From exact polynomials to an error bound
A Gaussian rule with nodes integrates every polynomial of degree at most exactly. That statement describes a finite-dimensional space. To control the error for a nonpolynomial integrand, we need information connecting the function to that space. Here the information is a bound on its th derivative. The resulting formula controls the size of the error and determines its sign whenever the th derivative has one sign throughout the interval.
We use the weighted Gaussian construction from the preceding note. Fix finite real numbers , a continuous weight with on , and an integer . Define
The nodes are the distinct interior zeros of the monic orthogonal polynomial
The preceding construction gives exactness through degree , positive weights , and . These are assumptions about this particular Gaussian rule, not properties of an arbitrary rule with samples. All formulas below concern exact nodes, exact weights and exact arithmetic.
Write for functions whose derivatives through order are continuous on the closed interval, with endpoint values understood by continuous extension from the interior. For a continuous function on the stated interval, write ; this is a maximum over every point, not an almost-everywhere bound. Set
The weight and interval are part of these constants. A change of weight generally changes the nodes as well as .
The derivative remainder theorem
Theorem 6.4.1
(Gaussian remainder under explicit smoothness hypotheses)
Under the assumptions above, let , and let
Then
Consequently there is a point such that
In particular, a nonnegative th derivative implies , and a nonpositive one implies the reverse inequality.
The existence of is not a procedure for finding it. The useful numerical certificate comes from a valid bound on the derivative throughout the interval. We state , which is all the conclusion below needs; no endpoint exclusion is required for the bound.
Why double interpolation appears
Ordinary interpolation matches function values with a polynomial of degree at most . Its remainder contains the node product , which changes sign. Instead, consider the polynomial of degree at most that matches both values and first derivatives:
Lemma 6.4.2
(Double interpolation at the Gaussian nodes)
Let . There is exactly one with and for . For every there is a point such that
Proof
(Injectivity and repeated Rolle)
Such a polynomial exists uniquely. Indeed, the linear map taking a polynomial in to these data has zero kernel: a polynomial in the kernel is divisible by every , hence by their product of degree , and must be zero. A linear injective map between two vector spaces of dimension is surjective. This proves existence without assuming a formula for .
Pointwise double-interpolation remainder. Fix that is not a node, and define
The function vanishes at and has a double zero at every : both and vanish. Repeated Rolle's theorem therefore gives a point between the smallest and largest of these points such that .
Here is the zero count behind that step. A zero of multiplicity at least at means . If a function has zeros at distinct points with multiplicity lower bounds , its derivative has a zero of multiplicity at least at each point for which , and a zero in each of the intervening open intervals. Thus the total number of zeros counted with multiplicity decreases by at most one. Starting with zeros of and applying this argument times gives the required zero of . The double zeros lie in the interior, and the extra point may be an endpoint; Rolle's theorem applies on the intervening closed intervals.
Since and is monic of degree ,
At a node, both sides of the desired bounds on are zero, so no choice of is needed there.
Although the proof uses , the quadrature algorithm still uses only . The interpolant is an analytical device. Since and the rule is exact on ,
The lemma now finishes the proof of the theorem.
Proof
(Integrating the pointwise sandwich)
The point can vary with . We do not replace all these points by one point inside the integral. Instead, and give, for every ,
Multiply by the positive weight and integrate to obtain . Because , the quotient lies between the minimum and maximum of the continuous function . The intermediate value theorem supplies a point attaining that quotient. This last step proves the single-point remainder formula without requiring a measurable or continuous choice of .
Proof X-ray
Where the squared node polynomial does the work
Three hypotheses are spent at three different places. Exactness through degree is used once, on . Matching derivatives as well as values is what lifts the interpolant from degree to degree without adding samples, because involves only values. Positivity of is used when the pointwise sandwich is integrated; a sign-changing weight would destroy both inequalities. Continuity of is used to make and exist and, at the end, for the intermediate value theorem.
The key move is that the remainder carries rather than . Because this factor never changes sign, the unknown values can be trapped between and under the integral even though moves with . With ordinary interpolation the factor changes sign at every node and the same sandwich fails. If the final intermediate-value step is omitted, the two-sided bound survives, and that bound is all a numerical certificate uses.
The derivative bound is sharp for this fixed rule. Take . Its th derivative is identically one, its quadrature value is zero, and its integral is . Thus the constant multiplying cannot be decreased uniformly over this function class.
Concept lensStructural
The normalization changes the constant
If an orthogonal polynomial is written as with leading coefficient , its roots are unchanged, but its squared norm is . The error constant is therefore
Dropping gives the wrong constant unless the polynomial is monic. For example, and have the same nodes but different norms.
An exact two-node calculation
Worked example 6.4.3
(A sixth-degree integrand and an unknown remainder point)
For on , the two-node rule is
Direct integration gives
Take . Its integral and quadrature value are and , respectively. Hence
Since , the general bound gives . This bound is valid but much larger than the actual error. The single-point formula is consistent with the exact result:
There are two possible points with that square. The theorem never asserted uniqueness. Substituting the midpoint instead would wrongly predict zero error. Knowing that some point exists does not allow us to choose a convenient one.
Worked example 6.4.4
(Certified sign and size for )
For on the same interval, . Thus
This certifies the direction of the error as well as an upper bound. Computing the exact expression is a separate way to inspect this particular example.
Numerically, and , so the error is about . It lies between and , as certified, and the remainder formula is satisfied by , that is .
Rescaling and composite two-node rules
Suppose now that we integrate an unweighted function on a panel of width and midpoint . Set and . The transformed two-node rule is
The chain rule gives . The Jacobian contributes another factor , so for ,
The power is five for one panel. It includes four powers from differentiating the transformed function and one from transforming the integral.
Partition into equal panels of width , and sum their rules to obtain . Adding absolute bounds proves
For a fixed function on a fixed interval, this is an bound with samples. The decrease from the fifth power to the fourth occurs because there are panels. We have not asserted a fifth-order global bound or a relative error estimate near a zero integral.
Worked example 6.4.5
(An exact composite error for )
On with , the fourth derivative is the constant . Every local error formula is exact with this same derivative, giving
For one panel the integral is and the rule gives , whose difference is . For two panels the error is , exactly one sixteenth of the one-panel error.
Where the hypotheses matter
Counterexample mode
An almost-everywhere zero derivative is not enough
Take on . Its fourth derivative is zero at every nonzero point where that derivative is evaluated classically, but is not even differentiable at zero. The two-node rule gives
Treating the almost-everywhere value zero as would falsely predict exactness; the theorem does not apply because . The repeated Rolle argument cannot pass through the corner. Splitting the integral at zero repairs this example: on each half the function is linear, so a two-node Gaussian rule is exact there. This observation does not establish a general singular-integrand error theorem.
A derivative certificate also differs from a sample-based error estimate. The displayed Gaussian sum cannot itself certify the size of an unsampled derivative. If sampled values satisfy independently justified bounds , positivity gives
The first term is quadrature error and the second is sample uncertainty. This statement still assumes exact nodes, weights and accumulation; it does not certify a floating-point implementation. Endpoint singularities, infinite intervals and signed weights require other hypotheses and arguments.
Common mistake
(Dropping the Jacobian when rescaling)
On the error of the mapped rule is times the reference error of , and , so the constant is . Keeping only the chain-rule factor gives . For on that predicts , twice the actual error .
Common mistake
(Reading the local fifth power as the global order)
Each panel error is proportional to , but there are panels, so the global bound is proportional to . A claimed global bound is false: for on with it gives , while the actual error is .
Exercises
Quick check
(For the one-node rule on with , find from . Apply the result to and decide whether the bound is attained.)
Compute directly and divide by .
Solution · Answer
Direct integration gives , so . The rule is the midpoint rule . For , . Since everywhere, the derivative bound is attained.
Quick check
(Starting from the two-node calculation on , find the error constant when using the conventional Legendre polynomial . Explain the role of its leading coefficient.)
Use .
Solution · Answer
The leading coefficient is and . Therefore , agreeing with the monic calculation. Using alone would miss division by .
Quick check
(You want the composite two-node rule for to have absolute quadrature error at most . Use to give a sufficient integer . Then verify whether the preceding integer is certified by this same conservative bound.)
The composite bound is .
Solution · Answer
Since , the error is less than . It suffices that . For , , so six panels, using twelve samples, suffice. For , , so this chosen bound does not certify five panels. Even using itself, exceeds the tolerance. Failure of these bounds to certify five panels is not a proof that five panels actually fail the tolerance.
Quick check
(Let be any polynomial of degree at most and let , with real . Find the exact error. Explain why no upper bound on the error follows just from the sampled values, even though every such has the same samples as .)
Use exactness on and .
Solution · Answer
Exactness gives and , so . Also , recovering the remainder formula exactly. At every node , independently of . Letting increase leaves the samples fixed and makes the error unbounded. A derivative bound rules out that freedom by bounding .