Abstract
We construct new polynomial interpolation schemes of Taylor and Hermite types in \(\mathbb {R}^{n}\). The interpolation conditions are real parts and imaginary parts of certain differential operators. We also give formulas for the interpolation polynomials which are of Newton form and can be computed by an algorithm.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Let \(\mathcal {P}_{d}(\mathbb {R}^{n})\) be the vector space of all polynomials of degree at most d in \(\mathbb {R}^{n}\). It is well-known that the dimension of \(\mathcal {P}_{d}(\mathbb {R}^{n})\) equals \(\left (\begin {array}{cc}n+d\\n \end {array}\right )\). The Lagrange interpolation problem associated to A asks whether there is a unique polynomial p in \(\mathcal {P}_{d}(\mathbb {R}^{n})\) which matches preassigned data on A, where A consists of \(\left (\begin {array}{cc}n+d\\n \end {array}\right )\) distinct points in \(\mathbb {R}^{n}\). Let \({\mathscr{B}}=\{p_{1},\ldots ,p_{N}\}\) be a basis for \(\mathcal {P}_{d}(\mathbb {R}^{n})\) with \(N=\left (\begin {array}{cc}n+d\\n \end {array}\right )\). Then, the Vandermonde determinant defined by
is a polynomial of interpolation points. Hence, it is non-zero for almost all choices of interpolation points. In other words, a subset \(A\subset \mathbb R^{n}\) of \(\left (\begin {array}{cc}n+d\\n \end {array}\right )\) distinct points is regular for almost all choices of A. On the other hand, it is difficult to check whether a given set of \(\left (\begin {array}{cc}n+d\\n \end {array}\right )\) points is regular as soon as n ≥ 2. Explicit interpolation schemes are available in literatures. Chung and Yao [5] gave a quasi-constructive description of locations of nodes in \(\mathbb {R}^{n}\) for which Lagrange interpolation is regular. Here, the interpolation points are suitably distributed on hyperplanes and form a so-called nature lattice. Another type of regular sets was discovered by Bos [2], where the interpolation points are taken from algebraic varieties in \(\mathbb {R}^{n}\).
We consider the problem of Hermite interpolation by polynomials in \(\mathbb {R}^{n}\). More precisely, the problem is to find a polynomial which matches, on a set of distinct points in \(\mathbb {R}^{n}\), the values of a function and its partial derivatives. We deal with the case where the number of interpolation conditions is equal to the dimension of \(\mathcal {P}_{d}(\mathbb {R}^{n})\). If the interpolation problem has a unique solution, then we say that the interpolation scheme is regular.
Roughly speaking, Hermite interpolation of type total degree defined below is the most natural generalization of univariate Hermite interpolation.
Problem 1 (Lorentz [8, 9])
Let m be a positive integer. Let d,d1,…,dm be natural numbers such that
Let {b1,…,bm} be a set of distinct points in \(\mathbb {R}^{n}\). Find conditions for which the interpolation problem
has a unique solution for given values cj,α.
Many regular (resp., almost regular, singular) interpolation schemes of type total degree can be found in literature. For instance, it was shown in [8] that the only multivariate interpolation of type total degree regular for all choice of nodes is Taylor interpolation. Also in [8], it was proved that the Hermite interpolation is singular if the number of nodes satisfies 2 ≤ m ≤ n + 1 with n ≥ 2 except for the case of Lagrange interpolation. Some interesting results focusing on almost regular interpolation schemes were given in [16].
Another general type of Hermite interpolation was defined in [16], which replaces derivatives in the coordinate directions by directional derivatives. More precisely, interpolation conditions at the point \(\mathbf {a}\in \mathbb {R}^{n}\) are given by chains of directional derivatives of consecutive order
In [16], the authors introduced the notations of trees and the blockwise structure to defined general Hermite interpolation problems. Using the Bézier representation of polynomials in barycentric coordinates, they proved necessary and sufficient conditions for almost regular interpolation problems. Moreover, they established the Newton formula and the remainder formula for the Hermite interpolation polynomials. For details, we refer the reader to [16].
Some explicit multivariate Hermite interpolation schemes in \(\mathbb {R}^{n}\) were constructed. In [1], the authors gave bivariate regular schemes, where the nodes are equidistant points on concentric circles and the derivatives are the normal derivatives at these points. In [3], Bos and Calvi constructed new regular interpolation schemes of Hermite type in \(\mathbb {C}^{n}\) and \(\mathbb {R}^{n}\). Here, the interpolation points are distributed on algebraic hypersurfaces and the discrete differential conditions come from certain least spaces of finite-dimensional spaces of analytic functions. For a recent account of the theory of Hermite interpolation, we refer the readers to [4, 9].
We now state another general problem. Associated with a polynomial \(Q(\mathbf {x})={\sum }_{\alpha } c_{\alpha }\mathbf {x}^{\alpha }\), \(c_{\alpha }\in \mathbb {R}\) in \(\mathbb {R}^{n}\), we define a differential operator P(D) by
In the case when Q(x) = c, we set Q(D)f = cf.
Problem 2
Let A = {a1,…,am} be m distinct points in \(\mathbb {R}^{n}\). Let n1,…,nm and d be positive integers such that \(n_{1}+n_{2}+\cdots +n_{m}=\left (\begin {array}{cc}n+d\\n \end {array}\right )\). Find differential operators Pjk(D) for j = 1,…,m and k = 0,…,nj − 1 for which the interpolation problem
has a unique solution \(P\in \mathcal {P}_{d}(\mathbb {R}^{n})\) for any given preassigned data {fjk}.
When \(m=\left (\begin {array}{cc}n+d\\n \end {array}\right )\) and all the differential operators are point-evaluation functionals, Problem 2 becomes a Lagrange interpolation problem.
Some special cases of Problem 2 were recently studied by the first author of this paper. In [11], we gave a solution of the problem. The differential operators are the real parts and imaginary parts of the complex derivatives \(\left (\frac {\partial }{\partial x_{1}}-i\frac {\partial }{\partial x_{2}}\right )^{k}\), k ≥ 0. In [12], we considered an analogous problem on the space of bivariate symmetric polynomials. It was showed that the differential operators can be taken as \(\left (-\beta \frac {\partial }{\partial x_{1}}+\alpha \frac {\partial }{\partial x_{2}}\right )^{k}\) with \(\alpha ,\beta \in \mathbb {R}\) and k ≥ 0. We showed in [13] a way to mix two types of differential operators mentioned above to get solutions of Problem 2. Our method relies on factorization results for polynomials and the construction of certain Taylor type polynomials. Moreover, we give a Newton formula for the interpolation polynomial and use it to prove that the Hermite interpolation polynomial is the limit of the Lagrange interpolation polynomials.
The aim of this paper is to generalize Hermite interpolation in [11]. The regular Hermite schemes are constructed as follows. We first give a criterion such that a polynomial in \(\mathcal {P}_{d}(\mathbb {R}^{n})\) is divisible by the polynomial qa(x) := |(xm − am) + c(xj − aj)|2 with a = (a1,…,an), j ≠ m and \(c\in \mathbb {C}\setminus \mathbb {R}\). The criterion contains differential operators of the forms
We use these differential operators to construct a Taylor type polynomial at a. The formula and properties of the new Taylor polynomial are similar to that of ordinary Taylor polynomial. Next, we use the above-mentioned divisibility criterion to get a factorization result which leads to the regularity of Hermite interpolation. The formula for the Hermite interpolation polynomial is of a Newton form and is written in terms of Taylor type polynomials. This enables us to create an algorithm to compute the interpolation polynomial. We also give some examples to illustrate our results. It is worth pointing out that if we take n = 2 and c = i, we recover the theory of Hermite interpolation in [11].
Finally, we note that an analogous problem is studied on the unit sphere \(\mathbb S\) in \(\mathbb {R}^{3}\). In recent works, we constructed some regular Hermite schemes on \(\mathbb S\). For more details, we refer the readers to [13,14,15].
Notations and conventions
We use bold symbols x,a, etc. to denote points in \(\mathbb {R}^{n}\). We always assume that n ≥ 2, 1 ≤ j,m ≤ n with m ≠ j. The constant c is a non-real number. The multi-index in \(\mathbb {N}^{n-1}\) is denoted by \(\alpha ^{\prime }=(\alpha _{1},\ldots ,\alpha _{m-1},\alpha _{m+1},\ldots ,\alpha _{n})\) in which the m-entry does not appear. Throughout this paper, f and g are real-valued functions. All polynomials are of real coefficients except for polynomials of two types \({\varPi }_{\alpha ^{\prime }}\) and \(B_{\beta ^{\prime }}\). By a suitably defined function f, we mean that the function f is sufficiently differentiable.
2 Taylor type polynomials
2.1 A divisibility criterion
Let \(\mathbf {a}=(a_{1},\ldots ,a_{n})\in \mathbb {R}^{n}\) and \(c\in \mathbb {C}\setminus \mathbb {R}\). Let 1 ≤ j,m ≤ n and j ≠ m. We define the following polynomial
Easy computations gives
Note that qa depends not only on a but also on j,m and c. We adopt this setting for simplicity of notations. Clearly qa is an irreducible polynomial of degree 2 in \(\mathbb {R}^{n}\) and
which is a flat of dimension n − 2 in \(\mathbb {R}^{n}\).
Theorem 1
Let \(\mathbf {a}=(a_{1},\ldots ,a_{n})\in \mathbb {R}^{n}\) and qa(x) = |(xm − am) + c(xj − aj)|2 with m ≠ j and \(c\in \mathbb {C}\setminus \mathbb {R}\). Let P be a polynomial of degree at most d in \(\mathbb {R}^{n}\) of real coefficients. Then P is a multiple of qa if and only if
Proof
Without loss of generality we assume that j = n − 1 and m = n. We write \(P(\mathbf {x})={\sum }_{k=0}^{d} C_{k}(\mathbf {x}^{\prime }) {{x}_{n}^{k}}\) with \(C_{k}\in \mathcal {P}_{d-k}(\mathbb {R}^{n-1})\) and \(\mathbf {x}^{\prime }=(x_{1},\ldots ,x_{n-1})\). We can regard qa and P as polynomials in \(\mathbb {C}^{n}\), that is \(\mathbf {x}\in \mathbb {C}^{n}\). Since
we see that qa divides P if and only if both (xn − an) + c(xn− 1 − an− 1) and \((x_{n}-a_{n})+\overline c(x_{n-1}-a_{n-1})\) divide P. By [10, Lemma 2.5], the condition reduces to
and
Note that (5) and (6) hold for every \(\mathbf {x}\in \mathbb {C}^{n}\) if and only if they are true for every \(\mathbf {x}\in \mathbb {R}^{n}\). Hence we can return to work with polynomials in \(\mathbb {R}^{n}\). We have
where
and
Note that both CkQk and CkRk belong to \(\mathcal {P}_{d}(\mathbb {R}^{n-1})\). In other words, the real part and the imaginary part of \(P\left (\mathbf {x}^{\prime },a_{n}- c(x_{n-1}-a_{n-1})\right )\) are polynomials of degree at most d in \(\mathbb {R}^{n-1}\).
We consider the canonical basis for \(\mathcal {P}_{d}(\mathbb {R}^{n-1})\)
Observe that relation (5) holds if and only if
By the chain rule, it is easily seen that
and
More generally, relation (8) can be rewritten as
Using similar arguments applying to relation (6), we get
Note that P is a polynomial of real coefficients, two relations (9) and (10) are equivalent, and the proof is complete. □
We define
Then,
Theorem 1 can be restated as: the polynomial \(P\in \mathcal {P}_{d}(\mathbb {R}^{n})\) is divisible by qa(x) = |(xm − am) + c(xj − aj)|2 if and only if it satisfies the following relations
In the case where αj≠ 0, the differential operator \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})\) can be decomposed into non-zero real and imaginary parts. More precisely \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})=\text {Re}\left ({\varPi }_{\alpha ^{\prime }}(\mathrm {D})\right )+i \text {Im}\left ({\varPi }_{\alpha ^{\prime }}(\mathrm {D})\right )\) where
and
The above remark enables us to compute the number of interpolation conditions in the real setting in (11).
Lemma 1
Let \(\mathbf {a}=(a_{1},\ldots ,a_{n})\in \mathbb {R}^{n}\) and \(c\in \mathbb {C}\setminus \mathbb {R}\). Let 1 ≤ j,m ≤ n and j ≠ m. Then the set of functionals
and
consist of \(\left (\begin {array}{cc}n+d-1\\n-1 \end {array}\right )+\left (\begin {array}{cc}n+d-2\\n-1 \end {array}\right )\) elements.
To prove Lemma 1, we need the following simple result. The proof are left to the reader.
Lemma 2
For any 1 ≤ k ≤ n, we have
Proof Proof (Proof of Lemma 1)
For simplicity, we assume that j = n − 1 and m = n. If αn− 1 = 0, then α1 + ⋯ + αn− 2 ≤ d, and hence, we get \(\left (\begin {array}{cc}n+d-2\\n-2 \end {array}\right )\) differential operators of the forms
Otherwise, if 0 < k = αn− 1 ≤ d, then α1 + ⋯ + αn− 2 ≤ d − k. In this case, we have \(\left (\begin {array}{cc}d-k+n-2\\n-2 \end {array}\right )\) choices of (α1,…,αn− 2) and two choices corresponding to the real part and the imaginary part of \(\left (\frac {\partial }{\partial x_{n-1}}-c \frac {\partial }{\partial x_{n}} \right )^{\alpha _{n-1}}\). Hence, we get \(2 \left (\begin {array}{cc}d-k+n-2\\n-2 \end {array}\right )\) differential operators. The number of functionals coincide with the number of differential operators which are equal to
where we use Lemma 2 in the second relation. The proof is complete. □
2.2 Construction of Taylor type operators
We give a dual basis for the differential operators and use it to construct a Taylor type operator.
Lemma 3
Let \(\mathbf {a}=(a_{1},\ldots ,a_{n})\in \mathbb {R}^{n}\) and \(c\in \mathbb {C}\setminus \mathbb {R}\). For 1 ≤ j,m ≤ n and j ≠ m, \(\beta ^{\prime }=(\beta _{1},\ldots ,\beta _{m-1},\beta _{m+1},\ldots ,\beta _{n})\in \mathbb {N}^{n-1}\) we set
Then,
Proof
By definition, we have
It is easily to check that
and, for 1 ≤ k ≤ n, k ≠ j,m,
The result follows directly from the above computations. □
Proposition 1
Let \(\mathbf {a}=(a_{1},\ldots ,a_{n})\in \mathbb {R}^{n}\) and \(c\in \mathbb {C}\setminus \mathbb {R}\). Let 1 ≤ j,m ≤ n and j ≠ m. For a suitably defined function f, we set
where
and
Then, \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(f)\) belongs to \(\mathcal {P}_{d}(\mathbb {R}^{n})\) and
The polynomial \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(f)\) is called a Taylor type polynomial of f at a corresponding to qa.
Proof
Firstly, observe that \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})(f)(\mathbf {a})\in \mathbb {R}\) and \(B_{\alpha ^{\prime }}\in \mathcal {P}_{d}(\mathbb {R}^{n})\) when αj = 0. By definition, \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})(f)(\mathbf {a})\) and \(B_{\alpha ^{\prime }}(\mathbf {x})\) are the complex conjugates of \(\overline {{\varPi }}_{\alpha ^{\prime }}(\mathrm {D})(f)(\mathbf {a})\) and \(\overline {B}_{\alpha ^{\prime }}(\mathbf {x})\) respectively. In addition, \(B_{\alpha ^{\prime }}\) is a polynomial of degree \(|\alpha ^{\prime }|\leq d\). It follows that \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(f)\) is a polynomial of degree at most d with real coefficients. It remains to check the relation (14). From the formula, we consider three cases.
If \(\alpha ^{\prime }=0\), then \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(f)(\mathbf {a})=f(\mathbf {a})\), because \(B_{\beta ^{\prime }}(\mathbf {a})=0\) for every \(|\beta ^{\prime }|>0\).
Next, we assume that \(|\alpha ^{\prime }|>0\) and αj = 0. Then, for any \(\beta ^{\prime }\) with \(|\beta ^{\prime }|>0\) and βj > 0, we have \(\alpha ^{\prime }\ne \beta ^{\prime }\). It follows from Lemma 3 that \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})\left (B_{\beta ^{\prime }}\right )(\mathbf {a})=0\) for such \(\beta ^{\prime }\). Moreover
On the other hand, if \(|\beta ^{\prime }|>0\) and βj = 0 then, using Lemma 3 again, we obtain
Hence, relation (14) holds in this case.
Finally, we treat the case where \(|\alpha ^{\prime }|>0\) and αj > 0. It is easily seen that
It follows that
Furthermore, we can use Lemma 3 to get
Combining the last two relations, we obtain the desired equation. The proof is complete. □
Corollary 1
For a suitably defined function f, \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(f)=0\) if and only if
Proof
One direction is trivial. We assume that \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})(f)(\mathbf {a})=0\) for every \(|\alpha ^{\prime }|\leq d\). Then, its conjugate \(\overline {{\varPi }}_{\alpha ^{\prime }}(\mathrm {D})(f)(\mathbf {a})\) also vanishes. Hence, the conclusion follows directly from (13). □
Remark 1
The definition of the Taylor type polynomials gives a recurrent relation which is useful in computations,
2.3 Some properties of Taylor type operators
In this subsection, we show that the Taylor type polynomial has some expected properties. In particular, the Taylor type polynomial of any multiple of qa is identically zero.
Lemma 4
The set of polynomials
is D-invariance. In other words,
Proof
It suffices to check that
Relation (16) is trivial when \(\alpha ^{\prime }=0\). Hence, we can assume that \(|\alpha ^{\prime }|>0\). Direct computations gives
and
where δkl is the Kronecker symbol. It follows from the above formulas for \(\frac {\partial }{\partial x_{k}} {\varPi }_{\alpha ^{\prime }}\) that relation (16) holds in any cases. This finishes the proof of the lemma. □
Corollary 2
For any \(\beta \in \mathbb {N}^{n}\) and \(|\alpha ^{\prime }|\leq d\), we have
Proof
By Lemma 4, \(\mathrm {D}^{\beta }{\varPi }_{\alpha ^{\prime }}\) belongs to \(\mathcal {F}\). Hence, \(\left (\mathrm {D}^{\beta }{\varPi }_{\alpha ^{\prime }}\right )(\mathrm {D})\) is a linear combination of the operators \({\varPi }_{\gamma ^{\prime }}(\mathrm {D})\), \(|\gamma ^{\prime }|\leq d\). The desired relation now follows directly from (14). □
The result below asserts that the Taylor type operators obey the weak Leibniz rule.
Lemma 5
For suitably defined functions f and g, we have
In particular, \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(g)=\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}\left (\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(g)\right ).\)
Proof
It is sufficient to show that
and, for αj > 0,
Since \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(g)(\mathbf {a})=g(\mathbf {a})\), relation (18) is trivial. For \(0<|\alpha ^{\prime }|\leq d\), we can use the Leibniz-Hörmander formula (see, e.g., [7, p. 177] or [6, p. 243]) to get
Now, it follows from Corollary 2 that
Consequently,
where, in the second relation, we use the Leibniz-Hörmander formula again. This proves (19) and the first relation in (20). The second relation in (20) follows directly from the first one since
The proof is complete. □
Corollary 3
If \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})(f)(\mathbf {a})=0\) for any \(|\alpha ^{\prime }|\leq d\), then \( {\varPi }_{\alpha ^{\prime }}(\mathrm {D})(fg)(\mathbf {a})=0\) for any \(|\alpha ^{\prime }|\leq d\).
Proof
By Corollary 1, we have \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(f)=0\). Hence, using Lemma 5, we can write
Corollary 1 now gives the desired relations. □
Lemma 6
If Q is a multiple of qa, then \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(Q)=0\).
Proof
Without loss of generality we assume that j = n − 1 and m = n. We first prove that \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(q_{\mathbf {a}})=0\). We see that
Here, ha(x) = (xn − an) + c(xn− 1 − an− 1) and \(\overline {h}_{\mathbf {a}}(\mathbf {x})=(x_{n}-a_{n})+ \overline c (x_{n-1}-a_{n-1})\). Evidently, \({\varPi }_{\alpha ^{\prime }}(\mathrm {D}) (q_{\mathbf {a}})(\mathbf {a})=0\) when \(\alpha ^{\prime }=0\) or \(|\alpha ^{\prime }|\geq 3\). Direct computations show that
and
for every
Indeed, relation (24) follows directly from the fact that \(\frac {\partial }{\partial x_{k}}q_{\mathbf {a}}(\mathbf {a})=0\) for every 1 ≤ k ≤ n. Moreover, we see that
and
It follows that
Consequently, \({\varPi }_{\alpha ^{\prime }}(\mathrm {D})(q_{\mathbf {a}})(\mathbf {a})=0\) for every \(\alpha ^{\prime }\). Now, by the definition in Proposition 1 we get \(\mathrm {T}_{\mathbf {a}, q_{\mathbf {a}}}^{d}(q_{\mathbf {a}})=0\).
Next, we write Q = qaQ1. Then, Lemma 5 enables us to write
The proof is complete. □
3 Hermite interpolation in \(\mathbb {R}^{n}\)
3.1 Hermite interpolation schemes
In the main theorem below, we show that interpolation conditions corresponding to Taylor type polynomials can be collected to obtain regular Hermite interpolation schemes in \(\mathbb {R}^{n}\).
Theorem 2
Let d ≥ 2 be a positive integer and m = [d/2] + 1. Let sk = d − 2k + 2 for k = 1,…,m. Let 1 ≤ jk,mk ≤ n with jk≠mk and \(c^{[k]}\in \mathbb {C}\setminus \mathbb {R}\) for k = 1,…,m. Let A = {a1,…,am} be m distinct points in \(\mathbb {R}^{n}\) such that \(q_{\mathbf {a}_{k}}(\mathbf {a}_{j})\ne 0\) for j > k, where
For each 1 ≤ k ≤ m, let
where \(\left (\alpha ^{[k]}\right )^{\prime }=\left (\alpha ^{[k]}_{1},\ldots ,\alpha ^{[k]}_{m_{k}-1},\alpha ^{[k]}_{m_{k}+1},\ldots , \alpha ^{[k]}_{n} \right ).\) Then, for any function f is of class \(C^{s_{k}}\) in neighborhoods of the ak’s, there exists a unique polynomial \(P\in \mathcal {P}_{d}(\mathbb {R}^{n})\) such that
Moreover, \(P={\sum }_{k=1}^{m} P_{k}\), where \( P_{1}(\mathbf {x})=\mathrm {T}^{s_{1}}_{\mathbf {a}_{1}, q_{\mathbf {a}_{1}}}(f)(\mathbf {x})\),
Proof
For each 1 ≤ k ≤ m, Lemma 1 shows that the number of interpolation conditions in (27) is
where we use Lemma 2 in the above binomial relation. It follows that the number of interpolation conditions in the Hermite scheme is equal to
which matches the dimension of \(\mathcal {P}_{d}(\mathbb {R}^{n})\). Hence, to prove the regularity of the Hermite scheme, it is sufficient to check that if \(H\in \mathcal {P}_{d}(\mathbb {R}^{n})\) and
then H = 0. Since s1 = d, relation (28) along with Theorem 1 asserts that H divides \(q_{\mathbf {a}_{1}}\). Hence, we can write \(H=q_{\mathbf {a}_{1}} H_{1}\) with \(\deg H_{1}\leq d-2=s_{2}\). Using Corollary 3 for f = H and \(g=\frac {1}{q_{\mathbf {a}_{1}}}\), we get from (28) the following relations
By similar arguments, we have \(H_{1}=q_{\mathbf {a}_{2}} H_{2}\) with \(\deg H_{2}\leq d-4\). We continue in this fashion to obtain
It follows from the last relation that H = 0. Conversely, suppose that H≠ 0. Then, the degree of the polynomial on the right hand side is at least 2m > d. This contradicts to the fact that \(\deg H\leq d\), and the proof the first part of the theorem is complete.
It remains to prove the formula for the interpolation polynomial. We first check that the polynomial \(P={\sum }_{k=1}^{m} P_{k}\) belongs to \(\mathcal {P}_{d}(\mathbb {R}^{n})\). By definition we have \(P_{1}\in \mathcal {P}_{s_{1}}(\mathbb {R}^{n})=\mathcal {P}_{d}(\mathbb {R}^{n})\). For 2 ≤ k ≤ m, since \(\mathrm {T}^{s_{k}}_{\mathbf {a}_{k}, q_{\mathbf {a}_{k}}}(g)\in \mathcal {P}_{s_{k}}(\mathbb {R}^{n}) \), we get \(\deg P_{k}\leq 2(k-1)+s_{k}=d\). It follows that \(\deg P\leq d\). By Corollary 1, it is sufficient to show that
To prove relation (30) we first treat the case k = 1. In this case, we see that \(q_{\mathbf {a}_{1}}\) divides Pk for any k ≥ 2. Hence, Lemma 6 gives \(\mathrm {T}^{s_{1}}_{\mathbf {a}_{1},q_{\mathbf {a}_{1}}}(P_{k})=0\) for 2 ≤ k ≤ m. This enables us to write
where we use Lemma 5 in the second relation. Next, we assume that 2 ≤ k ≤ m. Using the fact that Pj contains the factor \(q_{\mathbf {a}_{k}}\) for k + 1 ≤ j ≤ m and Lemma 6 we get \(\mathrm {T}^{s_{k}}_{\mathbf {a}_{k},q_{\mathbf {a}_{k}}}(P_{j})=0\) for k + 1 ≤ j ≤ m. Therefore
From Lemma 5, we have
Combining the last relation with (31) we finally obtain \(\mathrm {T}^{s_{k}}_{\mathbf {a}_{k},q_{\mathbf {a}_{k}}}(P)=\mathrm {T}^{s_{k}}_{\mathbf {a}_{k},q_{\mathbf {a}_{k}}}(f)\), which proves the claim. The proof is complete.
□
Corollary 4
The interpolation polynomial \(P\in \mathcal {P}_{d}(\mathbb {R}^{n})\) in Theorem 2 is determined by the following relation
Remark 2
The condition \(q_{\mathbf {a}_{k}}(\mathbf {a}_{j})\ne 0\) for j > k is used in the proof of Theorem 2. From (3), we see that it is equivalent to
In other words, the mk-coordinate and the jk-coordinate of ak and aj are not simultaneously identical for any j > k.
Definition 1
The interpolation polynomial \(P\in \mathcal {P}_{d}(\mathbb {R}^{n})\) in Theorem 2 is called a Hermite type interpolation polynomial of f at A. We write
From the Newton type formula in Theorem 2, we obtain an algorithm to compute the polynomial \(\mathbf H[\{(\mathbf {a}_{1}, q_{\mathbf {a}_{1}},s_{1}),\ldots ,(\mathbf {a}_{m}, q_{\mathbf {a}_{m}},s_{m})\};f ]\).
-
Step 1. Compute \(P_{1}=\mathrm {T}^{s_{1}}_{\mathbf {a}_{1}, q_{\mathbf {a}_{1}}}(f)\) by using (13);
-
Step 2. Compute \(P_{k}={\prod }_{j=1}^{k-1} q_{\mathbf {a}_{j}}\mathrm {T}^{s_{k}}_{\mathbf {a}_{k},q_{\mathbf {a}_{k}}}\left (\frac {f-P_{1}-\cdots -P_{k-1}}{{\prod }_{j=1}^{k-1} q_{\mathbf {a}_{j} }} \right )\) for k = 2,…,m respectively by using (13);
-
Step 3. Compute the sum \(\mathbf H[\{(\mathbf {a}_{1}, q_{\mathbf {a}_{1}},s_{1}),\ldots ,(\mathbf {a}_{m}, q_{\mathbf {a}_{m}},s_{m})\};f ]=P_{1}+\cdots +P_{m}\).
Remark 3
We have known that some kinds of Hermite interpolants are the limits of Lagrange interpolants when interpolation points coalesce (see, e.g., [11,12,13,14]). One may ask whether there are regular Lagrange interpolation schemes such that the corresponding Lagrange interpolation polynomials of sufficiently smooth functions converge to the Hermite interpolation polynomial constructed in Theorem 2.
3.2 Some examples
In this subsection, we compute the set differential operators of degree 2. We also give explicit formulas for Hermite interpolation polynomials of degree 2 and degree 3 in \(\mathbb {R}^{3}\).
Example 1
This example gives interpolation conditions for Hermite interpolation of degree 2 at two points in \(\mathbb {R}^{n}\). Let d = 2. Then m = 2, s1 = 2 and s2 = 0. Let \(\mathbf {a}_{k}=\left (a^{[k]}_{1},\ldots ,a^{[k]}_{n}\right )\) for k = 1,2. Set c[k] = uk + ivk with vk≠ 0, k = 1,2. Take 1 ≤ jk,mk ≤ n such that jk≠mk for k = 1,2. Set
We have
Since s2 = 0, we get
The interpolation conditions in (27) corresponding the following differential operators
Easy computations show that the above relations are equivalent to \(\left (\begin {array}{cc}n+2\\2 \end {array}\right )\) interpolation conditions in the real settings:
Example 2
We construct a formula for the Hermite interpolation polynomial of degree 2 at two points in \(\mathbb {R}^{3}\). Let n = 3 and d = 2. Then, m = 2,s1 = 2 and s2 = 0. Let a1 = (0,0,0) and a2 = (0,1,0). We choose c[1] = 2i and c[2] = 3i. Take j1 = 1,m1 = j2 = 2, m2 = 3. Then,
The polynomial \(P=\mathbf H[\{(\mathbf {a}_{1},q_{\mathbf {a}_{1}},2),(\mathbf {a}_{2},q_{\mathbf {a}_{2}},0) \}](f)\) belonging to \(\mathcal {P}_{2}(\mathbb {R}^{3})\) interpolates f at the following 10 conditions
By definition of Taylor type polynomial in (13), we have
where
and
Hence,
and
Combining the above computations we obtain
On the other hand,
It follows that
Example 3
We give a formula for the Hermite interpolation polynomial of degree 3 at two points in \(\mathbb {R}^{3}\). Let n = 3 and d = 3. We have m = 2,s1 = 3 and s2 = 1. Let a1 = (0,0,0) and a2 = (0,1,0). We choose c[1] = 2i and c[2] = 3i. Take j1 = 1,m1 = j2 = 2, m2 = 3. Then
The polynomial \(P=\mathbf H[\{(\mathbf {a}_{1},q_{\mathbf {a}_{1}},2),(\mathbf {a}_{2},q_{\mathbf {a}_{2}},1) \}](f)\) belonging to \(\mathcal {P}_{3}(\mathbb {R}^{3})\) satisfies the following conditions in the complex setting
and
They are equivalent to 20 conditions which consist of 10 conditions in (33)–(35) along with the following functionals
Theorem 2 gives the formula for the interpolation polynomial
We need to compute \(\mathrm {T}^{3}_{\mathbf {a}_{1}, q_{\mathbf {a}_{1}}}(f)\) and \(\mathrm {T}^{1}_{\mathbf {a}_{2}, q_{\mathbf {a}_{2}}}(g)\). From (15), we can write
where Σl denotes the (l + 1)-th term at the right hand side for 0 ≤ l ≤ 3,
and
Direct computations give
It follows that
Next, we calculate
From the formula for \(\mathrm {T}^{3}_{\mathbf {a}_{1}, q_{\mathbf {a}_{1}}}(f),\) we see that
Direct computations give
Hence,
The above computations along with (36) lead to a formula for \(\mathbf H[\{(\mathbf {a}_{1},q_{\mathbf {a}_{1}},3),\) \((\mathbf {a}_{2},q_{\mathbf {a}_{2}},1) \}]\). The precise formula is left to the readers.
Remark 4
The interpolation conditions in the previous two examples and the corresponding Hermite interpolation polynomials do not depend on the choices of j2 and m2.
References
Bojanov, B., Xu, Y.: On a Hermite interpolation by polynomials of two variables. SIAM J. Numer. Anal. 39, 1780–1793 (2002)
Bos, L.: On certain configurations of points in \(\mathbb {R}^{n}\) which are unisolvent for polynomial interpolation. J. Approx. Theory 64, 271–280 (1991)
Bos, L., Calvi, J.-P.: Multipoint Taylor interpolation. Calcolo. 45, 35–51 (2008)
Bos, L., Calvi, J.-P.: Taylorian points of an algebraic curve and bivariate Hermite interpolation. Ann. Scuola Norm. Sup. Pisa Cl. Sci. 7, 545–577 (2008)
Chung, K.C., Yao, T.H.: On lattices admitting unique Lagrange interpolations. SIAM J. Numer. Anal. 14, 735–743 (1977)
Duistermaat, J.J., Kolk, J.A.C.: Multidimensional real analysis I: differentiation. Cambridge University Press, Cambridge (2004)
Hörmander, L.: On the theory of general partial differential operators. Acta Math. 94, 161–248 (1955)
Lorentz, R.A.: Multivariate Birkhoff Interpolation, Lecture Notes in Mathematics, vol. 1516. Springer, Berlin (1992)
Lorentz, R.A.: Multivariate Hermite interpolation by algebraic polynomials: a survey. J. Comput. Appl. Math. 122, 167–201 (2000)
Milnor, J.: Singular Points for Complex Hypersurfaces. Ann. of Math. Stud. 61. Princeton University Press, Princeton (1968)
Phung, V.M.: On bivariate Hermite interpolation and the limit of certain bivariate Lagrange projectors. Ann. Polon. Math. 115, 1–21 (2015)
Phung, V.M.: Hermite interpolation with symmetric polynomials. Numer. Algorithms 76, 709–725 (2017)
Phung, V.M.: Polynomial in \(\mathbb {R}^{2}\) and on the unit sphere in \(\mathbb {R}^{3}\). Acta. Math. Hungar. 153, 289–317 (2017)
Phung, V.M.: Hermite interpolation on the unit sphere and limits of Lagrange projectors. IMA J. Numer. Anal. https://doi.org/10.1093/imanum/draa026 (2020)
Phung, V.M., Nguyen, V.K.: Polynomial interpolation on the unit sphere and some properties of its integral means. Filomat. 33, 4697–4715 (2019)
Sauer, T., Xu, Y.: On multivariate hermite interpolation. Adv. Comput. Math. 4, 207–259 (1995)
Acknowledgements
We are grateful to anonymous referees for their constructive comments. A part of this work was done when the first author was working at Vietnam Institute for Advanced Study in Mathematics (VIASM). He would like to thank the VIASM for providing a fruitful research environment and working condition.
Funding
This research is funded by the Vietnam Ministry of Education and Training under grant number B2021-SPH-16.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
About this article
Cite this article
Van Manh, P., Van Trao, N., Tung, P.T. et al. Taylor type and Hermite type interpolants in \(\mathbb {R}^{n}\). Numer Algor 89, 145–166 (2022). https://doi.org/10.1007/s11075-021-01109-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11075-021-01109-6