Progressive-iterative approximation method
≈ 21 min read
Names in other languages
No names in other languages are held for this entry yet.
Article bodies are English and are not machine-translated. A name in another language is not a translation of the article.
In mathematics, the progressive-iterative approximation method is an iterative method of data fitting with geometric meanings. Given a set of data points to be fitted, the method obtains a series of fitting curves (or surfaces) by iteratively updating the control points, and the limit curve (surface) can interpolate or approximate the given data points. It avoids solving a linear system of equations directly and allows flexibility in adding constraints during the iterative process. Therefore, it has been widely used in geometric design and related fields. The study of the iterative method with geometric meaning can be traced back to the work of scholars such as Dongxu Qi and Carl de Boor in the 1970s. In 1975, Qi et al. developed and proved the "profit and loss" algorithm for uniform cubic B-spline curves, and in 1979, de Boor independently proposed this algorithm. In 2004, Hongwei Lin and coauthors proved that non-uniform cubic B-spline curves and surfaces have the "profit and loss" property. Later, in 2005, Lin et al. proved that the curves and surfaces with normalized and totally positive basis all have this property and named it progressive iterative approximation (PIA). In 2007, Maekawa et al. changed the algebraic distance in PIA to geometric distance and named it geometric interpolation (GI). In 2008, Cheng et al.…
In mathematics, the progressive-iterative approximation method is an iterative method of data fitting with geometric meanings. Given a set of data points to be fitted, the method obtains a series of fitting curves (or surfaces) by iteratively updating the control points, and the limit curve (surface) can interpolate or approximate the given data points. It avoids solving a linear system of equations directly and allows flexibility in adding constraints during the iterative process. Therefore, it has been widely used in geometric design and related fields. The study of the iterative method with geometric meaning can be traced back to the work of scholars such as Dongxu Qi and Carl de Boor in the 1970s. In 1975, Qi et al. developed and proved the "profit and loss" algorithm for uniform cubic B-spline curves, and in 1979, de Boor independently proposed this algorithm. In 2004, Hongwei Lin and coauthors proved that non-uniform cubic B-spline curves and surfaces have the "profit and loss" property. Later, in 2005, Lin et al. proved that the curves and surfaces with normalized and totally positive basis all have this property and named it progressive iterative approximation (PIA). In 2007, Maekawa et al. changed the algebraic distance in PIA to geometric distance and named it geometric interpolation (GI). In 2008, Cheng et al. extended it to subdivision surfaces and named the method progressive interpolation (PI). Since the iteration steps of the PIA, GI, and PI algorithms are similar and all have geometric meanings, they are collectively referred to as geometric iterative methods (GIM). PIA is now extended to several common curves and surfaces in the geometric design field, including NURBS curves and surfaces, T-spline surfaces, and implicit curves and surfaces. \mathbf{P}(\mathbf{t})=\sum_{i=1}^n\mathbf{P}_iB_i(\mathbf{t}). For example: If \mathbf{P}(\mathbf{t}) is a B-spline curve, then \mathbf{t} is a scalar, B_i(t) is a B-spline basis function, and \mathbf{P}_i denotes the control point; If \mathbf{P}(\mathbf{t}) is a B-spline patch with n_u\times n_v control points, then \mathbf{t}=(u,v) and B_i(\mathbf{t})=N_i(u)N_i(v) , where N_i(u) and N_i(v) are B-spline basis functions; If \mathbf{P}(\mathbf{t}) is a trivariate B-spline solid with n_u \times n_v \times n_w control points, then \mathbf{t}=(u,v,w) and B_i(\mathbf{t})=N_i(u)N_i(v)N_i(w) , where N_i(u) , N_i(v) , and N_i(w) are B-spline basis functions. Additionally, this can be applied to NURBS curves and surfaces, T-spline surfaces, and triangular Bernstein–Bézier surfaces. Given an ordered data set \mathbf{Q}_i with parameters t_i satisfying t_1 for i=1,2,\cdots,n , the initial fitting curve is: \mathbf{P}^{(0)}(t)=\sum_{i=1}^n\mathbf{P}_i^{(0)}B_i(t) where the initial control points of the initial fitting curve \mathbf{P}_i^{(0)} can be randomly selected. Suppose that after the k th iteration, the k th fitting curve \mathbf{P}^{(k)}(t) is generated by To construct the (k+1) st curve, we first calculate the difference vectors, \mathbf{\Delta}^{(k)}_i=\mathbf{Q}_i-\mathbf{P}^{(k)}(t_i), \quad i=1,2,\cdots,n and use them to update the control points by \mathbf{P}_i^{(k+1)}=\mathbf{P}_i^{(k)}+\mathbf{\Delta}_i^{(k)} which leads to the (k+1) st fitting curve: \mathbf{P}^{(k+1)}(t)=\sum_{i=1}^n\mathbf{P}_i^{(k+1)}B_i(t). In this way, we obtain a sequence of curves \mathbf{P}^{(\alpha)}(t),\alpha=0,1,2,\cdots , which converges to a limit curve that interpolates the give data points, i.e., \lim \limits_{\alpha\rightarrow\infty}\mathbf{P}^{(\alpha)}(t_i)=\mathbf{Q}_i, \quad i=1,2,\cdots,n. \mathbf{P}^{(k)}(t)=\sum_{j=1}^n\mathbf{P}_j^{(k)}B_j(t). To generate the (k+1) th fitting curve, first compute the difference vectors for the data points \boldsymbol{\delta}^{(k)}_i=\mathbf{Q}_i-\mathbf{P}^{(k)}(t_i), \quad i=1,2,\cdots,m and then the difference vectors for the control points \mathbf{\Delta}^{(k)}_j=\frac{ \sum_{i \in I_j}{c_i B_j(t_i) \boldsymbol{\delta}_i^{(k)}}}{\sum_{i \in I_j}c_i B_j(t_i)}, \quad j = 1,2,\cdots,n where I_j is the index set of the data points in the j th group, whose parameters fall in the local support of the j th basis function, i.e., B_j(t_i)\ne0 . The c_i are weights that guarantee the convergence of the algorithm, usually taken as c_i = 1, i \in I_j . Finally, the control points of the (k+1) th curve are updated by \mathbf{P}_j^{(k+1)}=\mathbf{P}_j^{(k)}+\mathbf{\Delta}_j^{(k)}, leading to the (k+1) th fitting curve \mathbf{P}^{(k+1)}(t) . In this way, we obtain a sequence of curve, and the limit curve converges to the least-squares fitting result to the given data points. In the local-PIA method, the control points are divided into active and fixed control points, whose subscripts are denoted as I=\left\{i_1,i_2,\cdots,i_I\right\} and J=\left\{j_1,j_2,\cdots,j_J\right\} , respectively. Assume that, the k th fitting curve is \mathbf{P}^{(k)}(t)=\sum_{j=1}^n\mathbf{P}_j^{(k)}B_j(t) , where the fixed control points satisfy \mathbf{P}_j^{(k)}=\mathbf{P}_j^{(0)},\quad j\in J,\quad k=0,1,2,\cdots. Then, on the one hand, the iterative formula of the difference vector \mathbf{\Delta}_h^{(k+1)} corresponding to the fixed control points is \begin{aligned} \mathbf{\Delta}_h^{(k+1)}&=\mathbf{Q}_h-\sum_{j=1}^n\mathbf{P}_j^{(k+1)}B_j(t_h)\\ &=\mathbf{Q}_h-\sum_{j\in J}\mathbf{P}_j^{(k+1)}B_j(t_h)-\sum_{i\in I}\left(\mathbf{P}_i^{(k)}+\mathbf{\Delta}_i^{(k)}\right)B_i(t_h)\\ &=\mathbf{Q}_h-\sum_{j=1}^n\mathbf{P}_j^{(k)}B_j(t_h)-\sum_{i\in I}\mathbf{\Delta}_i^{(k)}B_i(t_h)\\ &=\mathbf{\Delta}_h^{(k)}-\sum_{i\in I}\mathbf{\Delta}_i^{(k)}B_i(t_h), \quad h\in J. \end{aligned} On the other hand, the iterative formula of the difference vector \mathbf{D}_l^{(k+1)} corresponding to the active control points is \begin{aligned} \mathbf{\Delta}_l^{(k+1)}&=\mathbf{Q}_l-\sum_{j=1}^n\mathbf{P}_j^{(k+1)}B_j(t_l)\\ &=\mathbf{Q}_l-\sum_{j=1}^n\mathbf{P}_j^{(k)}B_j(t_l)-\sum_{i\in I}\mathbf{\Delta}_i^{(k)}B_i(t_l)\\ &=\mathbf{\Delta}_l^{(k)}-\sum_{i\in I}\mathbf{\Delta}_i^{(k)}B_i(t_l)\\ &=-\mathbf{\Delta}_{i_1}^{(k)}B_{i_1}(t_l)-\mathbf{\Delta}_{i_2}^{(k)}B_{i_2}(t_l)-\cdots+\left(1-B_l(t_l)\right)\mathbf{\Delta}_l ^{(k)}-\cdots-\mathbf{\Delta}_{i_I}^{(k)}B_{i_I}(t_l),\quad l\in I. \end{aligned} Arranging the above difference vectors into a one-dimensional sequence, \mathbf{D}^{(k+1)}=\left[\mathbf{\Delta}_{j_1}^{(k+1)} ,\mathbf{\Delta}_{j_2}^{(k+1)},\cdots,\mathbf{\Delta}_{j_J}^{(k+1)},\mathbf{\Delta}_{i_1}^{(k+1)},\mathbf{\Delta}_{i_2}^{(k+1)},\cdots,\mathbf{\Delta}_{i_I}^{(k+1)}\right]^T,\quad k=0,1,2,\cdots, the local iteration format in matrix form is, \mathbf{D}^{(k+1)}=\mathbf{T}\mathbf{D}^{(k)},\quad k=0,1,2,\cdots, where \mathbf{T} is the iteration matrix: \mathbf{T}= \begin{bmatrix} \mathbf{E}_J & -\mathbf{B}_1\\ 0 & \mathbf{E}_I-\mathbf{B}_2 \end{bmatrix}, where \mathbf{E}_J and \mathbf{E}_I are the identity matrices and \mathbf{B}_1= \begin{bmatrix} B_{i_1}\left(t_{j_1} \right) & B_{i_2}\left(t_{j_1} \right) & \cdots &B_{i_I}\left(t_{j_1} \right) \\ B_{i_1}\left(t_{j_2} \right) & B_{i_2}\left(t_{j_2} \right) & \cdots &B_{i_I}\left(t_{j_2} \right) \\ \vdots & \vdots &\vdots & \vdots \\ B_{i_1}\left(t_{j_J} \right) & B_{i_2}\left(t_{j_J} \right) & \cdots &B_{i_I}\left(t_{j_J} \right) \\ \end{bmatrix}, \mathbf{B}_2= \begin{bmatrix} B_{i_1}\left(t_{i_1} \right) & B_{i_2}\left(t_{i_1} \right) & \cdots &B_{i_I}\left(t_{i_1} \right) \\ B_{i_1}\left(t_{i_2} \right) & B_{i_2}\left(t_{i_2} \right) & \cdots &B_{i_I}\left(t_{i_2} \right) \\ \vdots & \vdots &\vdots & \vdots \\ B_{i_1}\left(t_{i_I} \right) & B_{i_2}\left(t_{i_I} \right) & \cdots &B_{i_I}\left(t_{i_I} \right) \\ \end{bmatrix}. The above local iteration format converges and can be extended to blending surfaces and subdivision surfaces. The PIA format for implicit curve and surface reconstruction is presented in the following. Given an ordered point cloud \left\{\mathbf{Q}_i\right\}_{i=1}^n and a unit normal vector \left\{\mathbf{n}_i\right\}_{i=1}^n on the data points, we want to reconstruct an implicit curve from the given point cloud. To avoid a trivial solution, some offset points \left\{\mathbf{Q}_l\right\}_{l=n+1}^{2n} are added to the point cloud. They are offset by a distance \sigma along the unit normal vector of each point \mathbf{Q}_l=\mathbf{Q}_i+\sigma\mathbf{n}_i,\quad l=n+i,\quad i=1,2,\cdots,n. Assume that \epsilon is the value of the implicit function at the offset point f\left(\mathbf{Q}_l\right)=\epsilon,\quad l=n+1,n+2,\cdots,2n. Let the implicit curve after the \alpha th iteration be f^{(\alpha)}(x,y)=\sum_{i=1}^{N_u}\sum_{j=1}^{N_v}C_{ij}^{(\alpha)}B_i(x)B_j(y), where C_{ij}^{(\alpha)} is the control point. Define the difference vector of data points as \begin{aligned} \boldsymbol{\delta}_k^{(\alpha)}&=0-f^{(\alpha)}(x_k,y_k),\quad k=1,2,\cdots,n,\\ \boldsymbol{\delta}_l^{(\alpha)}&=\epsilon-f^{(\alpha)}(x_l,y_l),\quad l=n+1,n+2,\cdots, 2n. \end{aligned} Next, calculate the difference vector of control coefficients \boldsymbol{\Delta}_{ij}^{(\alpha)}=\mu\sum_{k=1}^{2n} B_i(x_k)B_j(y_k) \boldsymbol{\delta}_k^{(\alpha)},\quad i=1,2,\cdots,N_u,\quad j=1,2,\cdots,N_v, where \mu is the convergence coefficient. As a result, the new control coefficients are C_{ij}^{(\alpha+1)}=C_{ij}^{(\alpha)}+\boldsymbol{\Delta}_{ij}^{(\alpha)}, leading to the new algebraic B-spline curve f^{(\alpha+1)}(x,y)=\sum_{i=1}^{N_u}\sum_{j=1}^{N_v}C_{ij}^{(\alpha+1)}B_i(x)B_j(y). The above procedure is carried out iteratively to generate a sequence of algebraic B-spline functions \left\{f^{(\alpha)}(x,y), \quad \alpha=0,1,2,\cdots\right\} . The sequence converges to a minimization problem with constraints when the initial control coefficients C_{ij}^{(0)}=0 . Assume that the implicit surface generated after the \alpha th iteration is f^{(\alpha)}(x,y,z)=\sum_{i=1}^{N_u}\sum_{j=1}^{N_v}\sum_{k=1}^{N_w}C_{ijk}^{(\alpha)}B_i(x)B_j(y)B_k(z), the iteration format is similar to that of the curve case. \mathcal{F}_{r,j}(f) = \int_{t_1}^{t_m}B_{r,j}(t)fdt,\quad j=1,2,\cdots,n,\quad r=1,2,3, where B_{r,j}(t) represents the r th derivative of the basis function B_j(t) , (e.g. B-spline basis function). Let the curve after the k th iteration be \mathbf{P}^{[k]}(t)=\sum_{j=1}^nB_j(t)\mathbf{P}_j^{[k]},\quad t\in[t_1,t_m]. To construct the new curve \mathbf{P}^{[k+1]}(t) , we first calculate the (k + 1) st difference vectors for data points, \mathbf{d}_i^{[k]} = \mathbf{Q}_i - \mathbf{P}^{[k]}(t_i),\quad i=1,2,\cdots,m. Then, the fitting difference vectors and the fairing vectors for control points are calculated by \begin{align} \boldsymbol{\delta}_j^{[k]} &= \sum_{h\in I_j}B_j(t_h)\mathbf{d}_h^{[k]},\quad j=1,2,\cdots,n \\ \boldsymbol{\eta}_{j}^{[k]} &= \sum_{l=1}^n \mathcal{F}_{r,l}\left(B_{r,j}(t)\right)\mathbf{P}_l^{[k]},\quad j=1,2,\cdots,n \\ \end{align} Finally, the control points of the (k+1) st curve are produced by \mathbf{P}_j^{[k+1]} = \mathbf{P}_j^{[k]} + \mu_j \left[ \left(1-\omega_j\right)\boldsymbol{\delta}_j^{[k]} - \omega_j\boldsymbol{\eta}_{j}^{[k]} \right],\quad j=1,2,\cdots,n, where \mu_j is a normalization weight, and \omega_j is a smoothing weight corresponding to the j th control point. The smoothing weights can be employed to adjust the smoothness individually, thus bringing great flexibility for smoothness. The larger the smoothing weight is, the smoother the generated curve is. The new curve is obtained as follows \mathbf{P}^{[k+1]}(t)=\sum_{j=1}^nB_j(t)\mathbf{P}_j^{[k+1]},\quad t\in[t_1,t_m]. In this way, we obtain a sequence of curves \left\{\mathbf{P}^{[k]}(t),\;k=1,2,3,\cdots\right\} . The sequence converges to the solution of the conventional fairing method based on energy minimization when all smoothing weights are equal ( \omega_j=\omega ). Similarly, the fairing-PIA can be extended to the surface case. \left\{ \begin{aligned} \mathcal{L}u=f,&\quad \text{in}\;\Omega,\\ \mathcal{G}u=g,&\quad \text{on}\;\partial\Omega, \end{aligned} \right. where u:\Omega\to\mathbb{R} is the unknown solution, \mathcal{L} is the differential operator, \mathcal{G} is the boundary operator, and f and g are the continuous functions. In the isogeometric analysis method, NURBS basis functions are used as shape functions to solve the numerical solution of this boundary value problem. The same basis functions are applied to represent the numerical solution u_h and the geometric mapping G : \begin{aligned} u_h\left(\hat{\tau}\right) &= \sum_{j=1}^nR_{j}(\hat\tau )u_j,\\ G({\hat \tau }) &= \sum_{j=1}^nR_{j}(\hat\tau )P_j, \end{aligned} where R_j(\hat{\tau}) denotes the NURBS basis function, u_j is the control coefficient. After substituting the collocation points \hat\tau_{i} ,i = 1,2,...,{m} into the strong form of PDE, we obtain a discretized problem \left\{ \begin{aligned} \mathcal{L}u_{h}(\hat\tau_{i})=f(G(\hat\tau_{i})),&\quad i\in\mathcal{I_L},\\ \mathcal{G}u_{h}(\hat\tau_{j})=g(G(\hat\tau_{j})),&\quad j\in\mathcal{I_G}, \end{aligned} \right. where \mathcal{I_L} and \mathcal{I_G} denote the subscripts of internal and boundary collocation points, respectively. Arranging the control coefficients u_j of the numerical solution u_h(\hat\tau) into an 1 -dimensional column vector \mathbf{U}=[u_1,u_2,...,u_n]^T , the discretized problem can be reformulated in matrix form as \mathbf{AU}=\mathbf{b} where \mathbf{A} is the collocation matrix and \mathbf{b} is the load vector. Assume that the discretized load values are data points \left\{b_i\right\}_{i=1}^m to be fitted. Given the initial guess of the control coefficients \left\{u_j^{(0)}\right\}_{j=1}^n, n , we obtain an initial blending function U^{(0)}(\hat\tau) = \sum_{j=1}^nA_j(\hat\tau)u_j^{(0)},\quad\hat\tau\in[\hat\tau_1,\hat\tau_m], where A_j(\hat\tau) , j=1,2,\cdots,n , represents the combination of different order derivatives of the NURBS basis functions determined using the operators \mathcal{L} and \mathcal{G} A_j(\hat\tau) = \left\{ \begin{aligned} \mathcal{L}R_j(\hat\tau), &\quad \hat{\tau}\ \text{in}\ \Omega_p^{in},\\ \mathcal{G}R_j(\hat\tau), &\quad \hat{\tau}\ \text{in}\ \Omega_p^{bd}, \quad j=1,2,\cdots,n, \end{aligned} \right. where \Omega_p^{in} and \Omega_p^{bd} indicate the interior and boundary of the parameter domain, respectively. Each A_j(\hat\tau) corresponds to the j th control coefficient. Assume that J_{in} and J_{bd} are the index sets of the internal and boundary control coefficients, respectively. Without loss of generality, we further assume that the boundary control coefficients have been obtained using strong or weak imposition and are fixed, i.e., u_{j}^{(k)}=u_{j}^{*},\quad j\in J_{bd},\quad k=0,1,2,\cdots. The k th blending function, generated after the k th iteration of IG-LSPIA, is assumed to be as follows: U^{(k)}(\hat\tau) = \sum_{j=1}^nA_j(\hat\tau)u_j^{(k)},\quad\hat\tau\in[\hat\tau_1,\hat\tau_m]. Then, the difference vectors for collocation points (DCP) in the (k + 1) st iteration are obtained using \begin{align} \boldsymbol{\delta}_i^{(k)} &= b_i-\sum_{j=1}^{n}A_j(\hat\tau_i)u_j^{(k)}\\ &= b_i-\sum_{j\in J_{bd}}A_j(\hat\tau_i)u_j^{(k)} -\sum_{j\in J_{in}}A_j(\hat\tau_i)u_j^{(k)} ,\quad i=1,2,...,m. \end{align} Moreover, group all load values whose parameters fall in the local support of the j th derivatives function, i.e., A_j(\hat\tau_i)\ne 0 , into the j th group corresponding to the j th control coefficient, and denote the index set of the j th group of load values as I_j . Lastly, the differences for control coefficients (DCC) can be constructed as follows: d_j^{(k)}=\mu\sum_{h\in I_j}A_j(\hat\tau_h)\boldsymbol{\delta}_h^{(k)},\quad j=1,2,...,n, where \mu is a normalization weight to guarantee the convergence of the algorithm. Thus, the new control coefficients are updated via the following formula, u_j^{(k+1)}=u_j^{(k)}+d_j^{(k)},\quad j=1,2,...,n, Consequently, the (k + 1) st blending function is generated as follows: U^{(k+1)}(\hat\tau) = \sum_{j=1}^nA_j(\hat\tau)u_j^{(k+1)}. The above iteration process is performed until the desired fitting precision is reached and a sequence of blending functions is obtained \left \{ U^{(k)}(\hat\tau),k=0,1,\dots \right \}. The IG-LSPIA converges to the solution of a constrained least-squares collocation problem. If n=m , the PIA iterative format in matrix form is \begin{align} \mathbf{P^{(\alpha+1)}} &=\mathbf{P^{(\alpha)}}+\mathbf{\Delta}^{(\alpha)} \\ &=\mathbf{P}^{(\alpha)}+\mathbf{Q}-\mathbf{B}\mathbf{P}^{(\alpha)} \\ &=\left(\mathbf{I}-\mathbf{B}\right)\mathbf{P}^{(\alpha)}+\mathbf{Q} \end{align} where \begin{align} \mathbf{Q} &= \left[\mathbf{Q}_1,\mathbf{Q}_2,\cdots,\mathbf{Q}_m\right]^T \\ \mathbf{P^{(\alpha)}} &= \left[\mathbf{P}_1^{(\alpha)},\mathbf{P}_2^{(\alpha)},\cdots,\mathbf{P}_n^{(\alpha)}\right]^T \\ \mathbf{\Delta}^{(\alpha)} &= \left[\mathbf{\Delta}_1^{(\alpha)},\mathbf{\Delta}^{(\alpha)}_2,\cdots,\mathbf{\Delta}^{(\alpha)}_n\right]^T \\ \mathbf{B} &= \begin{bmatrix} B_1(t_1) & B_2(t_1) &\cdots &B_n(t_1)\\ B_1(t_2) & B_2(t_2) &\cdots &B_n(t_2)\\ \vdots & \vdots &\ddots & \vdots \\ B_1(t_m) & B_2(t_m) &\cdots &B_n(t_m)\\ \end{bmatrix}. \end{align} The convergence of the PIA is related to the properties of the collocation matrix. If the spectral radius of the iteration matrix \mathbf{I}-\mathbf{B} is less than 1 , then the PIA is convergent. It has been shown that the PIA methods are convergent for Bézier curves and surfaces, B-spline curves and surfaces, NURBS curves and surfaces, triangular Bernstein–Bézier surfaces, and subdivision surfaces (Loop, Catmull-Clark, Doo-Sabin). If n , the LSPIA in matrix form is \begin{align} \mathbf{P^{(\alpha+1)}}&=\mathbf{P^{(\alpha)}}+\mu\mathbf{B}^T\mathbf{\Delta}^{(\alpha)} \\ &=\mathbf{P}^{(\alpha)}+\mu\mathbf{B}^T\left(\mathbf{Q}-\mathbf{B}\mathbf{P}^{(\alpha)}\right) \\ &=\left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)\mathbf{P}^{(\alpha)}+\mu\mathbf{B}^T\mathbf{Q}. \end{align} When the matrix \mathbf{B}^T\mathbf{B} is nonsingular, the following results can be obtained: Proof Since \mathbf{B}^T\mathbf{B} is nonsingular, and \mu>0 , then \lambda(\mu\mathbf{B}^T\mathbf{B})>0 . Moreover, \lambda(\mu\mathbf{B}^T\mathbf{B}) =\mu\lambda(\mathbf{B}^T\mathbf{B}) In summary, 0 . Proof From the matrix form of iterative format, we obtain the following: \begin{align} \mathbf{P^{(\alpha+1)}}&=\left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)\mathbf{P}^{(\alpha)}+\mu\mathbf{B}^T\mathbf{Q},\\ &=\left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)\left[\left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)\mathbf{P}^{(\alpha-1)}+\mu\mathbf{B}^T\mathbf{Q}\right]+\mu\mathbf{B}^T\mathbf{Q},\\ &=\left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)^2\mathbf{P}^{(\alpha-1)}+\sum_{i=0}^1\left( \mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)\mu\mathbf{B}^T\mathbf{Q},\\ &=\cdots\\ &=\left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)^{\alpha+1}\mathbf{P}^{(0)}+\sum_{i=0}^{\alpha}\left( \mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)^{\alpha}\mu\mathbf{B}^T\mathbf{Q}.\\ \end{align} According to above Lemma, the spectral radius of the matrix \mu\mathbf{B}^T\mathbf{B} satisfies 0 and thus the spectral radius of the iteration matrix satisfies 0 When \alpha\rightarrow \infty \left(\mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)^{\infty}=0,\ \sum_{i=0}^{\infty}\left( \mathbf{I}-\mu\mathbf{B}^T\mathbf{B}\right)^{\alpha}=\frac{1}{\mu}\left(\mathbf{B}^T\mathbf{B}\right)^{-1}. As a result, \mathbf{P}^{(\infty)}=\left(\mathbf{B}^T\mathbf{B}\right)^{-1}\mathbf{B}^T\mathbf{Q}, i.e., \mathbf{B}^T\mathbf{B}\mathbf{P}^{(\infty)}=\mathbf{B}^T\mathbf{Q} , which is equivalent to the normal equation of the fitting problem. Hence, the LSPIA algorithm converges to the least squares result for a given sequence of points.
Text derived from https://en.wikipedia.org/w/index.php?oldid=1316305253 under CC BY-SA 4.0; modifications have been made.