Bernstein polynomials

1 Definition and Construction

1.1 Basic formula on \([0,1]\)

For a nonnegative integer \(n\), the \(n\)-th Bernstein polynomial associated with a function \(f\) on \([0,1]\) is constructed through the Bernstein operator \(B_n\) defined by \[ (B_n f)(x)=\sum_{k=0}^{n} f\!\left(\frac{k}{n}\right)\binom{n}{k}x^{k}(1-x)^{n-k}, \quad x\in[0,1]. \] The functions \[ b_{n,k}(x)=\binom{n}{k}x^{k}(1-x)^{n-k}, \quad k=0,\dots,n, \] are often called Bernstein basis polynomials; they form the building blocks for \(B_n f\).

1.2 Relationship to binomial coefficients

The appearance of \(\binom{n}{k}\) is not incidental: it matches the combinatorial weights that arise from expanding \((x+(1-x))^n\). In particular, the basis elements \(b_{n,k}(x)\) scale the monomials \(x^k(1-x)^{n-k}\) so that, as \(k\) varies, the collection behaves like a probability mass function in \(k\).

1.3 Normalization, degree, and endpoint values

Each \(b_{n,k}\) is a polynomial of degree \(n\). At the endpoints, \[ b_{n,k}(0)=\begin{cases}1, & k=0,\\ 0, & k\ge1,\end{cases} \qquad b_{n,k}(1)=\begin{cases}1, & k=n,\\ 0, & k\le n-1.\end{cases} \] Consequently, \[ (B_n f)(0)=f(0),\qquad (B_n f)(1)=f(1), \] and the operator preserves the boundary values of \(f\) on \([0,1]\) regardless of \(n\).

1.4 Basis properties and partition of unity

A key feature is that the Bernstein basis forms a partition of unity: \[ \sum_{k=0}^{n} b_{n,k}(x)=1 \quad \text{for all } x\in[0,1]. \] This implies that constant functions are fixed points of the operator: \[ B_n(1)=1,\quad\text{and more generally}\quad B_n(c)=c \] for any constant \(c\). Because the coefficients \(f(k/n)\) are averaged using weights \(b_{n,k}(x)\ge 0\), many qualitative properties of the original data transfer to \(B_n f\).

2 Bernstein Approximation Theorem

2.1 Statement of the uniform convergence result

Bernstein’s theorem asserts that if \(f\) is continuous on \([0,1]\), then the sequence of polynomials \(B_n f\) converges uniformly to \(f\): \[

\] This result holds without imposing additional smoothness assumptions beyond continuity.

2.2 Approximation of continuous functions

Uniform convergence means that for every \(\varepsilon>0\), there exists \(N\) such that for all \(n\ge N\), \[

\] The operator therefore provides an explicit polynomial approximation scheme directly from the values \(f(k/n)\) at a uniform grid.

2.3 Role of smoothness (e.g., differentiability)

While continuity suffices for convergence, the *rate* of convergence improves when \(f\) has additional regularity. If \(f\) is differentiable, the approximation error typically decreases faster than in the merely continuous case. For functions with higher derivatives, classical estimates relate the error to those derivatives, reflecting how well the operator reproduces local behavior.

2.4 Probabilistic viewpoint of approximation

A common way to understand Bernstein approximation is to interpret \(B_n f(x)\) as an average of \(f\) with respect to a binomial distribution parameterized by \(x\). From this viewpoint, the theorem resembles a law-of-large-numbers phenomenon: as \(n\) grows, the binomial mass concentrates near its mean, forcing the average to align with \(f(x)\) when \(f\) is continuous.

3 Probabilistic Interpretation

3.1 Binomial random variables and expected values

Let \(X\sim\mathrm{Bin}(n,x)\), so that \[ \mathbb{P}(X=k)=\binom{n}{k}x^k(1-x)^{n-k},\quad k=0,\dots,n. \] Then \(X/n\) takes values \(k/n\) with the same weights used in the Bernstein basis.

3.2 Bernstein operator as an expectation operator

With the binomial interpretation, \[ (B_n f)(x)=\sum_{k=0}^{n} f\!\left(\frac{k}{n}\right)\mathbb{P}(X=k) = \mathbb{E}\!\left[f\!\left(\frac{X}{n}\right)\right]. \] Thus, Bernstein approximation can be viewed as replacing \(f(x)\) by the expected value of \(f\) evaluated at a random point \(X/n\) whose mean equals \(x\).

3.3 Moment computations and limiting behavior

For \(X\sim\mathrm{Bin}(n,x)\), \[ \mathbb{E}\left[\frac{X}{n}\right]=x,\qquad \mathrm{Var}\left(\frac{X}{n}\right)=\frac{x(1-x)}{n}. \] These moments quantify how quickly the random input \(X/n\) concentrates around \(x\). As \(n\to\infty\), the variance tends to \(0\), and the expectation \(\mathbb{E}[f(X/n)]\) approaches \(f(x)\) for continuous \(f\).

3.4 Concentration near the mean

Concentration results formalize that \(X/n\) is typically close to \(x\). When \(f\) is continuous, values of \(f\) near \(x\) are close to \(f(x)\), so the contribution from the concentrated region dominates the expectation. This explains both convergence and the dependence of the error on the regularity of \(f\).

4 Algebraic and Analytic Properties

4.1 Positivity and monotonicity

Since \(b_{n,k}(x)\ge 0\) and the basis sums to \(1\), the Bernstein operator is positive in the sense that if \(f\ge 0\) on \([0,1]\), then \(B_n f\ge 0\). More generally, it preserves inequalities: if \(f\le g\) pointwise, then \(B_n f\le B_n g\).

4.2 Linearity of the Bernstein operator

The operator \(B_n\) is linear in \(f\): \[ B_n(\alpha f+\beta g)=\alpha B_n f + \beta B_n g \] for scalars \(\alpha,\beta\). This follows directly from the defining sum and is useful in deriving estimates and studying how Bernstein polynomials act on function spaces.

4.3 Shape-preserving features (convexity, etc.)

Because Bernstein polynomials form convex combinations of function values, \[ (B_n f)(x)=\sum_{k=0}^n b_{n,k}(x)\, f\!\left(\frac{k}{n}\right), \] they tend to preserve several shape properties. For instance, if \(f\) is convex on \([0,1]\), then \(B_n f\) is also above \(f\) in a suitable sense: the operator acts like a smoothing average that aligns with convexity constraints. Similar statements exist for concavity and related geometric features, often expressed through Jensen-type inequalities.

4.4 Derivatives and variation-diminishing aspects

Differentiating the basis functions yields explicit formulas for \((B_n f)'(x)\) and higher derivatives in terms of finite differences of \(f(k/n)\). The operator also exhibits variation-diminishing behavior under appropriate conditions, meaning that oscillatory features can be reduced rather than amplified. This connects Bernstein approximation to broader themes in regularization and geometric monotonicity.

5 Rates of Convergence and Error Estimates

5.1 Modulus of continuity

To quantify approximation quality for general continuous \(f\), a standard tool is the modulus of continuity \[

\] Because Bernstein averaging effectively samples \(f\) at points within a typical distance on the order of \(n^{-1/2}\), the error can be bounded in terms of \(\omega(f,\cdot)\).

5.2 Error bounds via smoothness parameters

If \(f\) has additional smoothness—such as a Lipschitz condition or a bounded derivative—then one can derive sharper inequalities. For twice differentiable functions, error estimates often involve bounds on \(f''\) and scale like \(1/n\), reflecting the operator’s second-order behavior around each \(x\).

5.3 Classical inequalities for approximation error

Classical results in approximation theory provide explicit bounds for \[

\] These bounds typically combine the concentration of the binomial distribution with Taylor expansions (when available) or with inequalities that relate smoothness to approximation by positive linear operators.

5.4 Asymptotic behavior as degree increases

As \(n\) grows, the Bernstein polynomials form increasingly accurate approximations. The dominant asymptotic term depends on the smoothness class of \(f\): rougher functions yield slower convergence, while smoother functions allow faster decay of the error. The limiting behavior is therefore not uniform across all continuous functions, but it is systematically controlled by regularity measures.

6 Computational Aspects

6.1 Efficient evaluation strategies

Direct evaluation of the defining sum may cost \(O(n)\) per point \(x\), which can be heavy when many \(x\)-values are required. Practical implementations often compute the basis values \(b_{n,k}(x)\) iteratively or reuse intermediate quantities across nearby points to reduce computational overhead.

6.2 Stability considerations

Numerical stability depends on how the basis polynomials are computed. For large \(n\), terms like \(x^k(1-x)^{n-k}\) can underflow or overflow in floating-point arithmetic. Stable strategies often rely on logarithms, rescaling, or recurrence relations to keep values in a manageable range.

6.3 Using recurrence relations

The Bernstein basis satisfies useful recurrences in \(k\) and in \(n\). These relations enable computing all \(b_{n,k}(x)\) with a controlled sequence of arithmetic operations. When coupled with standard tabulation of \(f(k/n)\), they support efficient generation of \(B_n f(x)\).

6.4 Implementation in numerical software

In computational libraries, Bernstein approximation appears in contexts such as curve evaluation and smoothing. Implementations typically accept control values \(f(k/n)\) (or control points for geometric uses) and return polynomial values at sampling points. Careful choice of evaluation method—such as converting to an alternative polynomial basis when beneficial—can improve accuracy and performance.

7 Generalizations and Variants

7.1 Bernstein polynomials on general intervals

The classical definition on \([0,1]\) can be transported to an interval \([a,b]\) by a linear change of variables. One evaluates \(f\) at the rescaled grid points \(a+k(b-a)/n\) and uses corresponding basis polynomials in the normalized coordinate \((x-a)/(b-a)\).

7.2 Multivariate Bernstein polynomials

For functions of several variables, multivariate Bernstein operators are built from products of one-dimensional Bernstein bases. If \(\mathbf{x}\) lies in a multidimensional box, the weights factor across coordinates, and the approximation averages \(f\) on the associated grid. These operators are central in multivariate shape-preserving approximation and in geometric modeling.

7.3 Weighted Bernstein polynomials

Variants introduce weights that modify the binomial basis, resulting in operators tailored to non-uniform sampling behavior or to particular boundary characteristics. Weighted forms can also be used to achieve better approximation under specific assumptions about the function’s behavior near the endpoints.

7.4 Connection to other approximation operators

Bernstein polynomials are part of a broader family of positive linear approximation operators, including schemes based on other classical kernels. Many structural properties—positivity, reproduction of certain low-degree polynomials, and convergence criteria—mirror those of other operators, providing a comparative framework within approximation theory.

8 Bernstein Polynomials in Geometry and Modeling

8.1 From Bernstein polynomials to Bézier curves

Bézier curves use Bernstein basis polynomials directly. Given control points \(P_0,\dots,P_n\), the curve is defined by \[ C(t)=\sum_{k=0}^{n} P_k\, b_{n,k}(t),\quad t\in[0,1], \] where the weights \(b_{n,k}(t)\) are exactly the Bernstein basis polynomials. This links the approximation framework to geometric construction.

8.2 Control points and geometric meaning

Each control point \(P_k\) influences the curve’s shape through the corresponding basis function \(b_{n,k}(t)\). At \(t=0\), only \(P_0\) contributes; at \(t=1\), only \(P_n\) contributes. For intermediate \(t\), the curve point is a weighted average of control points, which guarantees that the curve stays within the convex hull of the control polygon.

8.3 Degree elevation and refinement

A distinctive feature of Bézier representations is that one can increase the degree without changing the geometric curve. Degree elevation uses identities among Bernstein bases to rewrite a degree-\(n\) representation in terms of a degree-\((n+1)\) basis with new control points. This supports refinement and compatibility across different modeling resolutions.

8.4 Continuity and smoothness in curve construction

When Bézier curves are joined to form composite shapes, smoothness at junctions depends on relationships among neighboring control points. Because Bernstein bases are smooth inside \((0,1)\), the differentiability properties of the curve are determined by how derivatives at endpoints match across pieces, typically expressed through control point constraints.

9 Connections to Other Areas of Mathematics

9.1 Approximation theory and operators

In approximation theory, Bernstein operators serve as canonical examples of positive linear operators that converge to the identity on continuous functions. They are frequently used to illustrate general theorems about convergence, reproduction of polynomials of low degree, and quantitative estimates driven by moduli of continuity.

9.2 Functional analysis perspective

From a functional-analytic viewpoint, \(B_n\) acts on spaces such as \(C([0,1])\) and operates as a bounded linear operator with norm control inherited from positivity and the partition of unity. Uniform convergence can be interpreted in operator-topology terms, and approximation results relate to properties like equicontinuity and density.

9.3 Total positivity and related kernels

The Bernstein basis kernels are linked to total positivity phenomena. Positivity of minors in associated kernel matrices has consequences for variation diminishing, shape preservation, and structured monotonicity. These connections place Bernstein approximation within the theory of totally positive kernels and their applications.

9.4 Links to numerical integration and quadrature ideas

Because Bernstein operators average function values against binomial weights, they resemble quadrature schemes with nodes at \(k/n\) and weights depending on the evaluation point \(x\). While not a fixed quadrature rule for a single integral, the underlying weighted averaging ties Bernstein methods to numerical approximation techniques and to broader ideas about kernels and smoothing in computation.