one descends from one assumption about geometry: that a plane section stays plane. That assumption is what makes strain linear across a cut face, which makes stress linear, which produces the second moment of area, the section modulus, the deflection formulae and every capacity this site has computed.

It is justified by Saint-Venant’s principle, which says that the details of how a load is applied stop mattering at a distance of roughly one section depth from where it is applied. Beyond that distance the stress field settles into the smooth linear one the theory assumes.

Read that principle carefully and it says something else at the same time. Within about one section depth of a load application, a support, a change of section or a joint, the linear field does not hold. The theory does not merely become less accurate there; it is a statement about a region it explicitly excludes.

A connection is nothing but that region.

Two kinds of region, and the names for them #

The distinction is old enough to have shorthand. A B-region — B for Bernoulli, or for beam — is one where plane sections stay plane and beam theory applies. A D-region — D for disturbed, or discontinuity — is one where it does not.

The rule of thumb for the extent of a D-region is the same section depth Saint-Venant’s principle uses. A beam 400 mm deep has a D-region roughly 400 mm long at each end, at each point load, and around each hole. The rest is B-region, and the rest is where every formula in this collection works.

For a member of ordinary proportions the B-regions are most of it. For a short deep beam they may be none of it, which is the case where beam theory stops applying altogether. And for a connection, the D-region is the connection.

The block above is the clearest possible demonstration that the connection needs its own analysis. Nothing in bending, in shear flow or in buckling can express it. The plate is not bending. It is coming apart along two surfaces, and predicting when requires knowing where the surfaces are.

What the point was hiding #

Take the bracket at the top of this page and ask what the frame analysis knew about it.

It knew a force: 100 kN, vertical, applied at a node. It possibly knew a moment, if the frame model included one. What it did not know, and could not have known, is that the connection has a size — and that the size is what produces the answer.

The cleanest way to see that is to shrink the size and watch the answer come back to the frame’s own. Bring the bracket in until the load’s line of action almost passes through the bolts, and the group behaves the way a node behaves.

That figure is what the analysis assumed, and it is very nearly true when the connection is small compared with the forces crossing it. Nothing else about the drawing changed. Six bolts at the same pitch and gauge, the same steel, the same load — and the only quantity that moved was a length nobody put in the model.

The load arrives at the end of a bracket. The bolts are 150 mm away. That distance is not a modelling choice; it is set by how wide the beam is, how much room the bolts need, how a spanner reaches them. And that distance turns a 100 kN force into a 100 kN force plus a 15 kN·m torque, which the six bolts share in a way that gives one of them three times the average.

Lengthen the bracket and the same six bolts get worse without anything else being touched.

Three figures, one load, one bolt pattern, and a worst bolt that runs from 20.6 kN to 90.0 as a single distance goes from 25 mm to 300. The frame analysis reported 100 kN at a node for all three, and it was right all three times.

The general form is worth stating plainly:

A connection has dimensions, and every dimension is a lever arm for something.

This is the sentence that the whole field elaborates. The distance from a bolt to a web face is the lever arm that makes the flange bend and produces prying. The distance from a bolt to the edge of a plate is the lever arm that decides whether the plate tears out or crushes. The distance from a bolt row to the compression flange is the lever arm that turns bolt tension into moment capacity. The distance from the centroid of an angle to the leg it is bolted through is the lever arm that leaves part of the section not working.

None of those distances exists on a frame diagram. All of them decide the answer.

The three questions a connection has to answer #

Once the point becomes a region, the design question splits into three, and they are usually answered by different mechanisms with different failure modes.

Can the force get out of the first member? The force in a member is distributed over its section in the way beam theory says. It has to be collected out of that distribution and into whatever fastens the members together — bolts, welds, bearing. A section connected through only part of itself does not deliver all of its own capacity, which is shear lag, and the loss is geometry.

Can the fastenings carry it? This is the question that gets called “connection design” and it is only a third of the problem. Bolts in shear, bolts in tension, welds along a line, friction across an interface. Each has its own capacity and its own way of sharing between fasteners, and the sharing is rarely equal.

Can the second member take it? The force arrives at a small patch of the second member, which has to distribute it back into a smooth field. Web crippling, column web buckling, plate tear-out, base plate bending: every one of them is a member being asked to accept a load at a point.

A connection fails at whichever of the three is weakest, and there is no reason for that to be the one anybody checked. Each of the three is a whole essay elsewhere in this field, and each of them is a length on a drawing doing the deciding: the plate that bends and levers more into its bolt than was ever applied, the hole that elongates because the edge distance behind it was set from a table of minima, the angle bolted through one leg that never collects its whole section into the fastenings.

The middle question is the one the bracket at the top of this page is already answering, and it is worth pushing a little further, because the sharing between fasteners depends on the group’s shape and not only on its size.

Three bolts instead of six doubled the direct share and multiplied the torsional term by three, because taking a column of bolts away removed the whole $\sum x^2$ half of the polar second moment. A bolt group’s capacity is not the sum of its bolts, and the arithmetic that says so is one line long. Nothing in the frame analysis could have produced it, because the frame analysis has no $x$ and no $y$ for the bolts to be at.

Why this is a field and not a chapter #

There is a reason connections have not appeared until now, and it is not that they are advanced.

Every other field here can be entered from the middle. Bending stress can be read without knowing about buckling. Deflection can be read without knowing about plasticity. The results are separable because each concerns one member analysed one way.

Connections are not separable, because the connection is where every other question arrives at once. A moment end plate is simultaneously a bolt problem, a plate bending problem, a weld problem, a column web problem and a stiffness problem, and its answer is the worst of them. That is why this field comes eighth and why it needs the seven before it.

It is also why connection design in practice is so heavily codified. The number of possible failure modes is large, several of them are non-obvious, and the consequences of missing one are worse than for a member — a member that is 10% overstressed usually yields and redistributes, whereas a bolt group that is 10% overstressed may fracture without warning at all.

What replaces the smooth field #

If beam theory does not apply in a D-region, something has to.

The oldest answer, and still the most useful, is to abandon stress distributions entirely and go back to equilibrium. Draw a free body that cuts through the connection, sum the forces, and ask what mechanism carries each one. That is what a strut-and-tie model does for concrete and what the block-shear calculation does for steel: it identifies a path the load could take and checks whether every element of the path is adequate.

Its justification is the lower-bound theorem of plasticity, which is one of the most useful results in the subject and appears here in the same form it takes for masonry arches: if any distribution of internal forces can be found that is in equilibrium with the applied load and nowhere exceeds the material’s strength, the structure will carry that load. The right distribution does not have to be found. One does.

That theorem is what makes connection design possible at all. Nobody knows the real stress field in a bolted end plate — it is three-dimensional, involves contact, friction and plasticity, and changes as the joint is loaded. What can be done is to propose a mechanism, check every link in it, and rely on the material’s ductility to redistribute towards whatever mechanism was proposed.

Which is why ductility is a precondition rather than a bonus in this field. The lower-bound theorem requires that the material can deform enough to reach the assumed distribution. A connection whose components are all brittle has no guarantee at all, and this is exactly why the property that appears in none of the equations turns out to be doing the load-bearing work here as well.

The bracket at the top of this page has been solved that way three times over without saying so. The elastic vector method used for all three of those figures assumes a mechanism: it declares that the group rotates about its own centroid, and shares the torque in proportion to distance from it. That is one distribution in equilibrium with the load, so the lower-bound theorem says the group will carry it. It does not say it is the best one available.

Fourteen per cent of the group’s apparent capacity was hidden by an assumption, and the assumption was not wrong: both distributions are in equilibrium with 100 kN at 150 mm, both are admissible, and the theorem guarantees the group carries the smaller of the two loads they imply. The difference between them is what the plate’s ductility is spending. If the bolts and the plies could not deform enough to let the group find the second centre of rotation, the first answer would be the only one available — which is the whole of why ductility is a precondition here and not a bonus.

Why the idealisation is still right #

None of this is an argument for modelling connections in the frame analysis. It is worth being explicit about that, because the obvious reading of everything above is “so the frame model should include the joints”, and for almost all work the answer is no.

The frame analysis exists to answer a question the connection cannot: how the load divides between routes to the ground, which needs the whole structure at once and is a stiffness question rather than an equilibrium one. Adding six bolts and a plate to every node would multiply the model’s size by a large factor, would require knowing details of connections that have not been designed yet, and would answer the same question to no better accuracy — because the quantity the frame model produces is a set of member forces, and those are insensitive to almost everything the connection does except its rotational stiffness.

So the division of labour is deliberate and it is the right one. The frame model produces forces at nodes. The connection design takes those forces and asks how they get across, using a completely different method on a much smaller region. The two are checked against each other at exactly one property — the rotational stiffness, which is the only thing the connection does that the frame can feel — and everything else about the joint is invisible to the analysis by design rather than by omission.

What goes wrong is not the split. It is forgetting that the split happened, and reading the frame’s node force as a complete specification of what the connection has to do.

Notice which quantities have decided every answer on this page. Not the steel grade, not the bolt grade, not the thickness of anything: an eccentricity, a pitch, a gauge, a bolt count and a centre of rotation. Every one of them is a length or a tally on a drawing, and not one of them appears anywhere in the member analysis that produced the force being transferred. The same is true of the failures this field will spend its essays on — the position of a section’s own centroid against the leg it is bolted through, the distance from a hole to the edge of the plate behind it, the depth from a bolt row down to a compression flange.

The two idealisations, and the space between them #

Frame analysis offers two options for a joint: pinned or rigid. Both are convenient, both are used constantly, and both are false.

A pinned joint is one that transmits no moment. Real “pinned” connections — web cleats, fin plates, flexible end plates — transmit some. A rigid joint is one where the members’ ends rotate together. Real “rigid” connections deform, so they do not.

For most purposes the errors are acceptable and in the safe direction, but not always, and the figure above says why the question cannot be answered by looking at the connection alone. The boundaries are stiffnesses compared with EI/L. A joint is not rigid; a joint is rigid relative to a particular beam. Change the beam and the same connection changes class.

The consequence is a redistribution that nobody designed: a beam analysed as simply supported, with a connection that is actually semi-rigid, has real end moments the analysis never predicted and a smaller mid-span moment than it was designed for. Usually that is safe. It is not always, and it is never intentional.

And a real connection is all three questions at once, which is the reason none of them can be answered in isolation. A moment end plate has to get the force out of the beam’s flanges, share it between bolt rows that are not equally able to take it, and deliver the compression half of the couple into a column web that no bolt schedule mentions — and its capacity is the worst of those, with no reason at all for that to be the one somebody checked. The bolt group above is the same statement with two of the three questions held still.

The only part of a structure whose cost is not its weight #

There is one more reason connections deserve a field of their own, and it is not structural. It is that they are where the money is, and the arithmetic of that runs opposite to the arithmetic of everything else in the subject.

A member is priced by the tonne. Halve a beam’s weight and roughly halve its cost, which is why the whole of the preceding seven fields reads as an argument for efficiency: get the material far from the neutral axis, use the depth, pick the section that carries the most per kilogram. A connection is priced by the operation — a hole drilled, an edge cut, a stiffener fitted, a weld pass run, a coat of paint applied to a surface that is now more complicated than it was. Those costs scale with how many things have to be done, and barely at all with how much steel is in them.

The consequence is a standing inversion that surprises anybody arriving from member design. A heavier beam with a simpler connection is routinely cheaper than a lighter one with a complicated connection, and the crossover is nowhere near where an efficiency argument would put it. Two extra stiffeners in a column web can cost more than the several hundred kilograms of beam they were fitted to save. A moment connection costs several times a shear connection at the same node, so a frame braced somewhere else and connected with fin plates everywhere is cheaper than a rigid frame using less steel.

That is why the simple connection is the industry’s default rather than its compromise. Designing every joint as a pin, providing stability with a braced bay, and accepting the heavier beams that follow is not a failure to optimise — it is the optimisation, performed against the cost that dominates.

It also explains a habit that otherwise looks like laziness: connections are standardised, taken from tables, and repeated across a project even where a particular node could take something lighter. A detail used forty times is drawn once, checked once, fabricated with one setup and erected the same way every time. Variety is the expensive thing, and a connection schedule with four types in it beats one with fourteen even if the fourteen weigh less.

None of which changes a single calculation in this field. It changes what the calculations are for: the question a connection design answers is usually not “what is the lightest joint that works” but “which of the standard joints works here”, and the essays that follow are about being able to tell.

What this field will do #

Every essay in it takes a distance on a drawing and shows what that distance decides.

Some of them are about bolts: how a group shares an eccentric load, how a bolt in tension carries more than it was given, how a plate fails around a hole, and what a preloaded joint does that an ordinary one does not. Some are about welds, where the geometry is a line rather than a set of points, and where the capacity depends on the direction of the load in a way that comes out of the yield criterion rather than out of a test.

Some are about whole connections: the end plate that makes a moment cross a gap, the base plate where a structure meets its foundation, the classification that decides what a frame analysis is allowed to assume. The last of those runs out of the building entirely, and the arithmetic goes with it unchanged.

That is the same figure family and the same three terms as the bracket, two hundred pages of code apart and forty metres lower down, which is the reason this field is worth its own set of essays rather than a chapter appended to the others.

The thread through all of them is the sentence above. The point had a size, and the size is the answer.