Why choose the nodes?
An interpolatory rule fixes nodes first and integrates the polynomial interpolant. Gaussian quadrature chooses the nodes as well, using that freedom to integrate a larger polynomial space. The earlier unweighted construction on is one instance. Here we allow a nonuniform weight: a large value of says that errors near matter more in the integral. We will see exactly where positivity is used, and why an optimal degree of precision is not a universal error guarantee.
Let be finite and let be continuous and strictly positive on . Write
All polynomials considered have real coefficients. A nonzero polynomial has finitely many zeros, so is continuous, nonnegative and not identically zero; hence . This makes the bilinear form an inner product on every finite-dimensional polynomial space. More general measures are possible, but these assumptions suffice for the construction and proofs below.
Orthogonality determines the nodes
Theorem 6.3.1
(Weighted Gaussian construction)
For every integer there is a unique monic polynomial of degree orthogonal to every polynomial of degree at most . It has distinct roots in . Let be the Lagrange basis polynomial at these roots, and set
Then is exact on every polynomial of degree at most , all are strictly positive, and . No rule using real nodes can be exact on all polynomials of degree at most .
Proof
(Projection, sign changes and division)
Existence and uniqueness. Orthogonally project onto and subtract that projection. The result is monic and orthogonal to . If two such polynomials existed, their difference would lie in and be orthogonal to that space, including itself. Positive definiteness forces .
Why are the roots suitable nodes? Suppose has only distinct sign-changing roots in . Define , or if there are none. The product has a constant sign on the interval, apart from its zeros: each interior sign change has been canceled. It is nonzero on an open subinterval. Since , its weighted integral is nonzero, contradicting orthogonality because . Hence there are at least sign-changing roots. A degree- polynomial has at most roots, so these are all of them, and they are distinct and interior.
Take . Polynomial division by the monic gives with . The first term has weighted integral zero by orthogonality and has quadrature value zero because every node is a root of . The remainder is its own Lagrange interpolant, , so its integral equals . Thus .
Apply this conclusion to , of degree . At the nodes , giving
For arbitrary real nodes , the polynomial has degree , vanishes at all nodes, and has positive weighted integral. Hence no choice of weights can be exact on all of . The proof permits repeated nodes in this impossibility statement.
Proof X-ray
What strict positivity of the weight pays for
Positivity of is used at three separate points. In the uniqueness step it makes force . In the root count it makes the weighted integral of the nonzero, sign-constant polynomial nonzero; this is what places all roots in and makes them distinct, so that Lagrange interpolation at the nodes is possible at all. In the weight formula it makes strictly positive.
The degree bookkeeping is the other half of the argument. Division by helps for precisely because the quotient then has degree at most , the range on which orthogonality is known. For the quotient has degree and the argument stops, and the squared node polynomial shows that this stop is unavoidable. Since has degree , inside the exact range, positive weights come out as a consequence of the doubled precision rather than from a separate calculation.
There is also a converse useful for checking a proposed rule. If an -node rule with distinct real nodes is exact through degree , its monic node polynomial satisfies for every . Uniqueness therefore forces . Exactness on each then forces the corresponding weight to equal , so the whole rule is unique under these assumptions.
Concept lensStructural
The role of the weight
The weight defines both the integral and the inner product. Changing it generally changes the orthogonal polynomial, its roots and the quadrature weights. Reusing unweighted Legendre nodes without transforming the problem is not the same construction. Strict positivity guarantees the sign-change proof and positive quadrature weights used here.
A nonsymmetric two-node example
Worked example 6.3.2
(The weight on )
Here . Set . Orthogonality to and gives
Solving yields , and
Both nodes are interior, as the theorem predicts. Let correspond to . From exactness for and ,
so . The plus sign in the weight goes with . Although the nodes and weights are not symmetric about , the rule integrates every polynomial through degree three exactly.
To see the first failure, note that and are both monic of degree four, so has degree at most three. Exactness for their difference and give the first equality below; direct integration gives the numerical value:
Here , and the rule gives . This is an exact polynomial error calculation, not a general remainder formula.
Worked example 6.3.3
(One node for the same weight)
For write . Orthogonality to gives , so
the weighted mean of . The weight is , so . It reproduces . For the integral is and the rule gives , an error of . Since has degree one, this error equals , as in the two-node case. The node lies to the right of because the weight is larger there.
Worked example 6.3.4
(Edge case: endpoint nodes cannot be Gaussian)
Try nodes and for the same weight. Exactness on and forces and , so . The node polynomial is , which vanishes at both nodes, so for every choice of weights. But
Hence the rule is wrong on a quadratic, whatever its weights, and its degree of precision is one. This is the converse above in action: exactness through degree three would force , whose roots are interior.
What positivity does and does not buy
If sample values carry errors , then positivity gives . This bounds perturbations of an exact weighted sum, not floating-point arithmetic in computing the nodes, weights, or sum.
There is also a useful approximation bound. For any continuous and , exactness, positive weights and interior nodes imply
Thus one may take the infimum over such . For continuous on the fixed compact interval, the Weierstrass approximation theorem shows that this infimum tends to zero as tends to infinity, and hence . It gives no numerical error certificate until a particular polynomial and a valid uniform approximation bound are supplied. No derivative-based Gaussian remainder is asserted here.
Counterexample mode
Exactness does not see an unsampled feature
For any fixed Gaussian nodes, take . It is smooth and nonnegative, , and . Even a rule with maximal polynomial precision can miss a smooth nonzero function. For a fixed rule, multiplying by any positive constant makes this error arbitrarily large. A small error claim needs information about the particular integrand, not the degree of precision alone.
Common mistake
(Reusing Legendre nodes for a weighted integral)
For on one might take the mapped Legendre nodes and fit the weights to and , which gives . This rule happens to integrate exactly, giving , but on it gives instead of , an error of . Its degree of precision is two, not three. By the converse uniqueness statement, only the roots of the weighted give degree three.
Exercises
Quick check
(For on , construct the one-node Gaussian rule from orthogonality. Check it on .)
Orthogonality to the constant fixes the root.
Solution · Answer
Write . Orthogonality to gives . Its weight is , so . On it gives , while the integral is : the error is .
Quick check
(In the weighted example, show that errors in the two sampled values, each bounded by in absolute value, change the exact weighted sum by at most . Does this bound include quadrature error?)
The weights are positive and sum to .
Solution · Answer
The weights are positive and sum to . The triangle inequality gives . It bounds only the effect of sample errors on this weighted sum; quadrature and arithmetic errors are separate.
Quick check
(For on and nodes , use the approximation bound with to bound the error for . Taylor's theorem gives on this interval.)
The polynomial has degree at most .
Solution · Answer
Here , and has degree at most three, so . The bound is conservative; it illustrates how a justified approximation estimate becomes a rigorous quadrature bound.