Motivation
A value tells us where a curve passes; a derivative also tells us how it passes through that point. If a trajectory must arrive with a prescribed velocity, matching position alone does not solve the problem. Hermite interpolation treats these derivative conditions as genuine data. The degree count must include them, and a repeated node means additional derivative information, not another copy of the same equation.
A long table of values raises a different design question. Instead of increasing one polynomial's degree across the entire interval, we can use cubics on short intervals and join them smoothly. The joins impose relations between neighboring pieces, so choosing each cubic independently is insufficient. Natural cubic splines close the remaining freedom by setting the endpoint curvatures to zero. We will derive the resulting tridiagonal system and show that the curve minimizes the integral of squared second derivative among all twice continuously differentiable interpolants.
The earlier distinct-node interpolation notes are prerequisites.
Definitions
Definition 5.2.1
(Hermite interpolating polynomial)
Let be distinct real nodes. At node , prescribe derivatives of orders , where the multiplicities are integers. Put
A Hermite interpolating polynomial has degree at most and satisfies . The conditions include every derivative order from zero through at each node. More general lists that omit intermediate derivatives require a different existence analysis. The data can be arbitrary real numbers; an underlying smooth function is needed only when asking for an approximation remainder.
Repeated-node divided-difference notation records derivatives through
It does not instruct us to divide by a zero node separation. The derivative value is supplied. For a polynomial, the identity is also its Taylor coefficient at ; the usual distinct-node formula cannot simply be applied to a zero denominator.
Definition 5.2.2
(Natural cubic spline and bending functional)
For the spline construction use ordered knots , with integer , values , interval lengths and secant slopes
A cubic spline is in and restricts to a polynomial of degree at most three on each . It interpolates when . It is natural when . Natural conditions do not mean zero endpoint slopes. Write and define the bending functional
This penalizes curvature measured by the second derivative in these coordinates; it is not the arc-length integral of geometric curvature. We will minimize it in the explicitly stated class and need no unannounced weak-derivative or Sobolev-space theorem.
Theorem/Proposition
Theorem 5.2.3
(Hermite existence, uniqueness and the multiplicity remainder)
The Hermite conditions determine exactly one polynomial of degree at most . Suppose additionally that and , where the closed interval contains the nodes and the evaluation point . If is not a node, some in the interval spanned by the nodes and satisfies
At a node the error is zero. In particular, gives . A finite data table alone does not certify that derivative bound.
Theorem 5.2.4
(Natural spline construction and least bending)
For every strictly increasing knot sequence with and every real value vector, there is exactly one natural cubic interpolating spline. Its curvatures satisfy and
The interior system is nonsingular. For every with , even when does not satisfy natural endpoint conditions,
Equality holds exactly when . Thus the spline is the unique minimizer in this full class of interpolants, rather than just among piecewise cubics.
Concept lensStructural
Three different ways to choose the remaining freedom
Hermite interpolation uses supplied derivatives to fix a polynomial. Cubic spline joining links neighboring pieces, while boundary conditions fix the remaining global freedom. The minimum-bending theorem then explains why the natural boundary choice is variationally distinguished when only values are prescribed. This explanation neither supplies missing physical slopes nor promises shape preservation or a complete spline error theory.
Proof sketch or proof idea
Count derivative conditions as zeros with multiplicity
Proof
(Divisibility gives uniqueness, dimension gives existence)
A polynomial with for is divisible by . This follows by expanding its finite Taylor polynomial at . Distinct factors are relatively prime, so their product divides . Two candidate Hermite polynomials differ by such a of degree at most ; divisibility by a degree- polynomial forces .
For existence, map the -dimensional space of polynomials of degree at most to its requested derivative values. The map is linear, and the argument just given shows its kernel is zero. An injective linear map between spaces of the same finite dimension is onto, so every prescribed vector of data is attained. Merely counting equations would not establish existence without this independence argument.
Proof
(Counted Rolle and the remainder)
Here is the counted version of Rolle's theorem needed for the remainder. Suppose a sufficiently differentiable function has zeros at distinct ordered points with assigned multiplicities . Its derivative inherits multiplicity at any point where , and Rolle supplies a further zero in each of the open gaps. These are distinct from the inherited zeros. Thus a total of zeros counted with multiplicity yields at least such zeros for the derivative. Repeating this observation under the required differentiability is legitimate even when a node is an endpoint, using one-sided endpoint derivative values.
Fix a nonnode and set . The function has the prescribed zeros of total multiplicity and one additional zero at . Apply the counted Rolle argument times. At a resulting point ,
because and is monic of degree . This proves the formula. The point can depend on ; it is not a known differentiable function of , so differentiating the remainder as though were a constant is invalid.
Endpoint values and slopes give a concrete cubic basis
For two endpoints , put and . Define
Given values and slopes with respect to , the interpolant is
At , the value basis functions have the requested values and zero -derivatives; and have zero endpoint values and the respective unit -derivative. Since , the factors convert supplied slopes to the correct units. These checks and uniqueness prove the formula.
If the data come from , its remainder is
Indeed, writing gives . This bound is attained for : all four Hermite data vanish, and . The claim concerns exact endpoint derivative data, not derivative estimates from noisy values.
Integrate linear curvature to construct each spline piece
Proof
(Piece formula and slope matching)
The second derivative of a cubic is affine. If its endpoint second derivatives are , integrating twice and fitting the two endpoint values gives
Direct substitution gives , and the two required curvatures. These four conditions determine a cubic uniquely: their homogeneous difference has affine second derivative zero at both endpoints, hence is affine itself, and its two zero values then force it to vanish.
The endpoint slopes of this piece are
Equate with . Rearrangement gives precisely the displayed tridiagonal equations. Values already join continuously; second derivatives already agree at ; these equations make first derivatives agree as well. Thus the assembled function is genuinely . Its third derivatives may jump, which is allowed.
The nonuniform system has a unique solution
Proof
(Positive definiteness and nonvanishing pivots)
For , there are no interior unknowns: , and the spline is the line through the two values. Assume . The interior coefficient matrix is symmetric, with diagonal and neighboring entries . Extend an interior vector by . Its quadratic form is
for nonzero interior , since and every is positive. Hence is positive definite and nonsingular. The solved curvatures produce an existing spline by the piece formula. Conversely, every natural interpolating spline must obey these same equations and the same piece formula, proving uniqueness.
A direct exact-arithmetic construction costs a linear number of operations. Set and . Start and ; for , eliminate the preceding unknown by
Then and backward for . No pivot vanishes: , and implies . This establishes the divisions in exact arithmetic. A floating-point implementation still needs data-scaling and rounding analysis; the recurrence alone is not a universal accuracy guarantee.
Proof X-ray
Counting equations against unknowns
Count first: there are interior curvatures and one slope-matching equation at each of the interior knots, while the natural conditions remove the two endpoint unknowns, so the system is square. The slope equations are the only place where continuity enters; values and second derivatives agree by the construction of the pieces. Positivity of every is used twice, in the quadratic form and in the pivot bound , so no equal-spacing assumption is hidden. A repeated knot with would make the piece formula undefined, and dropping the natural conditions would leave two free parameters, as the one-interval counterexample below shows.
Minimum bending follows from an orthogonal cross term
Proof
(Integration by parts against the error)
Let be any interpolant and put , so . On one interval, integrate by parts twice, using the one-sided values of the piece's third derivative:
The remaining integral contains and therefore vanishes. This calculation requires only and a polynomial , not a third derivative of . The terms involving vanish because is zero at every knot. At an interior knot, the terms from adjacent intervals cancel because both and are continuous. The remaining endpoint terms vanish because . Thus .
Expanding now gives . If equality holds, the continuous nonnegative function has integral zero, so everywhere. Then is affine; its zero values at the two distinct endpoints force . Notice that the competitor's endpoint curvatures and slopes were never restricted. Natural conditions belong to the minimizing spline and remove the endpoint cross terms; imposing them on every competitor would prove a weaker statement.
Proof X-ray
Where each hypothesis removes a boundary term
The real idea is that the error is orthogonal to in the integral inner product, so splits like a Pythagorean sum. Each hypothesis removes one kind of boundary term: kills every term although jumps at the knots; continuity of and cancels the interior terms; kills the two outer ones. Without natural conditions the cross term is . If , take vanishing at the knots with ; then drops below for small of suitable sign. Finally, and the two endpoint zeros close uniqueness.
Worked examples
Worked example 5.2.5
(Four endpoint data for a quartic function)
For on , the endpoint data are , and . The Hermite cubic is
At , and . Since , the uniform remainder bound is , attained at that midpoint. The cubic even takes the negative value although is nonnegative. Matching values and slopes does not automatically preserve the sign between them.
Worked example 5.2.6
(A complete natural spline on three nonuniform knots)
Use knots and values . Here , , and . The single interior equation is , so with . Substitution gives
Both values at the join are one. Both first derivatives there are , and both second derivatives are . The endpoint second derivatives vanish. In contrast, and , confirming that natural does not mean horizontal endpoints. The spline is not one global cubic: its third derivatives are and on the two intervals.
Worked example 5.2.7
(Compare bending with a global interpolating quadratic)
For the same data the quadratic also interpolates, with and . The spline has on and on , hence
The variational identity therefore gives . The minimizing curve has less squared-second-derivative cost than this smoother global polynomial. Minimum bending compares this specific functional, not polynomial degree or the number of derivatives that exist.
Common mistakes
Counterexample mode
Too few boundary conditions give a family; incompatible extras give no solution
With just one interval and endpoint values zero, every cubic interpolates. There are two free parameters; no interior joining condition exists to remove them. Natural endpoint conditions select zero.
For endpoint values zero and one instead, natural conditions force to be the affine function zero at both ends, hence and . Its endpoint slopes are both one. Adding the two extra requirements is inconsistent. The cubic matching those clamped slopes is , whose endpoint second derivatives are and . Natural and clamped conditions describe different interpolation problems; imposing both need not be possible.
Even with consistent boundary conditions, the natural spline answers only the question it was built for.
Counterexample mode
Natural splines do not inherit the endpoint-Hermite error bound or preserve monotonicity
For on one interval , , the natural spline is the line . Its maximum absolute error is , although . The cubic Hermite remainder with a fourth derivative cannot apply: this line does not interpolate the true endpoint slopes.
Likewise, for knots and nondecreasing data , the natural spline has . Its first piece is , negative for , and equal to at . Minimum bending is not a nonnegativity or monotonicity constraint. Shape-preserving interpolation requires additional design choices.
Common mistake
(Treating an endpoint derivative as recoverable data)
An endpoint derivative is independent data, not information recoverable exactly from two nearby values without further assumptions. Replacing prescribed slopes by difference quotients changes the problem and its error analysis.
Common mistake
(Repeating a node to reconcile inconsistent values)
Repeated-node interpolation does not average inconsistent copies of a value: two contradictory conditions at the same derivative order are simply incompatible.
Common mistake
(Expecting an interpolating spline to remove noise)
Cubic splines interpolate the samples exactly, so they are not smoothing splines or a method of automatically removing measurement noise. If the samples carry errors, the spline reproduces them at the knots.
Summary
Hermite interpolation counts derivatives as independent conditions and zeros with multiplicity. Counted Rolle gives its remainder under an explicit global smoothness hypothesis. Natural cubic splines instead determine curvatures from a positive definite tridiagonal system on arbitrary positive knot lengths. The resulting curve uniquely minimizes integrated squared second derivative among all interpolants. Boundary choices, exact derivative data, approximation error and shape constraints remain separate questions.
Exercises
Quick check
(Prescribe , and . Determine the Hermite polynomial and node product. If the data come from a function with and , bound the error at . Explain why repeating the value condition at zero would not replace its derivative condition.)
Write and impose the three conditions.
Solution · Answer
There are conditions with multiplicities two at zero and one at one. Writing gives , , , hence . The product is . The remainder bound at is . Repeating adds the same linear equation again; it does not constrain and leaves a family of quadratic interpolants.
Quick check
(On , prescribe endpoint values one and three, and slopes zero and one, respectively. Write the Hermite cubic in , compute its value at three, and check its -derivative at both endpoints. If , give the uniform remainder bound under the required smoothness assumption.)
Use and keep the factor on the slope terms.
Solution · Answer
The interval length is two, so . At , and . The derivative is , giving zero at and one at . For data from , the uniform bound is . Omitting the interval-length multiplier on the right endpoint slope would violate the derivative check.
Quick check
(Construct the natural spline through . Find the interior curvature, both cubic pieces, their common slope at the join and the bending functional. Compare its bending with the quadratic .)
With the interior equation is .
Solution · Answer
Here , and . Thus and . The pieces are on and on . They have value zero, slope zero and curvature three at the join, and zero curvature at the external endpoints. Their second derivatives are and , so . The quadratic has second derivative two, hence bending eight. Its cost exceeds the spline's by two, exactly .
Quick check
(For the spline through in the worked example, let with real . Show that these are admissible competitors and compute exactly. Must their endpoint second derivatives vanish for the minimization theorem to apply?)
The added polynomial vanishes at every knot.
Solution
The polynomial vanishes at all three knots, so every is a interpolant. Since ,
The cross term with vanishes by the proved identity, so , uniquely smallest at zero. The competitor endpoint second derivatives are and , which generally do not vanish. They are still allowed: only the minimizing spline needs the natural endpoint conditions in the cross-term argument.