Interpolation is the construction of a function that takes prescribed values at specified points. Given observations , an interpolating function satisfies at every data point. It can then supply values at other arguments, usually within the region covered by the observations. Interpolation is fundamental to numerical analysis, scientific computing, geometric modeling, and the reconstruction of sampled signals.
The data alone rarely determine a unique function: infinitely many curves can pass through a finite collection of points. An interpolation method therefore also specifies a class of admissible functions, such as polynomials of bounded degree, piecewise polynomials, or combinations of radial kernels. Its usefulness depends on how that class reflects the underlying problem, how the sample points are distributed, and how accurately the resulting function can be computed.
Interpolation differs from fitting methods that permit discrepancies at the observed points, and from extrapolation, which evaluates a model outside the region constrained by the data. Exact agreement at the samples is an algebraic property, not a guarantee of accuracy between them. The central questions concern existence and uniqueness, approximation error, sensitivity to data, computational cost, and the preservation of relevant features such as monotonicity or periodicity.
Mathematical formulation and uniqueness
For distinct real nodes and prescribed values , the basic interpolation problem is to find a function in a chosen space such that
The nodes need not be equally spaced, and the values may be real, complex, or vector-valued. In several variables, each node is a point in a region rather than a position on a line. More general interpolation conditions can prescribe derivatives, averages, or other linear measurements instead of function values alone.
When is a finite-dimensional vector space with basis , write
The interpolation conditions become a system of linear equations,
There is a unique interpolant for every data vector precisely when the evaluation matrix is invertible. This property of the function space and sampling conditions is called unisolvency. Having the same number of coefficients and conditions is necessary in this square formulation but does not, by itself, establish uniqueness.
For example, the space of polynomials of degree at most has dimension , and distinct one-dimensional nodes are unisolvent for this space. If two such polynomials interpolate the same data, their difference has distinct zeros while having degree at most ; it must therefore be identically zero. Conversely, unrestricted continuous functions are not uniquely determined by finitely many values. If interpolates the data, then
does so as well for every continuous function .
Interpolation is thus a constrained form of function approximation. When the data come from an unknown function , the aim is usually to make close to throughout a domain, not merely at the nodes. Exactness at the nodes supplies constraints; smoothness, locality, or minimum-energy principles provide additional structure.
A simple example shows why the choice of space matters. The points and determine the linear interpolant . They also lie on for every constant . At , these curves take the values , which can be arbitrarily large or small. No bound on the intermediate value follows from the two samples without further assumptions.
Polynomial interpolation and derivative constraints
For distinct nodes, the interpolating polynomial of degree at most can be written in Lagrange form:
Each cardinal polynomial equals one at and zero at every other node. The formula therefore establishes existence directly. It also shows that interpolation is linear in the data: interpolating a linear combination of two data sets gives the same linear combination of their interpolants.
The Newton form organizes the same polynomial differently:
Here , and the divided differences are defined recursively by
This representation is particularly convenient when nodes are added sequentially: the earlier terms remain unchanged and one new term extends the interpolant. Nested multiplication evaluates the polynomial without expanding it into powers of .
For equally spaced nodes , Newton’s formula can instead use forward differences. If , then
This connection with the finite difference method was central to interpolation from printed numerical tables. It also explains why interpolating polynomials provide formulas for numerical differentiation and integration.
Hermite interpolation supplements values with derivatives. If and its first derivative are specified at each of distinct nodes, there is a unique polynomial of degree at most satisfying the conditions. More generally, prescribed derivatives through order at node give conditions and determine a polynomial of degree below that total. Uniqueness follows by counting zeros with multiplicity.
A Taylor polynomial is a special case in which all derivative conditions occur at a single point. Hermite interpolation distributes such information among several points. On one interval, values and first derivatives at both endpoints determine a cubic polynomial. This cubic Hermite construction is important in trajectory design, numerical solutions of differential equations, and piecewise curves where both position and tangent information must be respected.
Approximation error and the placement of nodes
Suppose has continuous derivatives on an interval containing the nodes and the evaluation point . The polynomial interpolation remainder is
for some in that interval. This formula follows from repeated applications of Rolle’s theorem, a consequence of the mean value theorem. It separates two sources of error: the behavior of the function’s derivatives and the geometry of the interpolation nodes.
The remainder does not imply that raising the degree always improves the approximation. Both the high derivative and the nodal product can grow rapidly. The classic Runge phenomenon occurs when the function
is interpolated on increasingly fine equally spaced grids in . The resulting polynomials exhibit growing oscillations near the endpoints, and their maximum errors do not tend to zero. This is a failure of the interpolation sequence itself, not merely of a particular implementation.
Node placement can substantially change the outcome. The Chebyshev polynomial , defined by
has roots
These nodes cluster near the interval endpoints. Their monic nodal polynomial is , whose maximum absolute value on is . Among monic polynomials of degree , this achieves the smallest possible maximum absolute value there. The associated endpoint-inclusive Chebyshev grids are also widely used.
For functions analytic in a sufficiently large complex neighborhood of an interval, interpolation at Chebyshev-type nodes can converge geometrically as the degree increases. The attainable rate is governed by the location of complex singularities, even when only real arguments are evaluated. For functions with limited smoothness, convergence is usually slower and is often algebraic in the polynomial degree.
A broader error estimate compares interpolation with the best polynomial approximation. If is the interpolation operator and its operator norm in the maximum norm, then
This links interpolation to approximation theory while exposing the importance of stability. The Weierstrass approximation theorem guarantees that every continuous function on a compact interval can be approximated arbitrarily well by polynomials. It does not guarantee convergence of interpolants at an arbitrary prescribed sequence of nodes.
Conditioning and reliable numerical evaluation
Interpolation involves two distinct numerical questions: how sensitive the mathematical answer is to changes in the data, and how much additional error a computational procedure introduces. These are conditioning and numerical stability, respectively. A stable implementation cannot eliminate the sensitivity of an ill-conditioned interpolation problem.
For polynomial interpolation, perturbations in the values produce
Consequently,
The expression in parentheses is the Lebesgue function; its maximum over the interval is the Lebesgue constant . It measures worst-case absolute amplification of errors in the sampled values. For equally spaced nodes it grows exponentially with degree, whereas for standard Chebyshev nodes it grows only logarithmically.
Representing the interpolant in monomials leads to a Vandermonde matrix,
Such matrices can be severely ill-conditioned, particularly at high degree or with poorly scaled nodes. Solving for monomial coefficients may therefore lose substantial accuracy in floating-point arithmetic. This does not mean that every value of the interpolating polynomial is equally inaccurate: the coefficient representation can be problematic even when evaluation in a more suitable basis is well behaved.
The barycentric formula usually avoids explicit coefficient construction. Define
Then, away from the nodes,
At a node the interpolated value is returned directly. Multiplying every weight by the same nonzero constant leaves the result unchanged, allowing useful rescaling. After weights have been computed, each evaluation takes a number of arithmetic operations proportional to the number of nodes.
Standard Chebyshev grids have simple explicit barycentric weights, avoiding potentially troublesome products of many node differences. For arbitrary nodes, weight construction requires attention to overflow, underflow, and scaling. Stability also depends on the evaluation region: a formula effective for interpolation inside the node interval can behave much worse during extrapolation.
Sensitivity to node locations is separate from sensitivity to function values. Closely spaced nodes, uncertain coordinates, and derivative estimation can amplify positional errors. Affine rescaling to a moderate interval helps numerical representation, but it does not restore information absent from the observations or remove intrinsic ill-conditioning.
Piecewise interpolation and splines
Rather than use one high-degree polynomial across an entire interval, piecewise interpolation uses a low-degree polynomial on each subinterval. Given ordered nodes , piecewise linear interpolation is
for . The weights are nonnegative and sum to one, so each interpolated value is a convex combination of its two neighboring values. It cannot overshoot their range.
The piecewise linear interpolant is continuous but generally has jumps in its first derivative. If has a bounded second derivative, then
Thus halving all interval widths reduces this bound by a factor of four. Unlike global polynomial interpolation on equally spaced nodes, refinement does not require increasing the local degree.
A spline joins polynomial pieces with specified smoothness at the knots. A cubic spline is cubic on each interval and usually has continuous first and second derivatives across interior knots. Interpolation at the nodes does not supply all the conditions needed to determine it; endpoint or periodic conditions complete the problem. Natural boundary conditions set . Clamped conditions prescribe endpoint slopes. Not-a-knot conditions require the first two pieces, and separately the last two, to be parts of single cubics.
Writing , the interior equations are
Together with suitable endpoint equations, these form a narrow-banded system that can be solved in linear time. For a sufficiently smooth function and compatible endpoint conditions, cubic spline interpolation can achieve fourth-order accuracy in the maximum norm. Endpoint mismatch can reduce that order.
Natural cubic splines have a variational characterization: among sufficiently regular functions interpolating the given values, they minimize
This measures bending rather than statistical error. Smoothness alone, however, does not ensure shape preservation. A cubic spline can overshoot positive or monotone data. Piecewise cubic Hermite methods address this by choosing nodal slopes subject to local inequalities. They generally retain continuity of the first derivative while sacrificing continuity of the second derivative, illustrating the trade-off between smoothness and preservation of observed shape.
Rational interpolation and local functional models
Rational interpolation seeks an interpolant of the form
where and are polynomials with prescribed degree bounds. Rational functions can represent poles and steep variations much more economically than polynomials. For example, , problematic for high-degree equally spaced polynomial interpolation, is itself a rational function with numerator degree zero and denominator degree two.
If and , then numerator and denominator coefficients have an overall scaling redundancy. A generic interpolation problem therefore uses value conditions. The equations
are linear in those coefficients. Nevertheless, a nonzero solution of these equations is not automatically a valid interpolant: the denominator must not vanish at a data node unless a common factor can be cancelled and the resulting value satisfies the condition. Degree defects and exceptional data complicate existence and uniqueness.
Poles are both an advantage and a hazard. They allow faithful representation of functions with nearby singularities, but unwanted poles can appear in the interpolation interval. Nearly cancelling numerator and denominator zeros can also arise, producing apparently harmless features that are numerically sensitive. A rational interpolant therefore needs assessment beyond checking its residuals at the data nodes.
Barycentric rational interpolation writes
This has the same form as barycentric polynomial interpolation, but more general choices of weights usually produce a rational function rather than a polynomial. Provided each weight is nonzero, the limiting value at node is . Floater–Hormann interpolation, introduced in 2007, blends local polynomial interpolants; for ordered distinct real nodes and its standard admissible parameters, it has no poles on the real axis.
Rational interpolation is related to, but distinct from, Padé approximation. A Padé approximant matches coefficients of a local power series, while ordinary multipoint rational interpolation matches values at distinct arguments. Both use numerator and denominator degrees to capture behavior that a polynomial may represent inefficiently.
These methods are important in frequency-response modeling and the approximation of functions with nearby complex singularities. Their flexibility comes with additional model choices: degree allocation, pole restrictions, and treatment of near cancellations. Exact interpolation alone does not resolve these choices, especially when the observations contain noise.
Periodic interpolation and sampled signals
For periodic functions, trigonometric polynomials are often more natural than ordinary polynomials. On a -periodic domain, a complex trigonometric interpolant has the form
where is a set of consecutive integer frequencies. At equally spaced nodes , the coefficients are obtained from the discrete Fourier transform:
The fast Fourier transform computes these coefficients in operations.
For odd , choosing frequencies gives a particularly symmetric representation. With even , the highest resolvable, or Nyquist, frequency requires a convention because positive and negative versions have identical samples. For real data, conjugate symmetry of the coefficients produces a real-valued interpolant when the frequency convention is handled consistently.
Uniform samples cannot distinguish every frequency. At these nodes,
for every integer . This is aliasing: different continuous oscillations produce identical sampled data. Trigonometric interpolation selects a representative from each class of indistinguishable frequencies; it cannot determine from the samples alone which higher-frequency components were originally present.
The sampling theorem gives an exact reconstruction result under a stronger assumption. If a signal has no frequencies beyond angular frequency , and the sampling interval satisfies , an idealized reconstruction is
with . Appropriate function-space assumptions specify the mode of convergence. Unlike finite-data polynomial interpolation, this result uses an infinite sequence of samples together with a global frequency restriction.
Finite records require truncation, windowing, or periodic extension. If a nonperiodic record is treated as periodic, a mismatch between its endpoint values creates an artificial discontinuity. The resulting Fourier approximation can show persistent oscillations near the join. Smooth periodic functions behave much better: sufficiently smooth functions give rapid coefficient decay, and periodic functions analytic in a strip around the real axis can yield geometric convergence.
In digital resampling, a finite interpolation filter approximates ideal sinc reconstruction. The kernel’s frequency response affects blur, ringing, and attenuation of unwanted frequencies. Downsampling also requires suppression of frequencies above the new Nyquist limit; merely interpolating the existing samples does not prevent aliasing.
Interpolation in several dimensions
Multivariate interpolation replaces scalar nodes by points . Geometry now has a central role. In one dimension, distinct nodes uniquely determine a polynomial of the appropriate degree. In several dimensions, distinctness and the correct number of points are not sufficient. The space of total-degree- polynomials in variables has dimension
but the points must also be positioned so that no nonzero polynomial in that space vanishes at all of them.
For example, an affine function is determined by its values at three noncollinear planar points. Three collinear points do not determine it uniquely, because information in the transverse direction is absent. This geometric failure is a loss of matrix rank in the evaluation system.
On rectangular grids, tensor-product interpolation applies one-dimensional rules in each coordinate. Bilinear interpolation on a rectangle is
where are normalized coordinates. It reproduces every function linear in each coordinate separately, including a term proportional to . It is therefore generally not an affine function of the pair . Bicubic interpolation uses cubic dependence in each coordinate and requires additional samples or derivative estimates.
Scattered points require different constructions. A Delaunay triangulation partitions a planar point set into triangles with an empty-circumcircle property. Linear interpolation on each triangle uses barycentric coordinates: the interpolated value is the same weighted combination of the three vertex values as the point is of the vertices. Adjacent triangles agree along shared edges, producing a continuous piecewise affine surface whose gradient generally jumps across edges. Such local shape functions are central to the finite element method.
Radial basis function interpolation instead uses
The kernel depends only on distance from a center. Some kernels yield invertible interpolation matrices for distinct sites; others require a supplementary polynomial and side constraints on the coefficients. Gaussian, multiquadric, and thin-plate spline kernels are prominent examples. Globally supported kernels often produce dense systems, while compactly supported kernels can produce sparse ones.
High-dimensional grids face the curse of dimensionality: samples in each of coordinates require points. Sparse grids, low-rank representations, and adaptive sampling reduce this cost only by exploiting additional structure. Geometry also limits interpretation: inside a simplex, barycentric interpolation is a convex combination; outside the sampled region, the same algebraic formula becomes extrapolation and loses that range-preserving property.
Noisy data, smoothing, and probabilistic interpretation
Interpolation treats the observed values as exact constraints. In many applications, however, the data follow a model such as
where represents measurement error. An exact interpolant reproduces both the underlying signal and the realized errors. Whether this is useful depends on the noise level, sampling density, function class, and intended prediction task.
Least-squares fitting instead permits residuals and chooses a function to minimize their squared sum. For an overdetermined linear model, the fitted values need not equal the observations. Interpolation and regression are therefore different constraints, not simply different names for smooth curve construction. A low-degree fitted polynomial may resist measurement noise better than a high-degree interpolating polynomial, even though its training residuals are larger.
Regularization makes the trade-off explicit. A cubic smoothing spline minimizes a criterion of the form
The first term measures agreement with the observations; the second penalizes bending. As tends to zero, the solution approaches an interpolating natural cubic spline under standard conditions. As grows, the curvature penalty increasingly favors an affine least-squares fit. The numerical scale of depends on the units and normalization used in the criterion.
A Gaussian process supplies a probabilistic counterpart. For a zero-mean process with covariance kernel , let and . With noiseless observations and invertible , the conditional mean is
which interpolates the observations. Its conditional variance is
At an observed node this variance is zero in the noiseless model. Independent observation noise with variance replaces by ; the conditional mean then generally smooths rather than interpolates.
The same kernel formula can thus be interpreted either as deterministic interpolation or as a conditional mean under a probability model. The probabilistic uncertainty remains conditional on the kernel, noise assumptions, and other model specifications; exact interpolation does not establish that those assumptions are correct.
In machine learning, models with enough flexibility can interpolate training data while displaying widely different predictive errors. Zero training error alone is therefore insufficient to determine generalization. Cross-validation evaluates predictions at observations withheld from fitting, though spatial and temporal dependence can make randomly withheld points much easier to predict than genuinely distant or future observations.
Historical development and computational uses
Interpolation developed from the practical need to use numerical tables. Astronomical and mathematical tables listed values only at discrete arguments, while calculations required intermediate ones. Linear interpolation was a natural first procedure; higher-order rules used neighboring tabulated values or differences to represent curvature. Such practices appeared in several mathematical traditions before modern function notation and approximation theory.
In the seventeenth century, James Gregory and Isaac Newton developed systematic formulas involving differences and divided differences. Edward Waring published a form of the polynomial interpolation formula in 1779, before Joseph-Louis Lagrange’s 1795 presentation of the expression now associated with his name. The alternative formulas represent the same unique polynomial, but they support different computational arrangements and reveal different mathematical properties.
During the nineteenth century, interpolation became closely connected with analysis, numerical quadrature, and polynomial approximation. Charles Hermite’s work gave a systematic setting for matching derivatives as well as values. Carl Runge’s 1901 example demonstrated that increasing the degree at equally spaced nodes could make interpolation worse, helping distinguish polynomial approximability from convergence of a particular interpolation procedure.
In the twentieth century, mechanical and electronic computation changed the preferred representations. Isaac Jacob Schoenberg introduced the term “spline” in mathematical work in 1946, drawing on the flexible drafting strips used to draw smooth curves. Piecewise polynomial methods combined local flexibility with efficient computation, while B-spline representations provided compact support and well-structured bases. Spectral methods, fast transforms, and kernel methods subsequently expanded the range of practical interpolation problems.
Interpolation also acts as a building block for other numerical operations. Integrating an interpolating polynomial produces an interpolatory quadrature rule. Integrating the linear endpoint interpolant yields the trapezoidal rule; integrating the quadratic interpolant through two endpoints and their midpoint yields Simpson’s rule. On one interval of width , Simpson’s approximation is
Differentiating interpolants similarly produces differentiation weights, although differentiation generally amplifies noise more strongly than evaluation or integration.
Modern applications impose distinct requirements. Image enlargement values locality and controls blur and ringing; geometric design requires tangent or curvature continuity; scientific table lookup balances error bounds against memory and evaluation cost. Numerical solvers for ordinary differential equations use interpolating “dense output” to estimate solutions between computed time steps, including for event detection. In mesh-based simulation, transferring values between grids must distinguish interpolation from conservation: matching point values does not necessarily preserve total mass, momentum, or energy, so conservative remapping may instead enforce agreement of cell integrals.