One polynomial, several representations
Given distinct real nodes and data , there is a unique polynomial of degree at most with . The Lagrange form makes existence and uniqueness clear, but repeatedly evaluating products for many values of can be cumbersome. Barycentric weights reorganize the same polynomial into a quotient. This representation also lets us discuss a separate issue: how strongly errors in the supplied data affect the answer.
Definition 7.9.1
(Barycentric weights and the Lagrange basis)
For distinct nodes define
The denominator in each is nonzero because the nodes are distinct. The basis satisfies .
Theorem 7.9.2
(Existence and uniqueness of the interpolating polynomial)
For distinct nodes and arbitrary data there is exactly one polynomial of degree at most with for all , namely .
Proof
(Lagrange basis and root counting)
Each has degree and , so takes the value at . If two polynomials of degree at most agree at all nodes, their difference has more roots than its degree unless it is zero; this proves uniqueness.
Deriving the quotient
Let . For different from every node,
The first barycentric formula is .
Theorem 7.9.3
(Second barycentric formula)
For different from every node,
Multiplying every weight by the same nonzero constant leaves the quotient unchanged, and the quotient tends to as .
Proof
(Dividing by the interpolant of the constant 1)
The Lagrange interpolant of the constant data is the constant polynomial, so . By the first formula applied to this data, , which gives the denominator identity. Dividing the first barycentric formula by this identity gives the quotient. A common factor in the weights cancels between numerator and denominator. Finally the quotient equals the polynomial at every non-node point, hence it has the limit as .
Proof X-ray
Why the denominator cannot vanish in exact arithmetic
The normalized weights are special: is not an arbitrary rational function. Multiplying it by gives , and this is the one place where the constant polynomial is used. Thus the denominator is exactly away from nodes, nonzero because no factor of vanishes there. This argument does not guarantee that a floating-point evaluation of the denominator has no cancellation.
This is the second barycentric formula in exact arithmetic. At , do not evaluate the displayed terms separately: the polynomial value is . Equivalently, its quotient has that limit as , because it equals the polynomial at every nearby non-node point.
Common mistake
(Plugging a node into the quotient)
At the term divides by zero, in the numerator and in the denominator. Treating the quotient as or dropping the term gives a wrong value or a floating-point exception. The correct value is ; check for a node hit before evaluating the sums.
Worked example 7.9.4
(Second formula on three nodes)
Take nodes and data , which come from . The weights are , and . At the numerator is and the denominator is . The quotient is . As a check, agrees with the denominator.
Worked example 7.9.5
(Approaching a node)
Take nodes with weights and data , again from . At the numerator is and the denominator is , so . At the middle term is undefined. Multiply numerator and denominator by and let . The numerator tends to , because the middle datum is and the other two terms stay bounded before the factor is applied. The denominator tends to . The limit is , as the theorem predicts.
Quick check
(Find barycentric weights for nodes and , up to a common nonzero factor.)
Solution · Answer
The weights are , or any common nonzero multiple such as .
How data errors are amplified
Keep the nodes and the evaluation point fixed, measure data perturbations in the maximum norm, and perturb the data to .
Definition 7.9.6
(Lebesgue function and constant)
The Lebesgue function of the nodes is . On a compact interval containing the nodes, is the Lebesgue constant for those nodes and that interval.
Proposition 7.9.7
(Amplification of data errors)
The change at a fixed is exactly linear:
Also and at a node. The bound is sharp: for the choice (any sign where ) changes the value by exactly .
Proof
(Linearity and aligned signs)
The interpolant depends linearly on the data, so exactly, not merely to first order. The triangle inequality gives the bound. Since , . At a node gives . With the stated sign choice every term equals , so they add to .
The bound controls worst-case absolute amplification, not relative error when is near zero.
Worked example 7.9.8
(Three equally spaced nodes)
For , , , one may take weights . At ,
Thus data errors of magnitude at most can change the interpolated value by at most there. The choice attains this bound. This is a property of the interpolation problem at that point, independent of which algebraically equivalent formula is used to compute it. At the node the basis values are and : there the data error is passed on unchanged and not amplified.
Quick check
(At for the three nodes , verify that the Lagrange basis values sum to one and that their absolute values sum to .)
Solution · Answer
, while .
Quick check
(If all data errors are bounded by , give the sharp worst-case absolute bound at that point.)
Solution · Answer
The bound is ; the signs attain it.
Counterexample mode
Exact interpolation does not erase inaccurate data
Claim: if a polynomial reproduces the table exactly, the table values are reliable. Take nodes with true values but a recorded middle value , so the table is . The interpolant is . At the middle node it equals the supplied value exactly, and at it equals , though the true function is everywhere. Exact agreement with the table is not evidence that the table itself is accurate.
Quick check
(Why can a small absolute data-error bound fail to provide a useful relative-error bound?)
Solution · Answer
Relative error divides by the exact value , which may be arbitrarily small or zero even when the absolute bound is small.
Approximation error is another question
Theorem 7.9.9
(Interpolation remainder)
Suppose on an interval containing the nodes and , and let interpolate the exact values . For some in that interval,
Proof
(Rolle's theorem applied times)
At a node the error is zero. Away from nodes, set and . The function has distinct zeros at the nodes and . Applying Rolle's theorem times gives for some between the smallest and largest of those points. Since and , the remainder follows.
It concerns the error of approximating the function by a polynomial with exact data; the Lebesgue bound concerns changes caused by perturbed data. Both can matter simultaneously.
Common mistake
(Using the remainder to bound data errors)
The remainder formula assumes the table contains exact values . It says nothing about errors already present in the supplied data, and the Lebesgue function says nothing about how well a polynomial approximates . A total error estimate needs both terms.
Barycentric form is often convenient and avoids explicitly solving a Vandermonde system. It does not make all node choices well conditioned, and exact algebra does not by itself prove a floating-point stability theorem. In computation, handle node hits explicitly and take care with weight scaling, overflow, underflow and cancellation.