The count that does not see it

The count is the first thing anybody learns about whether a pin-jointed frame will stand up. Count the members, count the reaction components, count the joints, and compare $m + r$ with $2j$. Fewer unknowns than equations and the frame is a mechanism; equal and it is determinate; more and it is redundant.

The count is correct as far as it goes, and how far it goes is the subject of this essay. It compares two numbers: how many unknown forces there are, and how many equilibrium equations the joints supply. What it cannot do is look at the equations. Two of them can be the same equation written twice, and the count will not notice, because the count never reads them.

Both frames in that figure pass the count. One of them is a building and the other is a parallelogram with pretensions.

What the count is actually comparing #

At each joint of a plane pin-jointed frame, two equations hold: the forces meeting there sum to nothing horizontally and nothing vertically. With $j$ joints that is $2j$ equations. The unknowns are one axial force per member and one component per restrained direction — $m + r$ of them.

Writing every joint equation at once gives a linear system, $\mathbf{A}\mathbf{x} = \mathbf{b}$, where $\mathbf{x}$ is the list of member forces and reactions, and each column of $\mathbf{A}$ holds the direction cosines of one member at the two joints it connects. This is the equilibrium matrix, and everything about the frame’s stability is a property of it.

The count compares the shape of $\mathbf{A}$: how many rows against how many columns. A square matrix, $m + r = 2j$, is the determinate case, and a square matrix usually has an inverse — which is what makes the count nearly always right.

Usually. A square matrix has an inverse when its rows are linearly independent, and a frame can be arranged so that they are not.

A frame that folds while the count says otherwise #

The clearest case has a name: the critical form, or in older books, the critical configuration. Three ways of producing one are worth knowing, and the first two are about the supports rather than the members.

Parallel reactions. Three vertical props under a triangle. The count is satisfied — three reaction components against three equations of overall equilibrium — and the horizontal equilibrium equation reads $0 = 0$ for the reactions no matter what happens, because none of them has a horizontal component. This is the equation that was never there, arriving as a stability failure rather than a curiosity.

Concurrent reactions. Three reaction lines meeting at a point. Taking moments about that point, every reaction has zero lever arm and drops out, so the moment equation cannot be satisfied by the reactions at all. The frame rotates about the meeting point.

The robbed panel. This one is internal, and it is the case that matters most in practice, because it is the one that can be built by accident. A two-panel frame needs one diagonal in each panel. Put both diagonals in the left panel and none in the right, and the total is unchanged: nine members, three reactions, six joints, $9 + 3 = 12 = 2 \times 6$. The left panel is now braced twice over, which is one redundancy, and the right panel is not braced at all, which is one mechanism. They cancel in the count and do not cancel in the structure.

The last is the most instructive because the redundancy and the mechanism are in different places, and no local inspection of either panel would call the frame determinate. Only the total does. It is also the one that gets built: a bracing layout is often set out on a drawing by someone counting braces per elevation rather than per panel, and the triangle that cannot fold is a statement about one triangle, not about a bag of them.

Why the joint-by-joint method never notices #

Hand analysis of a truss proceeds joint by joint: find a joint with no more than two unknown member forces, resolve there, move on. The procedure is the method of joints, and its convenience is exactly what hides the defect.

Walking the joints never assembles the whole system, so it never has the opportunity to discover that two of the equations agree. What happens instead is that the walk stalls: at some joint there are three unknowns and two equations, and no ordering of the joints avoids it. The stall is the symptom, and it is easily misread — as a frame that needs the method of sections, or as a redundant frame requiring compatibility, or as an arithmetic mistake made earlier.

There is a second, worse outcome. On the robbed panel the walk does not stall everywhere. Starting from the support it resolves the braced panel perfectly happily, because that panel really does have determinate-looking arithmetic locally, and only arrives at the impossible joint after several correct-looking steps. Numbers accumulate, they are all right, and the frame is a mechanism.

The test that does work #

The rank of the equilibrium matrix is the number of genuinely independent equations in it. For a frame that stands, the rank equals the number of rows: every joint equation says something the others do not. For a critical form, the rank is short.

$$\text{rank}(\mathbf{A}) < 2j \quad\Longrightarrow\quad \text{a mechanism, whatever the count says}$$

Rank is not a quantity a count can produce. It requires the elimination to be performed — the matrix reduced, the pivots found, and the columns without pivots identified. This is why the check is rare in hand analysis and automatic in any computer analysis: a solver that inverts $\mathbf{A}$ discovers the deficiency the moment it looks for a pivot and finds nothing but zeros.

The mechanism is in the null space #

Knowing that a frame folds is less interesting than knowing how, and the how is available from the same matrix at no extra cost.

A set of joint displacements that changes no member’s length is a motion the frame can make with no member developing any force — and a member with no force in it is not the same thing as a zero-force member, which carries nothing under one particular load case and is perfectly capable of carrying something under another. The condition for that is $\mathbf{A}^{\mathsf{T}}\mathbf{d} = \mathbf{0}$, where $\mathbf{d}$ lists the joint displacements — the transpose of the equilibrium matrix is the compatibility matrix, which is a duality worth pausing over. The same array of direction cosines, read across, says how forces balance at joints; read down, it says how member extensions follow from joint movements. Statics and kinematics are one matrix seen from two sides.

So the mechanism is a vector in the null space of $\mathbf{A}^{\mathsf{T}}$, and finding it is the same elimination that found the rank. The ghosted outlines in the figures above are that vector, scaled up until it is visible and drawn on the frame it belongs to. Nothing about them is artistic: the joints that do not move came back as zeros.

Which free body produced the number #

For the robbed panel, take the joint at the top of the unbraced panel as the free body: two chord members and one vertical meet there, and no diagonal.

Cut it out and the forces on it are three axial forces along three lines. Two of those lines — the top chord on either side — are collinear. So the equilibrium of that joint in the direction perpendicular to the chord involves exactly one member, the vertical, and the equation reads $N_{\text{vertical}} = 0$. That is a valid equation and it determines a force. What it does not do is restrain the joint horizontally in any way that the chord forces have not already covered, and the same is true at the joint below it.

Assemble the whole system and the deficiency appears as a rank one short of the row count. The frame’s own solver reports it: the equilibrium matrix for this frame has twelve rows and rank eleven, and the null vector of its transpose is the racking motion drawn above.

The site’s own gate refuses this case rather than drawing it: feeding a singular system to the truss solver has to return no answer, and the checks that back these figures include a frame of exactly this shape whose analysis must fail.

Where the model stops #

Small displacements. The mechanism is a motion the frame can make at its current geometry. Move it far enough and the geometry changes: the robbed panel, once racked over, is no longer a parallelogram of the same shape, and the chords are no longer collinear at the joint. A first-order mechanism can be a stable structure at second order, held by the geometry it deforms into rather than by the geometry it started in. That is the principle behind a cable net and behind a tent, and it is why “mechanism” is a statement about a configuration and not about a pile of parts.

Pin joints. Real joints are welded or bolted, so a critical form in a real frame does not fold freely — it deflects a great deal and then stops, held by the bending stiffness the analysis pretended was not there. The same reserve is what keeps a portal frame standing when its bracing is omitted, and it is bought at a price nobody costed. This is the least reassuring possible way for a structure to be safe, because the reserve depends on connections that were designed on the assumption they carried no moment, and the size of that secondary bending is a separate calculation nobody performed.

Two dimensions. The three-dimensional count is $m + r$ against $3j$, and the ways of arranging a rank deficiency are correspondingly richer. Space frames have critical forms that are genuinely hard to spot, and the useful ones — cable domes, tensegrity — are the deliberate exploitation of a first-order mechanism stiffened by prestress. The deliberate version of the same move at the level of one member is the hinge put in on purpose, where a release is added to make a structure determinate rather than to make it move.

Exactly critical, or nearly so. Rank is a yes-or-no property and structures are not. A frame whose diagonals are very nearly collinear, or whose reactions very nearly meet at a point, has full rank and is therefore stable by every test in this essay — and it will develop enormous member forces from small loads, because the matrix it is solved through is nearly singular and its inverse is correspondingly large. The forces are real, the analysis is right, and the frame is a bad frame. The useful working rule is that the near-critical case is more dangerous than the critical one, because the critical one refuses to be analysed and the near-critical one returns numbers.

That connects the stability question to a numerical one: the condition number of the equilibrium matrix measures how far the frame is from folding, in a way the binary rank test cannot. A well-triangulated truss with sensible panel proportions has a modestly conditioned matrix; a shallow one with long flat diagonals does not, which is another route to the same conclusion depth reaches from the direction of chord forces.

The figures share a limitation. A mechanism is drawn at an exaggeration, as a second outline superimposed on the first, and the drawing therefore shows the frame in a position it would only reach after moving a long way. The real motion begins at zero and is resisted by nothing, so there is no scale at which the picture is honest; what the picture can show truthfully is only the shape of the motion, and its magnitude is a decision made by the person drawing it.

The generalisation #

The pattern here — a count that is necessary and not sufficient, with the sufficient test being a rank — recurs across engineering wherever a mobility criterion appears. The Grübler and Kutzbach criteria for linkages have the same structure and the same failure mode, and mechanism designers exploit the exceptions deliberately: a linkage that the count says is rigid and that moves anyway is a paradoxical mechanism, and several of the most useful ones are exactly that.

What generalises is the epistemic shape rather than the formula. A count of constraints against a count of freedoms is a comparison of dimensions, and dimensions cannot see degeneracy. Whenever a criterion counts things that might coincide, there is a measure-zero set of arrangements on which it fails — and structural engineering is unusual in that the measure-zero set contains many of the arrangements a person would naturally draw, because people draw symmetrical things and symmetry is exactly what makes lines concurrent and parallel.

The history is short and slightly embarrassing. Maxwell gave the count in 1864 and stated its limitation in the same paper: his rule holds “unless the frame is in a critical position”, a caveat that was widely dropped in the transmission. Föppl and later Müller-Breslau worked through the exceptions in the 1880s and 1890s. Calladine returned to it in 1978 and showed that Maxwell’s rule is properly an equality between two defects — the number of independent mechanisms minus the number of states of self-stress — which is the modern statement and which makes the robbed panel exactly what it looks like: one of each, cancelling. The same defect count, applied to a structure with more restraints than equations, is the statement that a redundant frame has states of self-stress — internal force patterns in equilibrium with no load at all, which is the mirror image of a mechanism and turns up again the moment a support settles.

The ladder from here #

Later rungs on this anchor: the count in three dimensions, and the space frames that defeat it. States of self-stress, which are the other half of Calladine’s equality and the reason a frame can be redundant and a mechanism at once. Prestress stability, where a first-order mechanism is made stiff by tensioning it. Kinematic indeterminacy as the dual of static indeterminacy. And the numerical question of how nearly critical a frame has to be before its computed forces stop meaning anything — a matrix does not have to be singular to be useless, and a nearly critical frame produces enormous member forces from small loads well before the determinant reaches zero.