The evaluation of errors in computing by interpolation polynomials

Abstract

Authors

T. Popoviciu
Institutul de Calcul

Keywords

?

Paper coordinates

T. Popoviciu, Evaluarea erorilor de calcul în interpolarea prin polinoame. (Romanian), Stud. Cerc. Mat. (Cluj), 13 (1962) fasc. anexă, pp. 221-241.

PDF

About this paper

Journal

Studii si Cercetari Matematice

Publisher Name

Academy of the Republic of S.R.A

Print ISSN

1220-269X

Online ISSN

google scholar link

??

Paper (preprint) in HTML form

Original text
Rate this translation
Your feedback will be used to help improve Google Translate

EVALUATION OF CALCULATION ERRORS IN INTERPOLATION BY POLYNOMIALS

of

T. POPOVICIU

Member of the RPR Academy
(Cluj)

  1. 1.

    Practical use of Lagrange's interpolation formula

f(x)IT(x1,x2,,xn+1;fx),f(x)\approx L\left(x_{1},x_{2},\ldots,x_{n+1};f\mid x\right), (1)

where in the second term we have the Lagrange polynomial of degreennwhich takes the values ​​of the functionf(x)f(x)on the (distinct) interpolation nodesxand(and=1,2,x_{i}(i=1,2,\ldots,n+1n+1), depends largely on the form in which this polynomial is put. This is especially important when it comes to the actual calculation of the values ​​of the polynomial for different values ​​of the variablexxHere we only deal with the case when the variablexxand the functionf(x)f(x)are real. In this case, to find the value of the polynomial, a finite number of elementary operations of addition, subtraction, multiplication and division must be performed on real numbers, usually limited decimal fractions, and in a determined order. The form in which the Lagrange polynomial is put is related precisely to this order of operations. As for the operations, they are executed, based on known procedures, directly or by machine, and which practically come down to the successive determination of the decimal digits of the result of each partial calculation, exact or approximate, in part.

The calculation procedure used involves inherent errors due to the fact that the successive calculations are made with approximation, for example due to the limitation to a certain number of decimal places of the partial results obtained. These errors are reflected in the final result, determining a correction which, in the interpolation problem considered, will have to be added to the corrections determined by the errors affecting the data of the problem.

We will first assume that the nodes are not necessarily equidistant and we will deal with the calculation of the second term of formula (1) forx=x0x=x_{0}, using Newton's formula

IT(x1,x2,,xn+1;fx0)=and=0n(x0x1)(x0x2)(x0xand)D1and,;L\left(x_{1},x_{2},\ldots,x_{n+1};f\mid x_{0}\right)=\sum_{i=0}^{n}\left(x_{0}-x_{1}\right)\left(x_{0}-x_{2}\right)\ldots\left(x_{0}-x_{i}\right)D_{1}^{i},*; (2)

where the coefficientsDandj,(and=0,1,,n)D_{i}^{j},(i=0,1,\ldots,n), are obtained from table 1 of divided differences in which we generally have

Dandj=Dandj[f]=[xand,xand+1,,xand+j;f],Dand0=fand=f(xand),D_{i}^{j}=D_{i}^{j}[f]=\left[\begin{array}[]{cccc}x_{i},&x_{i+1},&\ldots,&x_{i+j};f\end{array}\right],D_{i}^{0}=f_{i}=f\left(x_{i}\right),

using the usual notations for divided differences of the functionf(x)f(x)
The formation of table 1 of the divided differences is done by successively calculating the columns using the recurrence formula

Dandj[f]=Dand+1j1[f]Dandj1[f]xand+jxand(Dand0[f]=fand)\displaystyle D_{i}^{j}[f]=\frac{D_{i+1}^{j-1}[f]-D_{i}^{j-1}[f]}{x_{i+j}-x_{i}}\quad\left(D_{i}^{0}[f]=f_{i}\right) (4)
(and=1,2,,n+1j;j=1,2,,n).\displaystyle(i=1,2,\ldots,n+1-j;\quad j=1,2,\ldots,n).

It is seen that this calculation requires divisions (with node differences) due to which, in general, the divided differences cannot be calculated, for example, in the practical form of limited decimal fractions, except with a certain approximation, even if the nodes and the function values ​​at these nodes are given in this form.

Table 1: Table 1
xx f(x)f(x) D1D^{1} D2D^{2} D3D^{3} ... DnD^{n}
x1x_{1} f1f_{1}
D11D_{1}^{1}
x2x_{2} f2f_{2} D12D_{1}^{2}
x3x_{3} f3f_{3} D21D_{2}^{1} D22D_{2}^{2} D13D_{1}^{3}
\vdots \vdots \vdots D2n1D_{2}^{n-1} D1nD_{1}^{n}
xn1x_{n-1} fn1f_{n-1} Dn22D_{n-2}^{2}
Dn11D_{n-1}^{1} D23D_{*-2}^{3}
xnx_{n} tnt_{n} Dn12D_{n-1}^{2}
xn+1x_{n+1} fn+1f_{n+1} Dn1D_{n}^{1}
0 0 footnotetext: *) Forand=0i=0the corresponding term of the amount is reduced toDand0D_{i}^{0}An analogous convention

However, if the elements of array 1 are calculated approximately, the errors made are transposed onto the interpolation polynomial.

On the other hand, the multiplications involved in calculating the value of the interpolation polynomial are also made with errors. In what follows we propose to examine in turn the effect of these two types of errors, which overlap. For simplicity we will assume that the nodes and the pointx0x_{0}, where interpolated, are not affected by errors.
2. We deal first with the effect of errors that occur in the calculation of divided differences, assuming that the multiplications indicated in formula (2) are done exactly.

The columns of table 1 are calculated successively with certain approximations.
Letf~and,(and=1,2,,n+1)\widetilde{f}_{i},(i=1,2,\ldots,n+1), some approximate values ​​of the function values ​​on the nodes,cand(0),(and=1,2,,n+1)c_{i}^{(0)},(i=1,2,\ldots,n+1), the respective corrections, sofand=,~and+cand(0),(and=1,2,,n+1)f_{i}=\tilde{,}_{i}+c_{i}^{(0)},(i=1,2,\ldots,n+1). Let then, in general,D~andj=D~andj[f]\widetilde{D}_{i}^{j}=\widetilde{D}_{i}^{j}[f],(and=1,2,,n+1j)(i=1,2,\ldots,n+1-j), approximate values ​​of the column elementsjj, calculated using the already calculated elements of the columnj1j-1andcand(j),(and=1,2,,n+1j)c_{i}^{(j)},(i=1,2,\ldots,n+1-j), the respective corrections. We will then have

D~andj+cand(j)=\displaystyle\widetilde{D}_{i}^{j}+c_{i}^{(j)}= D~and+1j1D~andj1xand+jxand(D~and0=f~and)\displaystyle\frac{\widetilde{D}_{i+1}^{j-1}-\widetilde{D}_{i}^{j-1}}{x_{i+j}-x_{i}}\left(\widetilde{D}_{i}^{0}=\widetilde{f}_{i}\right) (5)
(and=1,2,,n+1j,j=1,2,\displaystyle(i=1,2,\ldots,n+1-j,\quad j=1,2, (.,n).,n))

Table 1 of the divided differences is then replaced with table 2 of the calculated approximate values ​​of the divided differences.

Table 2: Table 2
xx f~(x)\widetilde{f}(x) D~1\widetilde{D}^{1} D~2\widetilde{D}^{2} D~3\widetilde{D}^{3} D~n1\widetilde{D}^{n-1} D~n\widetilde{D}^{n}
x1x_{1} f~1\widetilde{f}_{1} D~11\widetilde{D}_{1}^{1}
x2x_{2} f2f_{2} D~12\widetilde{D}_{1}^{2}
D~21\widetilde{D}_{2}^{1} D~13\widetilde{D}_{1}^{3}
x3x_{3} f3~\widetilde{f_{3}} D~22\widetilde{D}_{2}^{2} D~1n1\widetilde{D}_{1}^{n-1}
\vdots xn1x_{n-1} f~n1\vdots\begin{gathered}\vdots\\ \tilde{f}_{n-1}\end{gathered} \vdots \vdots \vdots D~2n1\widetilde{D}_{2}^{n-1} D~1n\widetilde{D}_{1}^{n}
xn1x_{n-1} f~˙n1\dot{\tilde{f}}_{n-1} D~n22\widetilde{D}_{n-2}^{2}
xnx_{n} f~n\widetilde{f}_{n} D~n11\widetilde{D}_{n-1}^{1} D~n12\widetilde{D}_{n-1}^{2} D~n23\widetilde{D}_{n-2}^{3}
xn+1x_{n+1} f~n+1\widetilde{f}_{n+1} D~n1\widetilde{D}_{n}^{1}

An approximate valueIT~\widetilde{L}of expression (2) is then obtained using table 2 with the formula

IT~=and=0n(x0x1)(x0x2)(x0xand)D~1and\widetilde{L}=\sum_{i=0}^{n}\left(x_{0}-x_{1}\right)\left(x_{0}-x_{2}\right)\cdots\left(x_{0}-x_{i}\right)\widetilde{D}_{1}^{i} (6)

and we will have

IT(x1,x2,,xn+1;fx0)=IT~+λL\left(x_{1},x_{2},\ldots,x_{n+1};f\mid x_{0}\right)=\widetilde{L}+\lambda (7)

λ\lambdabeing the respective correction.
In order to delimit the correctionλ\lambdalinear functionals are introducedDandj,k[f]D_{i}^{j,k}[f], defined as follows(k<n)(k<n):

Dandj,k[f]=Dand+1j1,k[f]Dandj1,k[f]xand+j+kxand,D_{i}^{j,k}[f]=\frac{D_{i+1}^{j-1,k}[f]-D_{i}^{j-1,k}[f]}{x_{i+j+k}-x_{i}}, (8)

forj=1,2,,nk,and=1,2,,n+1kjj=1,2,\ldots,nk,i=1,2,\ldots,n+1-kjand

Dand0,k[f]=fand,D_{i}^{0,k}[f]=f_{i}, (9)

forand=1,2,,n+1ki=1,2,\ldots,n+1-kIn the paper [1] the formula is established

Dandj,k[f]=DandR,k+jR[Dandj1,k[f]](0Rj),D_{i}^{j,k}[f]=D_{i}^{r,k+jr}\left[D_{i}^{j-1,k}[f]\right]\quad(0\leqslant r\leqslant j), (10)

which is also written symbolically

Dα+β,k=Dα,β+kDβ,kD^{\alpha+\beta,k}=D^{\alpha,\beta+k}D^{\beta,k}

The notation used is based on the observation that any string ofn+1n+1NUMBERSφ1,φ2,,φn+1\varphi_{1},\varphi_{2},\ldots,\varphi_{n+1}can be considered as formed by the valuesφand=φ(xand)\varphi_{i}=\varphi\left(x_{i}\right),
=1,2,,n+1=1,2,\ldots,n+1of a functionφ(x)\varphi(x), defined on the nodesx1,x2,,xn+1x_{1},x_{2},\ldots,x_{n+1}
Taking into account formula (10), the following expression for the correction was obtained in the cited work : λ\lambda:

λ=and=0n(x0x1)(x0x2)(x0xand)[R=0andD1R1,andR[c(andR)]].\lambda=\sum_{i=0}^{n}\left(x_{0}-x_{1}\right)\left(x_{0}-x_{2}\right)\ldots\left(x_{0}-x_{i}\right)\left[\sum_{r=0}^{i}D_{1}^{r_{1},ir}\left[c^{(ir)}\right]\right]. (11)

To delimit the correctionλ\lambdawe need a convenient delimitation forDandj,k[f]D_{i}^{j,k}[f].

Divided differenceD1n=[x1,x2,,xn+1;f]D_{1}^{n}=\left[x_{1},x_{2},\ldots,x_{n+1};f\right]is easily delimited taking into account the well-known formula

[x1,x2,,xn+1;f]=and=1n+1fandit(xand),\left[x_{1},x_{2},\ldots,x_{n+1};f\right]=\sum_{i=1}^{n+1}\frac{f_{i}}{l^{\prime}\left(x_{i}\right)},

whereit(x)=(xx1)(xx2)(xxn+1)l(x)=\left(x-x_{1}\right)\left(x-x_{2}\right)\ldots\left(x-x_{n+1}\right)We have

|D1n|N(x1,x2,,xn+1)M,\left|D_{1}^{n}\right|\leqslant N\left(x_{1},x_{2},\ldots,x_{n+1}\right)M,

where

N(x1,x2,,xn+1)=and=1n+11|it(xand)|N\left(x_{1},x_{2},\ldots,x_{n+1}\right)=\sum_{i=1}^{n+1}\frac{1}{\left|l^{\prime}\left(x_{i}\right)\right|} (12)

andM=MAXand=1,2,,n+1(|fand|)M=\max_{i=1,2,\ldots,n+1}\left(\left|f_{i}\right|\right)An analogous delimitation
ofD1nk,h[f]D_{1}^{n-k,h}[f]it can be written

|D1nk,k[f]|Nk(x1,x2,,xn+1)M\left|D_{1}^{n-k,k}[f]\right|\leqslant N_{k}\left(x_{1},x_{2},\ldots,x_{n+1}\right)M (13)

where, by the way,MMcan be replacedwithMAXand=1,2,,n+1k(|fand|)\mathrm{cu}\underset{i=1,2,\ldots,n+1-k}{\max}\left(\left|f_{i}\right|\right)
In formula (13) we have

Nk(x1,x2,,xn+1)=soupand=1,2,,n+1k|fand|1|D1nk,k[fand]|.N_{k}\left(x_{1},x_{2},\ldots,x_{n+1}\right)=\sup_{\begin{subarray}{c}i=1,2,\ldots,n+1-k\\ |fi|\leq 1\end{subarray}}\left|D_{1}^{n-k,k}\left[f_{i}\right]\right|.

and therefore

Nk(x1,x2,,xn+1)=D1nk,k[f]N_{k}\left(x_{1},x_{2},\ldots,x_{n+1}\right)=D_{1}^{n-k,k}[f]

provided we choose all the numbersfand,and=1,2,,n+1kf_{i},i=1,2,\ldots,n+1-k, equal to 1 or -1 , conveniently.

we

N0(x1,x2,,xn+1)=N(x1,x2,,xn+1)N_{0}\left(x_{1},x_{2},\ldots,x_{n+1}\right)=N\left(x_{1},x_{2},\ldots,x_{n+1}\right)

the second term being given by (12). Fork>0k>0, his expressionNk(x1,x2,,xn+1)N_{k}\left(x_{1},x_{2},\ldots,x_{n+1}\right)but it is more complicated.

In his formationD1nk,k[t]D_{1}^{n-k,k}[t]only the nodes actually intervenexandx_{i}forand=1,2,,nk,k+2,k+3,,n+1i=1,2,\ldots,n-k,k+2,k+3,\ldots,n+1. So if2kn12k\leqslant n-1, all nodes will intervene. However, if2k>n12k>n-1only the first ones will intervenenkn-kand the last onesnkn-kbetween the nodesx1,x2,,xn+1x_{1},x_{2},\ldots,x_{n+1}We deduce from this the following formula:

Nk(x1,x2,,xn+1)=Nnk1(x1,x2,,xnk,xk+2,xk+3,,xn+1)N_{k}\left(x_{1},x_{2},\ldots,x_{n+1}\right)=N_{n-k-1}\left(x_{1},x_{2},\ldots,x_{n-k},x_{k+2},x_{k+3},\ldots,x_{n+1}\right) (14)

if2k>n12k>n-1It is enough to assumek<nk<n, because we obviously have

Nn(x1,x2,,xn+1)=1N_{n}\left(x_{1},x_{2},\ldots,x_{n+1}\right)=1 (15)

In the particular case

x1<x2<<xn+1x_{1}<x_{2}<\ldots<x_{n+1} (16)

it is shown that the formula holds

Nk(x1,x2,,xj+k+1)=Nk(x1,x2,,xj+k)+Nk(x2,x3,,xj+k+1)xk+k+1x1N_{k}\left(x_{1},x_{2},\ldots,x_{j+k+1}\right)=\frac{N_{k}\left(x_{1},x_{2},\ldots,x_{j+k}\right)+N_{k}\left(x_{2},x_{3},\ldots,x_{j+k+1}\right)}{x_{k+k+1}-x_{1}} (17)

15 - Studies and research in mathematics (Cluj), annex/1961-1962

Formulas (14), (15), (17) completely determineNk(x1,x2,,xand)N_{k}\left(x_{1},x_{2},\ldots,x_{i}\right)and we have

|Dandj,k[f]|Nk(xand,xand+1,,xand+j+k)M\left|D_{i}^{j,k}[f]\right|\leqslant N_{k}\left(x_{i},x_{i+1},\ldots,x_{i+j+k}\right)M

whereM=MAXR=and,and+1,,and+j(|fR|)M=\max_{r=i,i+1,\ldots,i+j}\left(\left|f_{r}\right|\right)In particular ,
if all correctionscand(j)c_{i}^{(j)}remain, in absolute value at most equal toε\varepsilon, based on formula (15) we have

|D1R,andR[c(andR)]|NandR(x1,x2,,xand+1)ε\left|D_{1}^{r,i-r}\left[c^{(i-r)}\right]\right|\leqslant N_{i-r}\left(x_{1},x_{2},\ldots,x_{i+1}\right)\varepsilon

and we can write

|λ|V(x0)ε|\lambda|\leqslant V\left(x_{0}\right)\varepsilon ((18)

where

V(x)\displaystyle V(x) =and=1n|(xx1)(xx2)(xxand)|×\displaystyle=\sum_{i=1}^{n}\left|\left(x-x_{1}\right)\left(x-x_{2}\right)\ldots\left(x-x_{i}\right)\right|\times
×[R=0andNandR(x1,x2,,xand+1)].\displaystyle\times\left[\sum_{r=0}^{i}N_{i-r}\left(x_{1},x_{2},\ldots,x_{i+1}\right)\right]. (19)

When the values ​​of the function are exact, that is, whencand0=0,and=1,2,,n+1c_{i}^{0}=0,i=1,2,\ldots,n+1. can be taken even

V(x)\displaystyle V(x) =and=1n|(xx1)(xx2)(xxand)|×\displaystyle=\sum_{i=1}^{n}\left|\left(x-x_{1}\right)\left(x-x_{2}\right)\ldots\left(x-x_{i}\right)\right|\times
×[R=0and1NandR(x1,x2,,xand+1)].\displaystyle\times\left[\sum_{r=0}^{i-1}N_{i-r}\left(x_{1},x_{2},\ldots,x_{i+1}\right)\right]. (20)
  1. 3.

    We will continue to deal with the effect of errors arising from the multiplications indicated in formula (2), assuming that the divided differencesD1andD_{1}^{i}which intervene in this formula are calculated exactly. For this, certain theoretical considerations on the interpolation formula (1) are necessary.

At any permutation

xV1,xV2,,xVS+1x_{v_{1}},x_{v_{2}},\ldots,x_{v_{s+1}} (21)

of the nodes

x1,x2,,xn+1x_{1},x_{2},\ldots,x_{n+1} (22)

an interpolation formula of the form will correspond

f(x)\displaystyle f(x) =and=0n(xxV1)(xxV1)(xxVand)×\displaystyle=\sum_{i=0}^{n}\left(x-x_{v_{1}}\right)\left(x-x_{v_{1}}\right)\ldots\left(x-x_{v_{i}}\right)\times
×[xV1,xV2,,xVand+1;f]+R(x)\displaystyle\times\left[x_{v_{1}},x_{v_{2}},\ldots,x_{v_{i+1}};f\right]+R(x) (23)

Of course the restR(x)R(x)does not depend on the permutation considered and can be written

R(x)=it(x)[x,x1,x2,,xn+1;f]R(x)=l(x)\left[x,x_{1},x_{2},\ldots,x_{n+1};f\right] (24)

,where

it(x)=(xx1)(xx2)(xxn+1)l(x)=\left(x-x_{1}\right)\left(x-x_{2}\right)\ldots\left(x-x_{n+1}\right) (25)

The practical use of the interpolation formula (23) will depend on the speed and accuracy of the calculation of the coefficients

[xV1,xV2,xVand+1;f],and=0,1,,n\left[x_{v_{1}},x_{v_{2}},\ldots x_{v_{i+1}};f\right],i=0,1,\ldots,n (26)

of this formula.
These coefficients will be found on the upper (descending) side of the table of divided differences corresponding to the nodes taken in the order (21). However, for the formation of this table, it will be advantageous for the nodes (21) to be placed in the ascending order of their numerical values, because then all the divisions that are made with node differences in order to form the table, are made with positive numbers. In this way, errors that could come from inattention to the signs of the node differences are eliminated.

To simplify the notations, let us assume that the nodes, which we consider distinct, are taken in ascending order, so that

x1<x2<<xn+1x_{1}<x_{2}<\ldots<x_{n+1}

Then, based on the previous observation, table 1 is the most advantageous of all.(n+1)(n+1)! analogous arrays obtained by permuting in all the waysn+1n+1knots.

We observe that not only formula (2) enjoys the property that all its coefficients are contained in table 1. Let us then find all formulas (23) that have all their coefficients in table 2. We will say that these formulas belong to table 1. For this it is necessary and sufficient that for anyand=0,1,,ni=0,1,\ldots,npointsxn1,xn2,,xnand+1x_{\nu_{1}},x_{\nu_{2}},\ldots,x_{\nu_{i+1}}to beand+1i+1consecutive points (taken in a certain order) of the string (22), so as the permutation

V1,V2,Vn+1v_{1},v_{2},\ldots v_{n+1} (27)

of numbers1,2,,n+11,2,\ldots,n+1to enjoy the property that any section

V1,V2,,Vand+1,and=0,1,,nv_{1},v_{2},\ldots,v_{i+1},\quad i=0,1,\ldots,n

of the string (27) to be formed byand+1i+1consecutive natural numbers, taken in a certain order.

If we agree to say about such a permutation (27) that it is a permutation consecutive to the initial permutation1,2,,n+11,2,\ldots,n+1, we see that:
the necessary and sufficient condition for the interpolation formula (23) to belong to table 1 is that the permutation (27) is a permutation consecutive to the initial permutation1,2,,n+11,2,\ldots,n+1[the permutation with which tableau 1 was formed].

Let us call neighboring elements of an element of the array 1 those terms of the array, which are found in the immediately preceding column and in the
immediately following column, located immediately above and immediately below the element under consideration. The neighboring elements of the element[xand,xand+1,,xand+j;f]\left[x_{i},x_{i+1},\ldots,x_{i+j};f\right]are then depicted in the diagram

The coefficients (26) of formulas (23) belonging to table 1 are then characterized by the property that any two consecutive coefficients are neighboring elements in table 1.

The number of interpolation formulas (23) belonging to table 1 can be easily determined. A formula (23) is obtained by choosing coefficients, one from each column, two consecutive coefficients being neighbors. The numberNnN_{n}of these formulas is therefore equal to the number of paths that can be described starting from the first column to the last, passing only through neighboring elements.

Let's assume that(n+1)!(n+1)!divides by2n2^{n}and then(n+1)!=2np(n+1)!=2^{n}p. However, the question may be asked whether they cannot be foundppdivided difference tables, obtained by convenient permutations of the nodes (1), so that they contain all the(n+1)(n+1)! formulas (5). In this case, each of the(n+1)(n+1)! formulas (5) belong to one and only one of the following tableausppconsidered paintings. In the paper [2] it is shown that this is impossible, except for the trivial casen=1n=1.

Also in the paper [2] it is shown that the number of formulas (23) belonging to table no. 1 in which the firstjjnodes arexand,xand+1,,xand+j1x_{i},x_{i+1},\ldots,x_{i+j-1}in any order, it is equalwith2j1(nj+1and1)\mathrm{cu}2^{j-1}\binom{n-j+1}{i-1}.

With these introductory considerations, we propose to determine the permutation (21) of the nodes, so that the effect of the errors coming from the multiplications (2) is the smallest, when these multiplications are done as usual according to the scheme

α0(D10+α1)(D11++αn1(D1n1+αnD1n).)),\left.\alpha_{0}\left(D_{1}^{0}+\alpha_{1}\right)\left(D_{1}^{1}+\ldots+\alpha_{n-1}\left(D_{1}^{n-1}+\alpha_{n}D_{1}^{n}\right)\ldots.\right)\right),

where was it notedαand=x0xand,and=0,1,,n\alpha_{i}=x_{0}-x_{i},i=0,1,\ldots,nThis comes down to determining the permutation (21) such that the string

|xxV1|,|xxV2|,,|xxVn+1|\left|x-x_{v_{1}}\right|,\left|x-x_{v_{2}}\right|,\ldots,\left|x-x_{v_{n+1}}\right| (28)

be non-decreasing.
By this property, the permutation (2) is not necessarily uniquely determined, but:
xxbeing a given point on the real axis, any permutation (21) for which the sequence (28) is nondecreasing is a permutation consecutive to the initial permutation1,2,,n+11,2,\ldots,n+1.

This property subsists regardless of the fact thatxxcoincides with a node or not.

If the permutation (21) enjoys the property that it is possible to find axx, so that the sequence (28) is non-decreasing, we say that this permutation or the corresponding formula (23) verifies the minimum property.

The interpolation formulas (23), for which the sequence (28) becomes non-decreasing for axxconveniently, they all belong to the array 1 formed with the nodes taken in ascending order.

Not all of them2n2^{n}Formulas (23) belonging to table 1 verify the minimum property studied.

It is shown in the paper [2] that at least2n2nof these formulas verify the minimum property.

Forn=1,2n=1,2the 2 and 4 formulas belonging to table 1 respectively verify the minimum property.

Ifn=3n=3andx1+x4x2+x3,7x_{1}+x_{4}\neq x_{2}+x_{3},7of the 8 formulas belonging to table 1 check the minimum probability, and ifx1+x4=x2+x3x_{1}+x_{4}=x_{2}+x_{3}, all 8 formulas verify the minimum property.

In the case ofn=4n=4, of the 16 formulas (23) belonging to table 1 , at least 11 and at most 14 verify the minimum property.

It would be interesting to see, fornnany, what is the minimum number(2n)(\geqslant 2n)and what is the maximum number (<2n<2^{n}ifn>3n>3) of consecutive permutations to the initial permutation, which can verify the minimum property.

For a given configuration of nodes, the number of formulas that verify the minimum property can be easily determined. To determine this number, we do not need to examine the monotonicity of the sequence (28) for all values ​​of 1uixxfor reasons of continuity, only those values ​​ofxxfor the house of the maker. This comes back to the fact that we have to examine the values ​​(in finite number) ofxx, for which we have

|xxR|=|xxS|(RS),\left|x-x_{r}\right|=\left|x-x_{s}\right|\quad(r\neq s),

equality that can only occur if

x=xR+xS2x=\frac{x_{r}+x_{s}}{2}

It is therefore sufficient to examine only the pointsxxwhich coincide with the means of the segments determined by two nodes.

In particular, if the nodes (22) are equidistant, the pointsxxWhat intervene are the nodes themselves as well as the means of the segments determined by two consecutive nodes.

A detailed enumeration, which we will not reproduce, leads us to the result that if the nodes are equidistant, the number of formulas (23) belonging to table 1, which verify the minimum property, is equal to

4.2[n2]+2.3[n+12]2n64.2^{\left[\frac{n}{2}\right]}+2.3^{\left[\frac{n+1}{2}\right]}-2n-6

where[α][\alpha]is the whole contained inα\alpha.
To apply the interpolation formula, at pointxxwe will choose that formula (23) that verifies the property that the sequence (28) is non-decreasing.

Let us suppose, in particular, that the nodes are symmetrical about a point on the real axis. This point may be the origin 0. We must then distinguish two cases, according to whether the nodes are odd or even in number. If we suppose that we interpolate in the vicinity of the origin, we must also distinguish two cases according to whether the pointxxis it to the left or to the right of himAO.

Case I. Odd number of nodes. The nodes in ascending order can then be written (n=2kn=2k)

tk,tk1,,t1,0,t1,t2,,tk,-t_{k},-t_{k-1},\ldots,-t_{1},0,t_{1},t_{2},\ldots,t_{k},

and the permutations that verify the minimum property correspond to the following permutations of the nodes:

0,t1,t1,t2,t2,,tk,tk,0,t1,t1,t2,t2,,tk,tk,\begin{gathered}0,-t_{1},t_{1},-t_{2},t_{2},\ldots,-t_{k},t_{k},\\ 0,t_{1},-t_{1},t_{2},-t_{2},\ldots,t_{k},-t_{k},\end{gathered}

asxxis it to the left or to the right of himAOIf
xxis sufficiently close to the origin, these are the only permutations that verify the minimum property.

The corresponding interpolation formulas are well-known generalizations of some of Euler's formulas, to which they reduce when the nodes are equidistant. By taking the arithmetic mean of these formulas, a well-known generalization of Stirling's interpolation formula is deduced (it is unnecessary to write these formulas explicitly here).

Case II. Even number of nodes. The nodes in ascending order can then be written (n=2k1n=2k-1)

tk,tk1,t1,t1,t2,,tk,-t_{k},-t_{k-1},\ldots-t_{1},t_{1},t_{2},\ldots,t_{k},

and the permutations that verify the minimum property correspond to the following permutations of the nodes:

t1,t1,t2,t2,,tk,tk\displaystyle-t_{1},t_{1},-t_{2},t_{2},\ldots,-t_{k},t_{k}
t1,t1,t2,t2,,tk,tk\displaystyle t_{1},-t_{1},t_{2},-t_{2},\ldots,t_{k},-t_{k}

asxxis to the left or to the right of 0. These permutations are the only ones that verify the minimum property ifxxis close enough to the origin.

The corresponding interpolation formulas are well-known generalizations of other Euler formulas, to which they reduce if the nodes are equidistant. Taking the arithmetic mean of these two formulas, we obtain a well-known generalization of Bessel's interpolation formula (it is unnecessary to write these formulas explicitly).

Proprietatea de minimum pe care am stabilit-o constituie o justificare a utilizării acestor formule, în cazul cînd se interpolează în mijlocul tabloului diferentelor divizate.
4. Ne ocupăm în continuare cu cazul particular important al nodurilor echidistante, studiat în lucrarea [3]. Atunci printr-o transformare simplă, problema de interpolare se pune sub forma egalităţii aproximative.

f(a+xh)n=0rx(x1)(xv+1)v!Δhvf(a),f(a+xh)\approx\sum_{n=0}^{r}\frac{x(x-1)\ldots(x-v+1)}{v!}\Delta_{h}^{v}f(a), (29)

unde coeficienţii Δkvf(a)\Delta_{k}^{v}f(a) se calculează cu ajutorul tabloului de diferențe ale valorilor f(a+vh),v=0,1,,nf(a+vh),v=0,1,\ldots,n, ale functiei f(x)f(x) pe noduri.

Calculul valorii polinomului

L(x)=v=0nx(x1)(xv+1)v!Δkvf(a),L(x)=\sum_{v=0}^{n}\frac{x(x-1)\ldots(x-v+1)}{v!}\Delta_{k}^{v}f(a), (30)

pentru o valoare dată a lui xx, se face pe baza schemei

yv+1=Δhnvf(a)+xn+vnv+1yv,v=0,1,,n,(y0=0).y_{v+1}=\Delta_{h}^{n-v}f(a)+\frac{x-n+v}{n-v+1}y_{v},\quad v=0,1,\ldots,n,\quad\left(y_{0}=0\right). (31)

Atunci

yn+1=L(x)y_{n+1}=L(x) (32)

Într-o astfel de schemă de calcul însă, în mod practic, fiecare număr yv+1y_{v+1} se calculează aproximativ, folosind valorile aproximative deja obţinute ale numerelor precedente y1,,yv1,yvy_{1},\ldots,y_{v-1},y_{v}. Dacă notăm cu y¯v\bar{y}_{v} valoarea aproximativă astfel calculată a lui yvy_{v} și cu cvc_{v} corecţia respectivă, succesiunea de calcule se va face, în loc de (4), pe baza schemei
y¯v+1+cv+1=Δhnvf(a)+xn+vnv+1y¯v,v=0,1,,n(y¯0=0)\bar{y}_{v+1}+c_{v+1}=\Delta_{h}^{n-v}f(a)+\frac{x-n+v}{n-v+1}\bar{y}_{v},\quad v=0,1,\ldots,n\quad\left(\bar{y}_{0}=0\right).

Avem atunci

L(x)=y¯n+1+cL(x)=\bar{y}_{n+1}+c (34)

unde corectia cc este dată de formula

c=ν=0nx(x1)(xν+1)ν!cnν+1c=\sum_{\nu=0}^{n}\frac{x(x-1)\ldots(x-\nu+1)}{\nu!}c_{n-\nu+1} (35)

Dacă eroarea absolută maximă în calculul valorilor aproximative ale numerelor yνy_{\nu} este ε\leqslant\varepsilon, deci dacă

|cv|ε,v=1,2,,n+1\left|c_{v}\right|\leqslant\varepsilon,\quad v=1,2,\ldots,n+1 (36)

avem

|c|εK1(x)|c|\leqslant\varepsilon K_{1}(x) (37)

unde

K1(x)=v=0n|x(x1)(xv+1)v!|K_{1}(x)=\sum_{v=0}^{n}\left|\frac{x(x-1)\ldots(x-v+1)}{v!}\right| (38)

Formula de interpolare (29) este avantajoasă în special cînd xx este aproape de 0 . Această afirmatie este justificată, de ex., de cercetările noastre anterioare asupra utilizării practice a formulelor de interpolare [2]:

Este deci suficient să studiem aici cazul cînd 0<x<10<x<1. Avem atunci

K1(x)=2(1x)(2x)(nx)n!,x(0,1)K_{1}(x)=2-\frac{(1-x)(2-x)\ldots(n-x)}{n!},\quad x\in(0,1) (39)

și se vede că K1(x)K_{1}(x) este crescător în intervalul (0,1)(0,1). Rezultă că

1=K1(0)<K1(x)<K1(1)=2,x(0,1),1=K_{1}(0)<K_{1}(x)<K_{1}(1)=2,\quad x\in(0,1), (40)

deci

|c|<2ε.|c|<2\varepsilon. (41)

Un tablou al valorilor lui K1(x)K_{1}(x) poate fi utilizat în practică pentru obținerea unei delimitări mai bune.

Dăm mai jos tabloul valorilor lui K1(x)K_{1}(x) pentru x=v101x=v\cdot 10^{-1}, v=1,2,9v=1,2,\ldots 9, şi n=2,3,4,5,6n=2,3,4,5,6.

Table 3: Tabloul 3
x\nx\backslash n 2 3 4 5 6
0,1 145 1735 1941625 21027925 2234412625
0,2 28 328 3616 387136 4075648
0,3 405 4645 5046625 53438375 5576636125
0,4 52 584 6256 655552 6785152
0,5 625 6975 7265625 75390625 7744140625
0,6 72 776 8096 832448 8492032
0,7 805 8505 8766625 89392975 9063046125
0,8 88 912 9296 940864 9487488
0,9 945 9615 9701625 97553325 9792032625

În acest tablou valorile sînt calculate exact și nu figurează decît partea lor zecimală. Partea întreagă este evident 1 peste tot. Pentru o valoare a lui xx care nu figurează în tablou, se delimitează superior K1(x)K_{1}(x) prin valoarea sa pentru valoarea imediat superioară a lui xx care figurează în tablou. Este clar că pentru 0,9<x<10,9<x<1 nu putem obține pe această cale o delimitare mai buna decit (41). In practică este suficient să luăm niçte valori aproximative convenabile ale valorilor care figurează în tablou.

Formula (37) dă pentru cc delimitarea

εK1(x)cεK1(x)-\varepsilon K_{1}(x)\leqslant c\leqslant\varepsilon K_{1}(x) (42)

inferioară și superioară. Cînd natura calculului indică anumite condiții suplimentare verificate de corecţiile cνc_{\nu}, o analiză mai amănunțită permite à precizăm aceste delimitări.

Să presupunem că sîntem în cazul în care 0<x<10<x<1 și, ceea ce se întîmplă des în practică, că numerele yvy_{v} sînt pozitive. Aceasta va avea loc aproape totdeauna cînd Δhvf(a),v=0,1,,n\Delta_{h}^{v}f(a),\quad v=0,1,\ldots,n sînt numere pozitive. Pentru a calcula produsul

xn+vnv+1y\frac{x-n+v}{n-v+1}y (43)

se calculează totdeauna întîi o valoare aproximativă, de exemplu prin lipsă, a valorij sale absolute, care în cazul nostru este

|xn+vnv+1|y¯v.\left|\frac{x-n+v}{n-v+1}\right|\bar{y}_{v}. (44)

Acest caz are loc de exemplu dacă se calculează numerele (44) cu un număr oarecare de zecimale exacte.

Dacă presupunem că c1=0c_{1}=0, ceea ce practic este acceptabil, deoarece presupunem aici că valorile funcției pe noduri nu sînt afectate de erori, vedem că în conditiile de mai sus corecțiile c2,c3,,cnc_{2},c_{3},\ldots,c_{n} sînt 0(\leqslant 0( și ε)\geqslant-\varepsilon) iar cn+10(c_{n+1}\geqslant 0( și ε)\leqslant\varepsilon). Vom avea atunci

ε[K1(x)K2(x)]cεK2(x)-\varepsilon\left[K_{1}(x)-K_{2}(x)\right]\leqslant c\leqslant\varepsilon K_{2}(x) (45)

unde Pentru delimitarea lui K2(x)K_{2}(x) observăm că pentru n=2n=2 avem K2(x)=1K_{2}(x)=1, iar pentru n>2n>2 putem scrie

K2(x)=1+xK3(x),K_{2}(x)=1+xK_{3}(x), (47)

unde

K3(x)=v=1[n12](1x)(2x)(2v1x)(2v)!K_{3}(x)=\sum_{v=1}^{\left[\frac{n-1}{2}\right]}\frac{(1-x)(2-x)\ldots(2v-1-x)}{(2v)!} (48)

Funcția K3(x)K_{3}(x) este descrescătoare în intervalu1 ( 0,1 ) şi vom putea utiliza un tablou de valori ale functiei K3(x)K_{3}(x) pentru delimitarea lui K2(x)K_{2}(x) cu ajutorul formulei (47).

Dacă formăm un tablou al valorilor lui K3(x)K_{3}(x), pentru valorile ξv\xi_{v}, v=0,1,,m1(m>1)v=0,1,\ldots,m-1(m>1), ale variabilei, unde

0=ξ0<ξ1<<ξm1<1,ξm=10=\xi_{0}<\xi_{1}<\ldots<\xi_{m-1}<1,\quad\xi_{m}=1 (49)

și dacă xx nu figurează în acest tablou, formula (47) ne dă delimitarea

K2(x)<1+xK3(ξi),K_{2}(x)<1+xK_{3}\left(\xi_{i}\right), (50)

presupunînd că ξi<x<ξi+1\xi_{i}<x<\xi_{i+1}.

Trebuie să observăm că din (13) și din faptul că K1(x)K2(x)>0K_{1}(x)-K_{2}(x)>0 rezultă că

K2(x)<2,K_{2}(x)<2, (51)

aşa că pentru ξi<x<ξi+1\xi_{i}<x<\xi_{i+1}, delimitarea (50) nu ne va da o delimitare mai bună decit delimitarea imediata (51), decit daca xK3(ξ1)<1xK_{3}\left(\xi_{1}\right)<1. Deci price 0<x<10<x<1 pună decît (51) pentru orice 0<x<10<x<1, este necesar şi suficient ca să avem

ξi+1K3(ξi)<1,i=0,1,,m1.\xi_{i+1}K_{3}\left(\xi_{i}\right)<1,\quad i=0,1,\ldots,m-1. (52)

În particular, prima condiţie (52) se scrie

ξ1v==1[n12]12v<1,\xi_{1}\sum_{v==1}^{\left[\frac{n-1}{2}\right]}\frac{1}{2v}<1, (53)

care, din cauza divergenței seriei armonice, nu este verificată pentru nn destul de mare. Dacă deci numerele ξv,ν=0,1,,m1\xi_{v},\nu=0,1,\ldots,m-1, sînt date, conditii sînt însă verifica, condițille (52) nu sînt toate verificate. Aceste intervalul (0,1)(0,1) în 10 če daca, de exemplu, ξv\xi_{v} sînt punctele care împart lui ξ\xi_{\text{v }} si nn avem următori ale lu ξv\xi_{v} și nn avem urmatorul tablou de valori ale funcţiei K3(x)K_{3}(x).

Table 4: Tabloul 4
xx 3,5 5,6 7,8
0,0 5 75 916¯91\overline{6}
0,1 45 656625 788245125
0,2 4 568 670144
0,3 35 483875 561477875
0,4 3 404 461408
0,5 25 328125 369140625
0,6 2 256 2839253¯283925\overline{3}
0,7 15 187375 205053375
0,8 1 122 131856
0,9 05 059625 0637027916¯063702791\overline{6}

În acest tablou valorile sînt calculate exact și nu figurează decît partea lor zecimală. Partea intreagă este peste tot egală cu 0 . Pentru a delimita pe K2(x)K_{2}(x) cu ajutorul formulei (47) pentru o valoare a lui xx care nu figurează în tablou, se ia din tablou valoarea lui K3(x)K_{3}(x), pentru valoarea lui xx imediat inferioară și care figurează în tablou. Se verifică că condițiile (52) sînt satisfăcute în aceste cazuri. În practică se pot lua bineînțeles și aici niste valori aproximative prin adaos calorilor din tablou.

Pentru n=3,4n=3,4, în multe cazuri se poate obține destul de uşor direct, delimitarea lui K2(x)K_{2}(x) cu ajutorul formulei

K2(x)=1+x(1x)2=(1+x)(2x)2.K_{2}(x)=1+\frac{x(1-x)}{2}=\frac{(1+x)(2-x)}{2}. (54)

Pentru delimitarea numărului K1(x)K2(x)K_{1}(x)-K_{2}(x) observăm că el este egal cu xx pentru n=2,3n=2,3, iar pentru n>3n>3 se poate scrie

K1(x)K2(x)=xK4(x),K_{1}(x)-K_{2}(x)=xK_{4}(x), (55)

unde

K4(x)=v1[n22](1x)(2x)(2vx)(2v+1)!K_{4}(x)=\sum_{v-1}^{\left[\frac{n-2}{2}\right]}\frac{(1-x)(2-x)\ldots(2v-x)}{(2v+1)!} (56)

este o funcţie descrescătoare în intervalul ( 0,1 ). Se pot face observații analoage cu cele de mai sus, relativ la utilizarea unui tablou al valorilor lui K4(x)K_{4}(x) pentru delimitarea lui K1(x)K2(x)K_{1}(x)-K_{2}(x) prin formula (55). Din formula (40) şi din K2(x)>1K_{2}(x)>1 rezultă

K1(x)K2(x)<1,x(0,1)K_{1}(x)-K_{2}(x)<1,\quad x\in(0,1) (57)

şi condițiile pentru ca tabloul să dea o delimitare mai bună decît (57) pentru ξi<x<ξi+1\xi_{i}<x<\xi_{i+1} sînt ca xK4(ξi)<1xK_{4}\left(\xi_{i}\right)<1. Aici ξv\xi_{v} sînt numerele (49). Dacă i=m1i=m-1, această condiție cu siguranță nu este verificată pentru orice xx, deoarece K4(ξm1)>1K_{4}\left(\xi_{m-1}\right)>1. In cazul acesta nu poate fi deci vorba decit ca tabloul să dea o delimitare mai bună decît (57) pentru orice 0<xξm10<x\leqslant\xi_{m-1} şi orice ξm1x1K4(ξm1)\xi_{m-1}\leqslant x\frac{1}{K_{4}\left(\xi_{m-1}\right)}The necessary and sufficient conditions for this to be the case are written as

ξand+1K4(ξand)<1,and=0,1,,m2.\xi_{i+1}K_{4}\left(\xi_{i}\right)<1,\quad i=0,1,\ldots,m-2. (58)

In Table 5 we give the values ​​ofK4(x)K_{4}(x)for his valuesxxwhich divide the interval(0,1)(0,1)into 10 equal parts and for4n94\leqslant n\leqslant 9.

Table 5: Table 5
nx{}_{x}{}^{n} 4.5 6.7 8.9
0.0 3¯\overline{3} 53¯5\overline{3} 6761804¯\overline{6761804}
0.1 285 4461675 5571044625
0.2 24 38288 4675136
0.3 1983 29740083¯2974008\overline{3} 36059174583¯3605917458\overline{3}
0.4 16 23488 2808064
0.5 125 1796875 2119140625
0.6 093¯09\overline{3} 131413¯13141\overline{3} 15295573
0.7 065 0896675 1030525553571428¯103052555357142\overline{8}
0.8 04 05408 0614016
0.9 0183¯018\overline{3} 02430083¯0243008\overline{3} 02727179583

This table shows the exact decimal parts of the function values.K4(x)K_{4}(x). The integer part is everywhere equal to 1. The bound ofK1(x)K_{1}(x)K2(x)K_{2}(x)with the help of (55), using this table, it is done in the same way as it was done to delimit itK2(x)K_{2}(x)with the help of table 2.

Conditions (58) are fulfilled in these cases. As shown above, Table 5 gives a better delimitation than (57) for

{0<x<0,981996(n=4,5)0<x<0,976275(n=6,7)0<x<0,973452(n=8,9)}\left\{\begin{array}[]{l}0<x<0,981996\ldots(n=4,5)\\ 0<x<0,976275\ldots(n=6,7)\\ 0<x<0,973452\ldots(n=8,9)\end{array}\right\}

respectively.
5. The previous results can be extended to the case of interpolation by polynomials of two variables [4].

Let's consider(m+1)(n+1)(m+1)(n+1)puncture(xand,yj),and=1,2,,m+1\left(x_{i},y_{j}\right),i=1,2,\ldots,m+1,j=1,2,,n+1j=1,2,\ldots,n+1, in the plane, forming a rectangular network of nodes'. For the purpose of fixing the notations we will assume thatx1<x2<<xm+1x_{1}<x_{2}<\ldots<x_{m+1},y1<y2<<yn+1y_{1}<y_{2}<\ldots<y_{n+1}PermutationP(R1,R2,,Rm+1;S1,S2,,Sn+1)P\left(r_{1},r_{2},\ldots,r_{m+1};s_{1},s_{2},\ldots,s_{n+1}\right)of the network of nodes is defined by a permutationR1,R2,,Rm+1r_{1},r_{2},\ldots,r_{m+1}of indices1,2,,m+11,2,\ldots,m+1, of a permutationS1,S2,,Sn+1s_{1},s_{2},\ldots,s_{n+1}of indices1,2,,n+11,2,\ldots,n+1and renumbering the coordinates of the network nodes, so that we havexand=xRand,and=1,2,,m+1,y=ySj,j=1,2,,n+1x_{i}^{\prime}=x_{r_{i}},i=1,2,\ldots,m+1,y^{\prime}=y_{sj},j=1,2,\ldots,n+1.

Let's introduce the notations

DV,μand,j[f]=[xV,xV+1,,xV+andyμ,yμ+1,,yμ+j;f]D_{v,\mu}^{i,j}[f]=\left[\begin{array}[]{llll}x_{v}^{\prime},&x_{v+1}^{\prime},&\ldots,&x_{v+i}^{\prime}\\ y_{\mu}^{\prime},&y_{\mu+1}^{\prime},&\ldots,&y_{\mu+j}^{\prime}\end{array};f\right]

for the differences divided by the points(xV+α,yη+β),α=0,1,,and\left(x_{v+\alpha}^{\prime},y_{\eta+\beta}^{\prime}\right),\alpha=0,1,\ldots,i,β=0,1,,j\beta=0,1,\ldots,j, relative to the functionf(x,y)f(x,y).D
==(divided difference or divided differences).
DV,μand,j[f],V=1,2,,m+1and,μ=1,2,,n+1jD_{v,\mu}^{i,j}[f],\quad v=1,2,\ldots,m+1-i,\mu=1,2,\ldots,n+1-j,

and=0,1,,m,j=0,1,,ni=0,1,\ldots,m,j=0,1,\ldots,n (61)

constitutes the dd array corresponding to the permutation
of the network of nodes.

P(R1,R2,,Rm+1;S1,S2,,Sn+1)P\left(r_{1},r_{2},\ldots,r_{m+1};\quad s_{1},s_{2},\ldots,s_{n+1}\right)

The system(m+1)(n+1)(m+1)(n+1)d.d.

D1,1and,j[f],and=0,1,,m;j=0,1,,n,D_{1,1}^{i,j}[f],i=0,1,\ldots,m;j=0,1,\ldots,n, (62)

forms a chained system of dd
If each term of a chained system of dd belongs to the same array of dd we will say that the chained system considered belongs to this array of dd In particular, the chained system (€2) belongs to array 2. Of course, the chained system1 (62) belongs at the same time to other arrays of dd

In particular, tableau 2, corresponding to the initial permutationP(1,2,,m+1;2,,n+1)P(1,2,\ldots,m+1;2,\ldots,n+1)of the network of nodes, will be called
the normal array of dd and any linked system of dd belonging to this array will be called a normal linked system of dd. In the case of the initial permutation we have

Rand=and,Sj=j,DV,μand,j[f]=[xV,xV+1,,xV+andyμ,yμ+1,,yμ+j]r_{i}=i,s_{j}=j,D_{v,\mu}^{i,j}[f]=\left[\begin{array}[]{c}x_{v},x_{v+1},\ldots,x_{v+i}\\ y_{\mu},y_{\mu+1},\ldots,y_{\mu+j}\end{array}\right]

for all possible values ​​ofand,j,n,μi,j,\nu,\mu.

PermutationsP(R1,R2,,Rm+1;S1,S2,,Sn+1)P\left(r_{1},r_{2},\ldots,r_{m+1};s_{1},s_{2},\ldots,s_{n+1}\right)of the network of nodes, corresponds to the interpolation polynomial of two variables, which in its general form is [2]

IT(x,y)=\displaystyle L(x,y)=
=and=0mj=0nand(xxR1)(xxR3)(xxRand)(yyS1)(yyS2)(yySj)D1,1and,j[f]\displaystyle=\sum_{i=0}^{m}\sum_{j=0}^{n_{i}}\left(x-x_{r_{1}}\right)\left(x-x_{r_{3}}\right)\ldots\left(x-x_{r_{i}}\right)\left(y-y_{s_{1}}\right)\left(y-y_{s_{2}}\right)\ldots\left(y-y_{s_{j}}\right)\quad D_{1,1}^{i,j}[f] (63)

where0=nandn,and=0,1,,m0=n_{i}\leqslant n,i=0,1,\ldots,m, and forand=0i=0respectivelyj=0j=0the product(xxR1)(xxR2)(xxRand)\left(x-x_{r_{1}}\right)\left(x-x_{r_{2}}\right)\ldots\left(x-x_{r_{i}}\right)respectively(yyS1)(yyS2)(yySj)\left(y-y_{s_{1}}\right)\left(y-y_{s_{2}}\right)\ldots\left(y-y_{s_{j}}\right)is replaced by 1 .

On a given interpolation point (x,yx,y), determining an approximate value forf(x,y)f(x,y)with the help of the polynomial (63) requires the calculation of the numerical value of this polynomial. This calculation is made by executing, exactly or approximately, a certain sequence of elementary operations: additions, subtractions, multiplications and divisions. Since the operations are generally executed approximately, a calculation error will result that depends on the precision with which the calculations were executed as well as, of course, on the order of performing these calculations, that is, on the program on the basis of which they were executed. We will not take into account the errors that may affect the data of the problem, so the coordinates of the nodes and the values ​​of the functionf(x,y)f(x,y)on these nodes.

To calculate the numerical value of the polynomial (63), it may also be useful to modify its expression, putting

X0=1,Xand=xxRandkand,Y0=1,Y=yySanditjX_{0}=1,X_{i}=\frac{x-x_{r_{i}}}{k_{i}},Y_{0}=1,\quad Y=\frac{y-y_{s_{i}}}{l_{j}}

wherek0(=1),k1,k2,,km,it0(=1),it1,it2,,itnk_{0}(=1),k_{1},k_{2},\ldots,k_{m},l_{0}(=1),l_{1},l_{2},\ldots,l_{n}are some given numbers, different from 0. We then have

IT(x,y)=and=0mj=0nandX0X1XandY0Y1YjIt isand,jL(x,y)=\sum_{i=0}^{m}\sum_{j=0}^{n_{i}}X_{0}X_{1}\ldots X_{i}Y_{0}Y_{1}\ldots Y_{j}E_{i,j}

which is calculated, usually by adding up

and=0mX0X1XandAand\sum_{i=0}^{m}X_{0}X_{1}\ldots X_{i}A_{i} (64)

having previously calculated the amounts

Aand=j=0nandY0Y1YandIt isand,j,and=0,1,,mA_{i}=\sum_{j=0}^{n_{i}}Y_{0}Y_{1}\ldots Y_{i}E_{i,j},i=0,1,\ldots,m (65)

Each of the sums (64), (65) is of the form

c0C0+c0c1C1++c0c1CαCα.c_{0}C_{0}+c_{0}c_{1}C_{1}+\ldots+c_{0}c_{1}\ldots C_{\alpha}C_{\alpha}. (66)

Let us assume that additions and subtractions are error-free (they are performed exactly, or, in any case, with some negligible errors). Then if we calculate the sum (66) based on the scheme

c0(C0+c1(C1++cα1(Cα1+cαCα)))c_{0}\left(C_{0}+c_{1}\left(C_{1}+\ldots+c_{\alpha-1}\left(C_{\alpha-1}+c_{\alpha}C_{\alpha}\right)\ldots\right)\right) (67)

as is usually done, the error comes only from performing theα+1\alpha+1successive multiplicationswithcα,cα1,,c0\mathrm{cu}c_{\alpha},c_{\alpha-1},\ldots,c_{0}respectively. It is easy to see that if these multiplications are performed with an absolute error at most equal toε\varepsilon, the approximate result obtained will have an absolute error smaller than

ε(1+and=0α1|c0c1cand|)(=ε if α=0).\varepsilon\left(1+\sum_{i=0}^{\alpha-1}\left|c_{0}c_{1}\ldots c_{i}\right|\right)\quad(=\varepsilon\text{ dacă }\alpha=0).

In the cases that we will consider effectively, multiplication byc0c_{0}does not contain errors (we will always havec0=1c_{0}=1). In this case the absolute error of the result is smaller than

εand=0α1|c0,c1,,cand|(=0, if α=0).\varepsilon\sum_{i=0}^{\alpha-1}\left|c_{0},c_{1},\ldots,c_{i}\right|\quad(=0,\text{ dacă }\alpha=0).

Let us apply this calculation scheme to the sums (64) (65). The sum (65) is of the form (66), where we put

α=nand,cV=YV,CV=It isand,VV=0,1,,nand.\alpha=n_{i},c_{v}=Y_{v},C_{v}=E_{i,v}\quad v=0,1,\ldots,n_{i}.

So, if we perform the calculations based on the scheme (67), the multiplications being performed with an absolute errorε\leqslant\varepsilon, the result obtainedA~and\widetilde{A}_{i}is with an absolute errorεj=0nand1|Y0Y1Yj|(=0\leqslant\varepsilon\sum_{j=0}^{n_{i}-1}\left|Y_{0}Y_{1}\ldots Y_{j}\right|\quad\left(=0\right., ifnand=0)\left.n_{i}=0\right)It follows that the approximation of the sum (64),

and=0mX0X1XandA~and,\sum_{i=0}^{m}X_{0}X_{1}\ldots X_{i}\widetilde{A}_{i}, (68)

has an absolute errorεand=0m|X0X1Xand|(j=0nand1|Y0Y1Yj|)\leqslant\varepsilon\sum_{i=0}^{m}\left|X_{0}X_{1}\ldots X_{i}\right|\left(\sum_{j=0}^{n_{i}-1}\left|Y_{0}Y_{1}\ldots Y_{j}\right|\right).

The sum (68) is also of the form

α=m,cV=XV,CV=A~V,V=0,1,,m.\alpha=m,c_{v}=X_{v},C_{v}=\widetilde{A}_{v},\quad v=0,1,\ldots,m.

So, if here too we perform the calculations based on scheme (67), the multiplications being performed with an absolute errorε\leqslant\varepsilon, the approximate value obtained for (68) will have an absolute error at most equal toεand=0n1|X0X1Xand|\varepsilon\sum_{i=0}^{n-1}\left|X_{0}X_{1}\ldots X_{i}\right|(=0=0, ifm=0m=0).

Ultimately, therefore, making the calculations based on the indicated scheme, we obtain forIT(x,y)L(x,y)an approximate value with an absolute errorεM\leqslant\varepsilon M, where

M=and=0m|X0X1Xand|(j=0nand1|Y0Y1Yj|)+and=0m1|X0X1Xand|.M=\sum_{i=0}^{m}\left|X_{0}X_{1}\ldots X_{i}\right|\left(\sum_{j=0}^{n_{i}-1}\left|Y_{0}\quad Y_{1}\ldots Y_{j}\right|\right)+\sum_{i=0}^{m-1}\left|X_{0}X_{1}\ldots X_{i}\right|. (69)

We can ask ourselves the problem of finding, for the given interpolation point,(x,y)(x,y), that permutationP(R1,R˙2,,Rm+1;S1,S2,,Sn+1)P\left(r_{1},\dot{r}_{2},\ldots,r_{m+1};s_{1},s_{2},\ldots,s_{n+1}\right)of the network of nodes, for which the numberMMgiven by formula (69) is the smallest possible. If this condition is met, we will consider that polynomial (63) is the most advantageous for numerical calculations (made based on the indicated scheme).

Both the amounts

Band=j=0nand1|Y0Y1Yj|,B_{i}=\sum_{j=0}^{n_{i}-1}\left|Y_{0}Y_{1}\ldots Y_{j}\right|, (70)

and the sum (69) are of the form

S0+SV1S1+SV1SV2S2++SV1SV2SVαSα,S_{0}+s_{v_{1}}S_{1}+s_{v_{1}}s_{v_{2}}S_{2}+\ldots+s_{v_{1}}s_{v_{2}}\ldots s_{v_{\alpha}}S_{\alpha}, (71)

whereS0,S1,,SαS_{0},S_{1},\ldots,S_{\alpha}are positive numbers,andSn1,Sn2,,Snα\operatorname{iar}s_{\nu_{1}},s_{\nu_{2}},\ldots,s_{\nu_{\alpha}}is a permutation of a stringS1,S2,,Sαs_{1},s_{2},\ldots,s_{\alpha}of positive numbers.

The string (70) is of the form (71), where
α=nand1,Sj=|yyj|,Vand=Sj,Sj=1|it0it1itj|,j=1,2,,nand1,S0=1\alpha=n_{i}-1,s_{j}=\left|y-y_{j}\right|,v_{i}=s_{j},S_{j}=\frac{1}{\left|l_{0}l_{1}\cdots l_{j}\right|},j=1,2,\ldots,n_{i}-1,S_{0}=1,
and the sum (69) is of the form (71), where
α=m,Sand=|xxand|,Vand=Rand,Sand=1|k0k1kand|(1+j=0nand1|Y0Y1Yj|);\alpha=m,s_{i}=\left|x-x_{i}\right|,v_{i}=r_{i},S_{i}=\frac{1}{\left|k_{0}k_{1}\ldots k_{i}\right|}\left(1+\sum_{j=0}^{n_{i}-1}\left|Y_{0}Y_{1}\ldots Y_{j}\right|\right);

and=0,1,,m1,Sm=1|k0k1km|(j=0nm1|Y2Y1Yj|),Vm=Rmi=0,1,\ldots,m-1,S_{m}=\frac{1}{\left|k_{0}k_{1}\ldots k_{m}\right|}\left(\sum_{j=0}^{n_{m}-1}\left|Y_{2}Y_{1}\ldots Y_{j}\right|\right),v_{m}=r_{m}

We now have the following
Lemma. The sum (71) reaches its smallest value if the sequenceSV1,SV2,,SVαs_{v_{1}},s_{v_{2}},\ldots,s_{v_{\alpha}}is non-decreasing (in other words for a non-decreasing permutation of the stringS1,S2,,Sαs_{1},s_{2},\ldots,s_{\alpha}).

The proof is immediate. It is easy to see that the minimum is reached only for non-decreasing permutations of the sequenceS1,S2,,Sαs_{1},s_{2},\ldots,s_{\alpha}.

From the previous lemma the following
Theorem then follows. If the permutationP(R1,R2,,Rm+1;S1,S2,,Sn+1)P\left(r_{1},r_{2},\ldots,r_{m+1};s_{1},s_{2},\ldots,s_{n+1}\right)is such that the strings
|xxR1|,|xxR2|,,|xxRm|;|yyS1|,|yyS3|,,|yySn1|\left|x-x_{r_{1}}\right|,\left|x-x_{r_{2}}\right|,\ldots,\left|x-x_{r_{m}}\right|;\left|y-y_{s_{1}}\right|,\left|y-y_{s_{3}}\right|,\ldots,\left|y-y_{s_{n-1}}\right|
to be non-decreasing, the numberMMreaches its lowest value.

It follows from the preceding that if the strings

|xxR1|,|xxR2|,,|xxRm+1|;|yyS1|,|yyS2|,,|yySn+1|\left|x-x_{r_{1}}\right|,\left|x-x_{r_{2}}\right|,\ldots,\left|x-x_{r_{m+1}}\right|;\left|y-y_{s_{1}}\right|,\left|y-y_{s_{2}}\right|,\ldots,\left|y-y_{s_{n+1}}\right| (72)

are non-decreasing, the polynomial (63) is the most advantageous for calculating an approximate value off(x,y)f(x,y)In this case, based on a previous result [2], the permutationR1,R2,,Rm+1r_{1},r_{2},\ldots,r_{m+1}must be consecutive to the permutation1,2,,m+11,2,\ldots,m+1, and the permutationS1,S2,,Sn+1s_{1},s_{2},\ldots,s_{n+1}consecutive to permutation1,2,,n+11,2,\ldots,n+1This means that for anyt(t=1,2t(t=1,2,,m+1\ldots,m+1, respectivelyt=1,2,,n+1)t=1,2,\ldots,n+1)stringsR1,R2,,Rtr_{1},r_{2},\ldots,r_{t}andS1,S2s_{1},s_{2}, …S1s_{1}, are made up ofttconsecutive natural numbers (in a certain order).

When the sequences (72) are non-decreasing and, in this case, the polynomial (63) is advantageous by calculations in the sense shown above, the corresponding chained system of dd is a normal chained system. It is therefore sufficient to use the normal tableau of dd to be able to construct all the advantageous polynomials (63), according to the different positions of the interpolation point.

We can, as in the case of a single variable [2], study various particular cases. If the interpolation point is in the vicinity of one of the extreme nodes(x1,y1),(x1,yn+1),(xm+1,y1),(xm+1,xn+1)\left(x_{1},y_{1}\right),\left(x_{1},y_{n+1}\right),\left(x_{m+1},y_{1}\right),\left(x_{m+1},x_{n+1}\right), we find that the most advantageous polynomials ordered by ascending and descending differences with respect tox,yx,y, so those for which the chained system (62) corresponds to the permutationsP(1,2,,m+1;1,2,,n+1)P(1,2,\ldots,m+1;1,2,\ldots,n+1),P1,2,,m+1;n+1,n,,1),P(m+1,m,,1;1,2,,n+1)P1,2,\ldots,m+1;n+1,n,\ldots,1),P(m+1,m,\ldots,1;1,2,\ldots,n+1),P(m+1,m,,1;n+1,n,,1)P(m+1,m,\ldots,1;n+1,n,\ldots,1)If the pointsxandx_{i}andyjy_{j}are respectively equidistant, so if

xand=x1+(and1)h,and=1,2,,m+1,yj=y1+(j1)h\displaystyle x_{i}=x_{1}+(i-1)h,i=1,2,\ldots,m+1,y_{j}=y_{1}+(j-1)h^{\prime}
j=1,2,,n+1(h,h>0)\displaystyle j=1,2,\ldots,n+1\quad\left(h,h^{\prime}>0\right) (73)

and if the interpolation point is in the vicinity of the center of the network of nodes, we find that the most advantageous polynomials (63) are those ordered by the central differences with respect toxxandyy.

Finally, we note that instead of the normal array of dd, we can use the array formed with the numbersk0k1kjit0it1itjDV,μand,j[f]k_{0}k_{1}\ldots k_{j}l_{0}l_{1}\ldots l_{j}D_{v,\mu}^{i,j}[f], dd being related to the initial permutation of the network of nodes. If we takekand=andh,itj=jhk_{i}=ih,l_{j}=jh^{\prime}, for anythingand,j1i,j\geqslant 1, the numbers considered are reduced to the differencesΔh,hand,andf(x,y)\Delta_{h,h^{\prime}}^{i,i}f(x,y)of the functionf(x,y)f(x,y)The normal table of differences can then be replaced by the table of differences, the formation of which is particularly simple, since it only requires successive subtractions.

Finally, the previous considerations can be extended to the case of more than two independent variables.

BIBLIOGRAPHY

  1. 1.

    Popoviciu T, On the precision of numerical calculation in interpolation by polynomials. Bull. scient. - Acad. RP R., Section of mathematical and physical sciences, VIII, 954-961 (1955).

  2. 2.
    • Theoretical considerations on the practical use of some interpolation formulas. Scientific Bulletin - Acad. RPR, Section of Mathematical and Physical Sciences, III, 441-449 (1951).

  3. 3.
    • Desfre precision of numerical calculation in interpolation by Newton's polynomial with equidistant nodes. Scientific Studies and Research, Ser. I, VI, 3-4, 27-35 (1955).

  4. 4.
    • On the precision of numerical calculation in interpolation by polynomials of two variables. Studies and research in mathematics (Cluj), XI, annex fasciculă anexă, 159-165 (1960).

  5. 5.

    Steffensen JF, Interpolation, New York, 1950.

1962

Related Posts