Tensor Product Spline Surfaces

Spline in the x-direction:
$$s_1=\sum _{i=1}^{n_1}{c}_i{\varphi }_i(x)$$

Let $c_i$ be the spline in the y-direction:
$$s_2=c_i(y)=\sum _{j=1}^{n_2}{c}_{i,j}{\phi }_j(y)$$

Then, combine $s_1$ and $s_2$, we get the function of a Tensor Product Spline Surface:
$$f(x,y)=\sum _{i=1}^{n_1}\sum _{j=1}^{n_2}{c}_{i,j}{\varphi }_i(x){\phi }_j(y)$$

where the coefficients $(c_{i,j}{)}_{i,j=1}^{n_1,n_2}$ can be real numbers. This linear space of functions is denoted $s_1\otimes s_2$.

The space $s_1\otimes s_2$ is spanned by the functions ${{\varphi }_i(x){\phi }_j(y)}_{i,j=1}^{n_1,n_2}$ and therefore has dimension $n_1{n}_2$.

Tensor product surface in the matrix form:
$$f(X,Y)=\varphi (X{)}^{T}C\phi (Y)$$

where $\varphi =({\varphi }_1,{\varphi }_2,\dots ,{\varphi }_{n_1}{)}^T$, $\phi =({\phi }_1,{\phi }_2,\dots ,{\phi }_{n_2}{)}^T$, and $C=(c_{i,j})$ is the matrix of coefficients.

Advantages over other ways constructing surfaces:

Many standard operations that we wish to perform with the surfaces are very simple generalisations of corresponding univariate operations.

The approximation methods that we developed for functions and curves can be utilized directly from approximation for surfaces.

knots with multiplicity

• Each knot of multiplicity $k$ reduces at most $k-1$ basis functions’ non-zero domain.
• At each internal knot of multiplicity $k$, the number of non-zero basis functions is at most $p-k+1$, where $p$ is the degree of the basis functions.

Tensor Product Spline Interplotion

$$\sum _{p=1}^{m_1}\sum _{q=1}^{m_2}{c}_{p,q}{\varphi }_q(y_j){\phi }_p(x_i)=g(x_i,y_j)=f_{i,j}$$

This double sum can be split into two sets of simple sums:
$$\begin{array}{rl}\sum _{p=1}^{m_1}{d}_{p,j}{\varphi }_p(x_i)& =f_{i,j}\\ \sum _{q=1}^{m_2}{c}_{p,q}{\phi }_q(y_j)& =d_{p,j}\end{array}$$

We define the matrices:
$$\begin{array}{rl}\Phi & =({\varphi }_{i,p})\in {\Re }^{m_1,m_2},{\varphi }_{i,p}={\varphi }_p(x_i),\\ \Psi & =({\phi }_{j,q})\in {\Re }^{m_1,m_2},{\phi }_{j,p}={\phi }_q(y_j),\\ D& =(d_{p,j})\in {\Re }^{m_1,m_2},F=(f_{i,j})\in {\Re }^{m_1,m_2}\\ C& =(c_{p,q})\in {\Re }^{m_1,m_2}\end{array}$$

then
$$\begin{array}{rl}\sum _{p=1}^{m_1}{d}_{p,j}{\varphi }_p(x_i)& =\sum _{p=1}^{m_1}{\varphi }_{i,p}{d}_{p,j}=(\Phi D{)}_{ij}=(F{)}_{ij}\\ \sum _{q=1}^{m_2}{c}_{p,q}{\phi }_q(y_j)& =\sum _{q=1}^{m_2}{\phi }_{j,p}{c}_{p,q}=(\Psi C^T{)}_{jp}=(D^T{)}_{jp}\end{array}$$

Thus:
$$\Phi D=FandC{\Psi }^T=D$$

If the matrices $\Phi$ and $\Psi$ are nonsigular, there is a unique $g\in s_1\otimes s_2$ such that $g(x_i,y_j)=f_{i,j}$, for i ∈ { 1 , … , m 1 } i \in \{1,\dots,m_{1}\} i∈{1,…,m1​}, j ∈ { 1 , … , m 2 } j \in \{1,\dots,m_{2}\} j∈{1,…,m2​}.

Least Square for Gridded Data

Given data
$$(x_i,y_j,f_{i,j}{)}_{i=1,j=1}^{m_1,m_2}$$

positive weights $(w_i{)}_{i=1}^{m_1}$ and $(v_j{)}_{j=1}^{m_2}$, and univariate spline spaces $s_1$ and $s_2$, find a spline surface $g$ in $s_1\otimes s_2$ which solves the minimization problem
$$\underset{g\in s_1\otimes s_2}{\min }\sum _{i=1}^{m_1}\sum _{j=1}^{m_2}{w}_i{v}_j[g(x_i,y_j)-f_{i,j}{]}^2$$

Its matrix form
$$\underset{C\in {\Re }^{n_1,n_2}}{\min }\parallel AC{B}^T-G{\parallel }^2$$

where
$$\begin{array}{rl}A& =(a_{i,p})\in {\Re }^{m_1,n_1}{a}_{i,p}=\sqrt{w_i}{\varphi }_p(x_i)\\ B& =(b_{j,q})\in {\Re }^{m_2,n_2}{b}_{j,q}=\sqrt{v_j}{\phi }_q(y_j)\\ G& =(\sqrt{w_i}\sqrt{v_j}{f}_{i,j})\in {\Re }^{m_1,m_2}\\ C& =(c_{p,q})\in {\Re }^{n_1,n_2}\end{array}$$

Here, the norm $\parallel \cdot {\parallel }^2$ is the Frobenius norm,
$$\parallel E\parallel =(\sum _{i=1}^m\sum _{j=1}^n|e_{i,j}{|}^2{)}^{\frac{1}{2}}$$

for any rectangular $m\times n$ matrix $E=(e_{i,j})$.

${\min }_{C\in {\Re }^{n_1,n_2}}\parallel AC{B}^T-G{\parallel }^2$ always has a solution $C=C^{\ast }$, and the solution is unique iif both matrices $A$ and $B$ have linearly independent columns. The solution $C^{\ast }$ can be found by solving the matrix equation
$$A^{T}A{C}^{\ast }{B}^{T}B=A^{T}GB$$

Parametric Surfaces

A function of two variables $E=h(x,y)$ can always be considered as a parametric surface through the representation $f(u,v)=(u,v,h(u,v))$.

A parametric representation of class $C^m$ of a set $\varsigma \subseteq {\Re }^3$ is a mapping $f$ of an open set $\Omega \subseteq {\Re }^2$ onto $\varsigma$ such that

(1) $f$ has continous derivatives up to order $m$,

(2) the Jacbian Matrix of $f$ given by
$$J(f)=\left(\begin{array}{cc}{D}_u{f}^1(u,v)& D_v{f}^1(u,v)\\ D_u{f}^2(u,v)& D_v{f}^2(u,v)\\ D_u{f}^3(u,v)& D_v{f}^3(u,v)\end{array}\right)$$

has full rank $k$ for all $(u,v)$ is $\Omega$.

The unit normal of the regular parametric representation $f$ is the vector
$$N(u,v)=\frac{D_{u}f(u,v)\times D_{v}f(u,v)}{\parallel D_{u}f(u,v)\times D_{v}f(u,v)\parallel }$$

The normal vector plays an important role when we start analysing the curvature of surfaces.

Let $\varsigma$ be a surface with a regular parametric representation $f$. The tangent space or tangent plane $T_{f(u,v)}$ of $\varsigma$ at $f(u,v)$ is the plane in ${\mathbb{R}}^3$ spanned by the two vectors $D_{u}f(u,v)$ and $D_{v}f(u,v)$.i.e., all vectors on the form ${\delta }_1{D}_{u}f(u,v)+{\delta }_2{D}_{v}f(u,v)$. The normal of the tangent plane $T_{f(u,v)}$ is the normal vector $N(u,v)$.

Parametric Tensor Product Spline Surfaces

A parametric tensor product spline surface is given by a parametric representation on the form
$$f(u,v)=\sum _{i=1}^m\sum _{j=1}^n{c}_{ij}{B}_{i,d,\delta }(u)B_{j,l,\tau }(v)$$

where the coefficients $(c_{ij}{)}_{i,j=1}^{m,n}$ are point in space,
$$c_{ij}=(c_{ij}^1,c_{ij}^2,c_{ij}^3)$$
and $\delta =({\delta }_i{)}_{i=1}^{m+d+1}$ and $\tau =({\tau }_j{)}_{j=1}^{n+l+1}$ are knot vectors for splines of degree $d$ and $l$.