When Trapezoidal Samples Miss Oscillation

Why a Smooth Graph Can Still Fool a Numerical Estimate

The trapezoidal rule estimates an integral by joining sampled values with straight line segments and adding the areas of the resulting trapezoids. If the function changes gradually between sample points, this can be an effective approximation. But a smooth function can also oscillate rapidly between those points. Its sampled values may then give little indication of what happens between them.

This is an important limitation: a numerical estimate is based on both the sampled values and assumptions about the integrand. The samples alone do not generally certify that an estimate is accurate.

The Composite Trapezoidal Rule

Let \(f\) be continuous on \([a,b]\), and divide the interval into \(n\) equal subintervals, where \(n\) is a positive integer. The step size and nodes are

The composite trapezoidal estimate for \(\int_a^b f(x)\,dx\) is

The endpoints receive half-weight in the formula, and each interior sample receives full weight. Equivalently, on each subinterval the rule uses the trapezoid whose parallel sides have heights given by the function values at that subinterval’s endpoints.

Worked Example: Three Oscillations Hidden at Every Node

Why Node Values Alone Cannot Guarantee Accuracy

The example is not just an unusual calculation. It shows a basic limitation of finite sampling. Consider \(f_A(x)=A\sin(3\pi x)\) on \([0,1]\), where \(A\) is any positive constant. At the same four nodes, every value is still zero, so \(T_3=0\) for every \(A\). Yet

For any chosen positive number \(M\), taking \(A=\frac{3\pi M}{2}\) makes the exact integral equal to \(M\), while the estimate remains zero. Thus the same node values can be consistent with integrals of arbitrarily large size. These functions are all smooth; smoothness by itself does not prevent the problem.

A Curvature Bound Gives an Error Guarantee

One useful kind of additional information is a bound on the function’s second derivative. Curvature matters because the trapezoidal rule replaces each piece of the graph by a straight line. If the graph bends sharply between nodes, that line can differ substantially from the function.

Why the Error Bound Works

Here is a derivation. On one subinterval \([x_j,x_{j+1}]\), let \(L(x)\) be the line joining the endpoint values of \(f\). The area under \(L\) is exactly the trapezoid’s area on that subinterval. The linear interpolation remainder says that, for each interior \(x\), there is a point \(\xi_x\) in the subinterval such that

Since \(|f''(\xi_x)|\leq M\), and the two factors have opposite signs inside the subinterval, this gives

Write \(t=x-x_j\), so \(0\leq t\leq h\) and \(x_{j+1}-x=h-t\). The difference between the exact area and the trapezoid’s area is at most

There are \(n\) subintervals. Adding their error bounds gives \(nMh^3/12\). Since \(h=(b-a)/n\), this is

This proves the stated bound. The important condition is the available bound \(M\) on \(|f''|\). If the function oscillates more rapidly, its second derivative may be much larger, so the bound may be large even when the step size is small.

Worked Example: Estimating an Error Using Curvature

A Common Mistake: Treating a Small Estimate as Proof of a Small Integral

A tempting but incorrect conclusion from the first example is: “All four samples are zero, so the integral must be zero.” The calculation only establishes that the trapezoidal estimate is zero:

It does not establish that the function is zero between the nodes. In fact, direct integration gives \(\int_0^1\sin(3\pi x)\,dx=2/(3\pi)\), not zero. The correction is to distinguish the estimate from the integral and seek information—such as a second-derivative bound—that controls their difference.

This issue matters in practice whenever a graph or function is sampled at a fixed spacing. If the spacing is coarse relative to the oscillation, important changes can occur between measurements. A plausible-looking numerical output is not, on its own, evidence that those changes are negligible.