---
title: Kansa-like Method for Meshfree Collocation
url: https://www.emergentmind.com/topics/kansa-like-method
type: topic
---

# Kansa-like Method for Meshfree Collocation

Searching arXiv for the cited Kansa-like method papers to ground the article in current arXiv records.
Kansa-like method denotes a family of meshfree, kernel-based strong-form collocation schemes in which an approximate solution is expanded in translates of a kernel or radial basis function and the governing differential operator is enforced at collocation points. In the classical square formulation, the trial and test sets coincide; in later asymmetric, over-tested, or constrained variants, the test set is enlarged, the collocation matrix becomes rectangular, and the discrete problem is solved by least squares or by a constrained optimization procedure. In the cited literature, this framework is developed for second-order strongly elliptic PDEs on spheres and manifolds, parabolic surface diffusion, Hamiltonian wave equations, spatiotemporal fractional diffusion, and systems of ODEs in biology [2507.15137][2109.03409][2507.04361][1610.04698][1705.09381].

## 1. Classical strong-form collocation

In its basic form, the method seeks an approximation of the form
\[
u_N(x)=\sum_{j=1}^N \xi_j\,\phi(\|x-x_j\|;c),
\]
where $\phi(r;c)$ is an RBF depending on a scalar shape parameter $c>0$, $\{x_j\}_{j=1}^N$ are distinct centers, and the coefficients $\{\xi_j\}$ are unknown. For a differential operator problem
\[
L[u](x)=f(x),\qquad B[u](x)=g(x),
\]
the strong form is collocated by enforcing
\[
L[u_N](x_i)=f(x_i),\qquad B[u_N](x_i)=g(x_i),
\]
at interior and boundary or initial points. This yields the algebraic system
\[
A(c)\,\Xi = F,
\]
with entries obtained by applying $L$ or $B$ directly to the basis functions at the collocation points [1705.09381].

On compact manifolds, including the sphere, the same idea is expressed with a zonal kernel $\Phi:M\times M\to\mathbb R$, $\Phi(x,y)=\phi(x\cdot y)$, and a trial space
\[
S_X(\Phi)=\operatorname{span}\{\Phi(\cdot,x_j):j=1,\dots,N\}.
\]
Writing
\[
v(z)=\sum_{j=1}^N a_j\,\Phi(z,x_j),
\]
classical Kansa collocation imposes
\[
L\,v(x_i)=f(x_i),\qquad i=1,\dots,N,
\]
so that the square matrix has entries
\[
K_{i,j}=L^{(1)}\Phi(y_i,x_j)=L[\Phi(\cdot,x_j)](y_i),
\]
with $y_i=x_i$ in the classical case. The theoretical treatment on spheres emphasizes that this square Kansa matrix need not be invertible, nor well-conditioned, unless the kernel $\Phi$ and operator $L$ satisfy special compatibility [2507.15137].

A persistent issue in the classical form is the interaction between approximation quality and conditioning. In Gaussian-RBF implementations, small $c$ produces very flat, globally supported basis functions and severe ill-conditioning, whereas large $c$ gives strongly peaked bases and poorer global approximation [1705.09381]. This trade-off is one of the main motivations for later Kansa-like extensions.

## 2. Asymmetric, over-tested, and least-squares formulations

A central development is the replacement of the square collocation system by an overdetermined one. In the asymmetric formulation for elliptic PDEs on manifolds, the trial set $X=\{x_1,\dots,x_N\}$ is retained, but the test set is enlarged to $Y=\{y_1,\dots,y_M\}$ with $M\ge N$, producing
\[
K_{Y,X}\,a=f|_Y.
\]
Here
\[
(K_{Y,X})_{k,j}=L[B_j](y_k),
\]
and $\{B_j\}_{j=1}^N$ may be kernel translates, a Lagrange basis, or a local Lagrange basis. The analysis introduces the notion of an $L_p$-norming set $Y$ for a finite-dimensional space $W_N$, meaning that
\[
\|w\|_{L_p(M)} \le C_N\,M^{-1/p}\,\|w|_Y\|_{\ell_p(Y)}\qquad\text{for all }w\in W_N.
\]
When the trial-related space satisfies a Bernstein inequality, a quasi-uniform minimal $\varepsilon$-net $Y$ of cardinality $M\sim N$ can be chosen so that this norming property holds with $C_N\sim1$ [2507.15137].

The same oversampling principle appears in the surface-diffusion setting. With trial centers $Z=\{z_j\}_{j=1}^N\subset\Gamma$ and collocation nodes $Y=\{y_i\}_{i=1}^M\subset\Gamma$, the semi-discrete method of lines produces
\[
A\,\dot c(t)+B\,c(t)=F(t),
\]
with $M=|Y|>N=|Z|$. Because the system is overdetermined, it is not solved exactly; instead one minimizes the spatial $\ell^2$ residual at fixed $t$, or the space-time functional
\[
J(c)=\int_0^T \|A\dot c(t)+B c(t)-F(t)\|_2^2\,dt,
\]
subject to an initial condition obtained by a discrete Tikhonov-regularized fit of $g$ [2109.03409].

The cited analyses make stability the principal justification for oversampling. In the elliptic manifold theory, if $Y$ is an $L_2$-norming set and the trial basis is a Riesz basis with stability ratio $r_2(X)$, then
\[
\|K_{Y,X}a\|_{\ell_2(Y)} \ge \frac{1}{r_2(X)}\,\frac{1}{C_N}\,\sqrt{\frac{M}{2}}\,\|a\|_{\ell_2(X)}.
\]
Hence $K_{Y,X}$ is injective and $G=K^*K$ is invertible on $\ell_2(X)$ [2507.15137]. In the surface-diffusion theory, discrete sampling inequalities and a time-dependent stability estimate control the $H^1$, $H^2$, and $\partial_t$ components of the error, and the paper explicitly contrasts this with the square Kansa method, which is described as ill-conditioned and subject to solvability and stability difficulties [2109.03409].

## 3. Kernel spaces, trial spaces, and geometric operators

Kansa-like methods are defined by a trial space together with a strong-form discretization of the operator. For surface diffusion on a smooth, closed, orientable Riemannian manifold $\Gamma\subset\mathbb R^d$, the trial kernel is chosen from a positive-definite ambient-space kernel $\Phi_{m+\frac12}(x,y)$ whose Fourier transform decays like $(1+\|\omega\|^2)^{-(m+\frac12)}$. Restricting it to $\Gamma\times\Gamma$ gives
\[
\Psi_m(x,y)=\Phi_{m+\frac12}(x,y)|_{\Gamma\times\Gamma},
\]
which reproduces $H^m(\Gamma)$. The trial space is
\[
U_Z=\operatorname{span}\{\Psi_m(\cdot,z_j):z_j\in Z\},
\]
and the approximate solution is
\[
u_h(x,t)=\sum_{j=1}^N c_j(t)\,\Psi_m(x,z_j).
\]
The Laplace–Beltrami operator is implemented extrinsically through the closest-point extension and the projection $P=I-nn^T$:
\[
\Delta_\Gamma v(y)=\operatorname{tr}\!\bigl(P\,D^2(v\circ cp)(y)\,P\bigr)
\]
[2109.03409].

A closely related manifold construction appears in the Trajectory-Based RBF Collocation method for surface advection-diffusion. There the trial centers $Z\subset\mathcal M$ and collocation points $P\subset\mathcal M$ lie on a smooth, compact manifold $\mathcal M\subset\mathbb R^3$, and the approximation takes the form
\[
u_Z(\mathbf x,\tau)=\sum_{j=1}^{N_Z}\lambda_j(\tau)\,\Phi(\|\mathbf x-\mathbf z_j\|).
\]
The numerical tests use the Whittle–Matérn–Sobolev kernel
\[
\Phi_m(r)=\frac{2^{1-m+\tfrac d2}}{\Gamma(m-\tfrac d2)}\,r^{\,m-\tfrac d2}\,K_{m-\tfrac d2}(r),\qquad d=3,
\]
and the description explicitly states that no additional shape parameter is introduced beyond the native length-scale $r=\|\mathbf x-\mathbf z\|$. The surface Laplacian is again treated extrinsically,
\[
\Delta_{\mathcal M}^{(x)}\Phi(\|x-z\|)= (\mathcal P(x)\nabla_x)\!\cdot\!(\mathcal P(x)\nabla_x)\,\Phi(\|x-z\|),
\]
now coupled to a characteristic back-tracking map in time [2601.18186].

On spheres and other closed compact manifolds, the elliptic theory uses either positive-definite or strictly conditionally positive-definite zonal kernels. In the conditionally positive-definite case, the trial space is augmented by spherical harmonics and side conditions of the form $\sum c_j p(x_j)=0$ are imposed for all harmonics of degree at most $L-1$ [2507.15137]. This indicates that the term “Kansa-like method” covers a broader algebraic structure than a single choice of kernel or basis.

## 4. Time discretization and structure-preserving variants

For parabolic surface problems, the Kansa-like method is naturally cast as a method of lines. Substituting the trial expansion into
\[
\partial_t u(y,t)-\nabla_\Gamma\!\cdot\!(A(y,t)\nabla_\Gamma u(y,t))=f(y,t)
\]
and collocating in space gives the overdetermined ODE system
\[
A\,\dot c(t)+B\,c(t)=F(t),
\]
where $A_{i,j}=\Psi_m(y_i,z_j)$ and $B_{i,j}=-[\Delta_\Gamma \Psi_m(\cdot,z_j)](y_i)$. Backward Euler yields
\[
A\,\frac{c^n-c^{n-1}}{\Delta t}+B\,c^n=F^n,
\]
or equivalently
\[
(A/\Delta t+B)c^n = F^n+(A/\Delta t)c^{n-1},
\]
which is solved in the least-squares sense through the normal equations
\[
(A/\Delta t+B)^T(A/\Delta t+B)c=(A/\Delta t+B)^T rhs.
\]
The paper also notes that higher-order backward-difference formulas and quadrature in time lead to decoupled least-squares solves at each time slice, and that an equivalent ODE form is
\[
\dot c=A^\dagger(F(t)-Bc)
\]
[2109.03409].

In Hamiltonian wave equations, the Kansa framework is modified so that the discrete evolution inherits energy conservation. After time discretization of
\[
\ddot u-\Delta u+F'(u)=0,
\]
the step at time level $k$ is written as a nonlinear least-squares problem
\[
\alpha^k = \arg\min_{\eta\in\mathbb R^N}\|A\eta-b_k(\eta)\|_2^2,
\]
subject to the energy constraint
\[
E_w(\alpha^k)=E_0,
\]
or, in quadratic-plus-nonlinear form,
\[
\frac12(\alpha^k)^TQ\alpha^k + N(\alpha^k)-E_0=0.
\]
The constrained problem is handled through a Lagrangian and a fast iterative solver based on a generalized SVD of the pair $(A,B)$ together with a one-dimensional Newton iteration for the Lagrange multiplier $\lambda$. The cited complexity estimates are one GSVD at cost $O(N_Z^3+(N_X+N_Y)N_Z^2)$ and $O(N_Z)$ per Newton iteration, with a typical requirement of $10$–$50$ iterations per step. The discrete energy is maintained to within the solver tolerance, reported as $10^{-8}$–$10^{-12}$ in experiments, and comparisons with a secant-based Lagrange-multiplier solver show a $5\times$–$10\times$ speed-up [2507.04361].

The characteristic-splitting surface advection-diffusion method provides a different time-dependent variant. After back-tracking the advection term along characteristics, the diffusion stage is advanced by Crank–Nicolson or BDF. For Crank–Nicolson the coefficients satisfy
\[
A\,\boldsymbol\lambda^{\,n+1}=B\,\boldsymbol\lambda^n,
\]
where the matrix $B$ evaluates the kernel and its surface Laplacian at the departure points $\mathcal X(t_n;\mathbf p_i,t_{n+1})$. The theoretical contribution is an exact equivalence between the operator-split characteristic system and the original surface PDE, so that no operator-splitting error arises; the resulting time discretization is described as unconditionally stable under Crank–Nicolson [2601.18186].

## 5. Fractional, biological, and surface applications

The method has been adapted to spatiotemporal fractional diffusion by combining Kansa collocation in space with an analytical treatment in time. For the two-dimensional FADE
\[
{}^C\!D_t^\alpha u - \nabla\!\cdot\!(\mathbf V u)+k\,\mathcal D_M^\beta u = f,
\]
the approximation uses Hardy’s multiquadric kernel
\[
\phi(r)=\sqrt{r^2+c^2},
\]
and a global expansion
\[
u_N(x,y,t)=\sum_{j=1}^N \lambda_j(t)\,\phi(\|(x,y)-(x_j,y_j)\|).
\]
Collocating the spatial operators gives
\[
A\,{}^C\!D_t^\alpha\boldsymbol\lambda(t)=B\,\boldsymbol\lambda(t)+\mathbf F(t).
\]
When $\mathbf F(t)\equiv0$, setting $M=A^{-1}B$ yields the analytical solution
\[
\boldsymbol\lambda(t)=E_{\alpha,1}(M t^\alpha)\,\boldsymbol\lambda_0.
\]
The implementation therefore avoids time-stepping; the main cost is formation and inversion of $A$ and the eigendecomposition of $M$, each independent of the final time horizon $T$. Reported one-dimensional maximum absolute errors at $t=10$, $\alpha=0.6$, $\beta=1.6$ are $4.94\times10^{-3}$, $1.41\times10^{-3}$, $8.28\times10^{-4}$, and $2.91\times10^{-4}$ for $\Delta x=1/10,1/20,1/25,1/50$, respectively. In the two-dimensional rectangular test, the MAEs at $t=10$ are $1.00\times10^{-2}$, $4.70\times10^{-3}$, $2.81\times10^{-3}$, and $2.79\times10^{-3}$ for $\Delta=1/10,1/15,1/20,1/25$, and random versus perturbed grids are reported to give comparable MAEs within a factor of two [1610.04698].

In biological ODE models, Kansa collocation has been combined with a genetic strategy for selecting the Gaussian shape parameter
\[
\phi(r;c)=\exp(-c^2 r^2).
\]
For the HIV CD4$^+$ T-cell model, the three ODEs are collocated at $N$ time points in $[0,1]$, the three initial conditions are imposed at $t=0$, and the resulting $3(N+1)$ nonlinear equations are solved by Newton–Raphson. The GA searches $c\in(0.1,5)$ and, for $N=20$, converges to $c\approx0.774$. The reported Gaussian-RBF solution matches an eighth-order Runge–Kutta solution to $O(10^{-8})$ or better for $T$, $I$, and $V$. For the Influenza SIRC model, four ODEs plus four initial conditions produce $4(N+1)$ equations, the GA searches $c\in(1,200)$, and for $N=20,40,60$ the selected values are $c\approx27$, $138$, and $123$, respectively. Relative errors of order $10^{-4}$ are reported even at $N=20$, improving as $N$ increases [1705.09381].

Surface applications highlight the geometric flexibility of over-tested Kansa-like methods. On the sphere, the surface-diffusion paper studies
\[
\partial_tu-0.1\Delta_\Gamma u+3u=f
\]
with manufactured solution
\[
u^*(x,t)=e^{x_1+1/(1+t)}.
\]
Using Wendland-type or Matérn-type $\Psi_m$ with $m=4$, the method shows second-order temporal accuracy in $\Delta t$ and spectral-like spatial decay in $h_Z$ until stagnation. A tensor-diffusion example with
\[
A(y)=P(y)\,\operatorname{diag}(x_1^2+1,1,1)
\]
demonstrates anisotropic diffusion, and an Allen–Cahn computation compares the least-squares RBF method with a meshless Galerkin solver of Kunemund–Narcowich–Ward–Wendland, reporting comparable accuracy and superior robustness for long-time integration. The same work emphasizes that surface geometries beyond spheres, including the torus, require no parameterization, only closest-point extension and normal-vector projection [2109.03409].

## 6. Error estimates, conditioning, and methodological scope

The modern theory of Kansa-like methods is largely a theory of stability and error control under oversampling. For second-order elliptic PDEs on spheres, the discrete least-squares solution based on a thin-plate spline spherical basis satisfies
\[
\|u-u_N\|_{L_2(S^d)} \le C\,\rho_X^{2s+d}\,q_X^{2s+d-2}\,\|u\|_{H^{2s+d}(S^d)},
\]
provided $u$ and $f$ are sufficiently smooth and $Y$ is a suitable $L_2$-norming set with $M\sim N$. The RRQR-based thinning procedure, which reduces an oversampled system to a square one, leads to
\[
\|u-u_N^{\text{thin}}\|_{L_2(S^d)} \le C\,\rho_X^{2s+d}\,q_X^{2s-2}\,\|u\|_{H^{2s+d}(S^d)}.
\]
The same framework shows that the reduced matrix $K_{\mathrm{red}}$ can be chosen so that
\[
\sigma_{\min}(K_{\mathrm{red}})\ge c\,N^{-1}\kappa^{3/2},
\]
with $\kappa=M/N$ [2507.15137].

For parabolic surface diffusion, the analysis is organized into regularity of the exact solution, regularization of the discrete initial condition, stability of the discrete least-squares formulation, and consistency with a kernel interpolant. With $\alpha\asymp h_Y^{m-d_/2}$, the regularized initial state satisfies
\[
\|g_{Z,\alpha}-g\|_{H^1}=O(h_Z^{m-1}+h_Y^{m-1}),
\]
and the final bound is
\[
E[u_{Z,\alpha}-u^*]=O(h_Z^{2m-4}+h_Y^{-2}h_Z^{2m-d_}+h_Y^{d_-2}\alpha^2),
\]
where
\[
E[u]=\sup_t\|u(t)\|_{H^1}^2+\|u\|_{L^2(0,T;H^2)}^2+\|\dot u\|_{L^2(0,T;L^2)}^2.
\]
The stated theorem assumes $m\ge 3+\lfloor d_/2\rfloor$, quasi-uniform $Y$ and $Z$, and $|Y|>|Z|$ [2109.03409].

For manifold advection-diffusion with Whittle–Matérn–Sobolev kernels, the cited estimate is the standard Sobolev-space bound
\[
\|u-u_Z\|_{L^2(\mathcal M)}=\mathcal O(h^{\,m-d/2}),
\]
while the numerical section reports superconvergent rates in $\ell^\infty$ [2601.18186]. For Hamiltonian wave equations, the oversampled Kansa discretization is stated to recover nearly $2m$-order in $L^2$ when the oversampling rate $\gamma=(N_X+N_Y)/N_Z\ge1$ and to preserve discrete energy by construction, up to solver tolerance [2507.04361].

Several recurrent limitations are explicit in the literature. Global collocation matrices are dense, so direct dense solution is typical and can incur $O(N^3)$ cost per solve or per precomputation stage [1610.04698][2601.18186]. Conditioning remains sensitive to the kernel choice, the smoothness order $m$, the node density $h$, and, when present, the shape parameter $c$ [1705.09381][2601.18186]. At the same time, the cited papers consistently present the meshfree character, absence of triangulation, elimination of surface integrals in strong-form discretizations, and direct handling of complex geometries as the main practical scope of Kansa-like methods [2109.03409][2507.04361].

A plausible implication is that “Kansa-like” is best understood not as a single algorithm, but as a family of asymmetric or strong-form kernel collocation strategies whose defining questions are the same across applications: choice of trial space, selection of test points, stabilization by oversampling or constraints, and the management of dense linear algebra.

Source: https://www.emergentmind.com/topics/kansa-like-method