The material far from the middle does nearly all the work

Take a steel ruler and try to bend it flat-on. It bends easily. Turn it on edge and it will not bend at all by hand.

Nothing about the steel changed. The same amount of the same material is present, and the difference in stiffness is a factor of several hundred. It comes entirely from where the material sits relative to the axis it is being bent about.

Why the square, and not the distance #

Bending a beam makes one face longer and the other shorter, and somewhere between them is a surface whose length does not change — the neutral axis.

A strip of material at distance $y$ from that axis is stretched in proportion to $y$. Being elastic, its stress is proportional to its strain, so the stress is also proportional to $y$. That is one factor.

The second factor is the lever arm. The strip’s force acts at distance $y$ from the axis, so its contribution to the resisting moment is force times $y$.

Stress proportional to $y$, times area, times a lever arm of $y$, gives a contribution proportional to $y^2$. That is the whole derivation, and the quantity summed over the section is the second moment of area:

$$I = \int y^2 \, dA.$$

The two factors are worth keeping separate, because each on its own would give a linear relationship. It is the coincidence of the outer material being both more highly stressed and further from the axis that produces the square, and the square is what makes the whole of structural shaping worthwhile.

The hero figure’s section is 220 deep by 90 wide, so it holds 19,800 mm² of material and its second moment is $90 \times 220^3/12 = 79.86 \times 10^6$. Reshaping that same 19,800 mm² — not adding to it — says what the square is worth.

The exponent in that sweep is worth stating precisely, because it is not the one usually quoted. A rectangle’s $I$ is $bd^3/12$, so at constant breadth the depth enters as a cube. At constant area it cannot, because $b$ must fall as $d$ rises: $I = Ad^2/12$, and the depth enters as a square. Both are true and they answer different questions — the cube is what a deeper beam of the same width buys, and the square is what a designer with a fixed tonnage of steel can buy. The second is nearly always the question being asked.

What it buys #

A rectangle of breadth $b$ and depth $d$ has

$$I = \frac{bd^3}{12},$$

which grows as the cube of the depth. Doubling the depth multiplies the stiffness by eight and the strength by four, at twice the material.

That is the ruler. On edge, the depth is the ruler’s width and the breadth is its thickness; flat-on, the two swap. For a ruler 25 mm by 1 mm the ratio of the two second moments is $(25/1)^2 = 625$.

The I-section is the conclusion drawn from the strip diagram, and it is the same argument as truss depth at the scale of a cross-section. If the outer material does nearly all the work, the middle can be reduced to whatever holds the outer material apart and carries the shear. What remains is two flanges and a thin web — a shape that would be structurally absurd for any purpose except this one.

The theorem that makes it computable #

Real sections are assembled from pieces, and the second moment of a piece about the section’s axis is not the same as about its own. The parallel-axis theorem converts between them:

$$I = I_{\text{own}} + A d^2,$$

where $d$ is the distance from the piece’s own centroid to the section’s neutral axis.

For a flange of an I-section, the two terms are wildly unequal. A flange 110 mm wide and 12 mm thick has $I_{\text{own}} = 15{,}800\,\text{mm}^4$ about its own axis, and sitting 94 mm from the section’s neutral axis it contributes $A d^2 = 1320 \times 94^2 = 11.7\,\text{million}$. The transfer term is seven hundred times the flange’s own contribution.

That ratio is the argument in a number. A flange’s own bending stiffness is irrelevant; everything it does comes from being somewhere else. Which is why a flange can be thin and wide, and why moving it further out is worth far more than making it thicker.

Every figure on this site that quotes a second moment computed it this way — summing $I_{\text{own}} + Ad^2$ over the rectangles that make up the profile — rather than looking it up.

From stiffness to strength #

The second moment governs stiffness. Strength needs one more step, because failure happens at the most stressed fibre rather than on average.

The stress at distance $y$ is $\sigma = My/I$, so the largest stress is at the extreme fibre, $y = c$. Setting that to the material’s limit gives the moment the section can carry:

$$M = \frac{\sigma I}{c} = \sigma Z, \qquad Z = \frac{I}{c}.$$

$Z$ is the section modulus, and it is the strength property while $I$ is the stiffness property. For a rectangle, $Z = bd^2/6$ — the square rather than the cube, because dividing by $c$ removes one power of depth.

So depth is worth more for stiffness than for strength: doubling it gives eight times the stiffness and four times the strength. That difference is the reason deflection and strength do not fail at the same span, and the reason long-span beams are almost always governed by how far they move.

Where the neutral axis is #

For a symmetric section the neutral axis is at mid-depth. For an asymmetric one it is at the centroid of the area, and it is not obvious by eye.

That matters because $c$ — the distance to the extreme fibre — is then different on the two faces, so the section has two section moduli. A tee-section has a large $Z$ for the flange side and a small one for the stem side, and it is efficient when the moment always acts in the direction that puts the flange in compression and hopeless when it reverses.

This is why a tee is used for a beam that always sags and an I-section for one that might do either, and why the reversal of load direction during construction has caused more than one failure in a member sized for the other sign.

The free body the integral is taken on #

$I$ is written as an integral over an area, which makes it look like a property of a shape drawn on paper. It is a property of a cut face, and the two equations that produce it are equilibrium equations on the piece of beam behind that face.

Cut the beam. On the exposed face there is a distribution of direct stress, and the piece behind it must satisfy the same two conditions everything else on this site satisfies.

The force equation locates the axis. If the beam carries no axial force, the stresses on the cut face must sum to zero:

$$\int \sigma \, dA = 0.$$

With $\sigma$ proportional to $y$ — which is what plane sections staying plane, plus elasticity, delivers — this becomes $\int y\,dA = 0$, and the only line about which the first moment of an area vanishes is the one through its centroid. The neutral axis is not chosen, and it is not at mid-depth except by symmetry. It is forced to the centroid by the requirement that a beam in pure bending carries no net push or pull.

The moment equation produces $I$. Summing the moments of those same stresses about that same axis gives the moment on the face:

$$M = \int \sigma y \, dA = k\int y^2\,dA = kI,$$

where $k$ is the constant of proportionality between stress and distance. Eliminating $k$ gives $\sigma = My/I$, which is the formula everything else in this field rests on.

So both halves of the elastic bending theory are one free body and two sums, and the second moment of area is what the second sum happens to be called. There is nothing about shapes in it at all — the integral would be the same if the section were a probability distribution or a plate of soup, which is why the identical mathematics turns up as the moment of inertia in dynamics with a mass density in place of an area.

It is worth registering how long the first of those two equations took to be believed. Galileo, in 1638, assumed the entire cut face was in tension and the beam rotated about its lower edge, which puts the neutral axis at the bottom face and overstates a rectangular beam’s capacity by a factor of three. Parent had the correct answer in 1713 and was ignored. Navier established it in 1826. The obstacle was not the integral — it was accepting that a cut face carries compression on one side and tension on the other simultaneously, which no experiment of the period could show and which the force equation demands.

Everything out there has to be held there #

The square is an invitation to push material outward, and the invitation has a limit that arrives from a different subject entirely.

A flange far from the neutral axis is a wide, thin plate carrying compression, and a wide thin plate in compression buckles locally — it ripples, in waves whose length is comparable to the plate width, at a stress that has nothing to do with the strength of the material and everything to do with the ratio of the width to the thickness. Once the flange has rippled, its outer portions can no longer carry their share, the effective section is smaller than the drawn one, and the neat $\sigma = My/I$ has stopped applying because the material it integrated over is no longer participating.

The profession handles this by classifying sections rather than by computing them: a section whose plate elements are stocky can reach its full plastic capacity, one a little more slender can reach first yield, and one more slender still fails locally before that. Three different capacities for three sections whose $I$ differs hardly at all. A web is classified separately from a flange, and the worse of the two governs the member.

The web has its own version of the same problem. Thin enough, it buckles in shear — diagonal ripples running corner to corner — and the answer is transverse stiffeners, which are pieces of steel that carry no bending moment whatsoever and exist solely so that something else keeps the shape it was drawn with. A plate girder is largely an argument about how many of them are needed.

The general principle underneath is worth stating plainly, because it recurs across the whole of structural design: efficiency and robustness pull opposite ways. Material moved to where it works hardest is material that has less to hold it in place, and the failure that eventually governs is not the one being optimised against. Every step of the sequence from a solid rectangle to an I-section to a plate girder buys section properties and sells stability, and the transaction is only worth it as long as the stability can be bought back more cheaply than the material was saved.

The best bending shape is the worst torsion shape #

There is a price for the I-section that the strip diagram cannot hint at, and it is severe.

Take a $300 \times 150$ I-section with $15$ mm flanges and a $10$ mm web — around $7{,}200\,\text{mm}^2$ of steel. Its second moment of area about the major axis is about $108 \times 10^6\,\text{mm}^4$. Now take a square hollow section of the same area, roughly $200$ mm square with a $9.5$ mm wall. Its second moment of area is about $44 \times 10^6\,\text{mm}^4$ — the I-section is two and a half times better in bending, exactly as the argument of this essay predicts.

Their torsional constants go the other way, and not by a similar margin. An open section resists twist only by each of its plates bending through its own thickness, giving roughly $\frac{1}{3}\sum bt^3$, or about $0.43 \times 10^6\,\text{mm}^4$ for the I. A closed section resists twist with a shear flow running continuously round the tube, giving about $65 \times 10^6\,\text{mm}^4$ for the box. The hollow section is about a hundred and fifty times stiffer in torsion, using the same steel, while being worse in bending.

That is not a small correction to the story; it is a different ranking produced by the same material. The reason is that the two problems reward opposite things. Bending wants material far from one axis, and an open section obliges. Torsion wants a closed loop of material, and cutting a tube open lengthways destroys nearly all of its torsional stiffness however far the material has been moved.

The consequence shows up wherever twist matters: a beam curved in plan, a spandrel carrying load on one side only, a crane runway, a bridge deck under an eccentric vehicle. All of them are built from closed sections, and the section catalogue is really two catalogues, chosen between by asking which of two quite different integrals the member has to satisfy.

The same quantity, doing a different job #

The second moment of area appears in two completely different failure calculations, and it is worth noticing that it is the same number.

Those last two figures are the essay in two pictures. Both hold 39,600 mm² — 220 × 180 in one and 440 × 90 in the other. Identical area, identical material, identical cost, and one of them is four times the other, because the strips in one are twice as far from the axis as the strips in the other and the contribution goes as the square of that distance. Everything else on this page is a consequence.

In bending, $I$ sets how far a beam moves under load. In buckling, it sets the load at which a column stops being able to stay straight. A section chosen for bending stiffness about one axis may have very little about the other, and as a column it is judged by the worse of the two.

Span dominates everything and $I$ is the only term available to fight it, which is why the entire catalogue of structural sections is an argument about where to put the material.

The centroid is the axis that minimises it #

There is a corollary of the parallel-axis theorem worth stating because it decides the sign of a common error.

$I = I_c + Ad^2$ with $d$ the distance between the two axes, and $Ad^2$ cannot be negative. So the second moment about the centroidal axis is the smallest of all the parallel ones, and every other axis gives a larger number.

Which means a mislocated centroid always overestimates the stiffness. A section whose centroid is taken a little wrong — an unsymmetric shape computed carelessly, a composite section whose transformed areas were wrong, a built-up member assumed symmetric — returns an $I$ that is too large, a deflection that is too small, and a natural frequency that is too high, in every case in the unsafe direction.

The error is quadratic in the misplacement, so a small one costs nothing and a large one costs a great deal, and there is no arrangement in which it is conservative.

Where the model stops #

Plane sections stay plane. The whole derivation assumes a flat cross-section remains flat as the beam bends, which makes the strain proportional to distance. That is very accurate for slender beams and progressively wrong for deep ones. A beam whose span is less than about twice its depth does not obey it at all, and needs a different theory.

Linear elastic material. Stress is taken as proportional to strain everywhere. Past yield it is not, and the section then has a plastic capacity larger than $\sigma Z$ — for a rectangle, half as much again, because the stress block becomes rectangular rather than triangular.

Bending about one axis. A section has a second moment about every axis through its centroid, and the two principal values can differ enormously. Loading a section about an axis that is not principal produces bending in two directions at once, which for an angle section is the normal case and a well-known trap.

Shear ignored. The whole calculation concerns direct stress from bending, and shear is the other thing a cut reveals. Shear stress has its own distribution, peaks at the neutral axis where the bending stress is zero, and governs short deep members.

The figures have a distortion worth stating: the strips in the first figure are drawn at equal thickness and their contribution bars at true relative length, but there are only eleven of them. The real integral is continuous, and the outer contributions are consequently even more dominant than eleven strips make them look — the top strip of a genuine integration carries a larger share than the diagram can show.

The ladder from here #

Later rungs on this anchor: the parallel-axis theorem derived. Section modulus, elastic and plastic. Principal axes and unsymmetric bending. The shear centre, which is where load must act to avoid twisting. Shear stress distribution and shear flow. Composite sections and the transformed-area method. Cracked sections in concrete. Torsional constants, which follow a completely different rule. And the radius of gyration, which is the same quantity again in the language of buckling.

The $bd^3/12$ result is in Parent’s work of 1713 and in Navier’s of 1826, and it took the intervening century for the profession to believe that the neutral axis sits at the centroid rather than at the bottom face.