You are on page 1of 14

Noname manuscript No.

(will be inserted by the editor)

Generalized Duffy Transformation for Integrating Vertex Singularities
S. E. Mousavi · N. Sukumar

Received: date / Accepted: date

Abstract For an integrand with a 1/r vertex singularity, the Duffy transformation from a triangle (pyramid) to a square (cube) provides an accurate and efficient technique to evaluate the integral. In this paper, we generalize the Duffy transformation to power singularities of the form p(x)/rα , where p is a trivariate polynomial and α > 0 is the strength of the singularity. We use the map (u, v, w) → (x, y, z) : x = uβ , y = xv, z = xw, and judiciously choose β to accurately estimate the integral. For α = 1, the Duffy transformation (β = 1) is optimal, whereas if α = 1, we show that there are other values of β that prove to be substantially better. Numerical tests in two and three dimensions are presented that reveal the improved accuracy of the new transformation. Higher-order partition of unity finite element solutions for the Laplace equation with a derivative singularity at a re-entrant corner are presented to demonstrate the benefits of using the generalized Duffy transformation. Keywords weakly singular integrand · numerical quadrature · partition of unity enrichment · FEM · BEM.

1 Introduction Elliptic boundary-value problems can admit solutions u ∼ rλ (0 < λ < 1) with derivative singularities— Laplace equation in a domain with a re-entrant corner or when there is an abrupt change in the boundary condition (Dirichlet to Neumann) [1,2]; problems in elasticity such as analysis of plates with sharp notches
National Science Foundation, CMMI-0626481, DMS-0811025. S. E. Mousavi · N. Sukumar Department of Civil and Environmental Engineering, University of California, Davis, CA 95616. E-mail: nsukumar@ucdavis.edu

and cracks [3], wedge-shaped bimaterials [4, 5], crack impinging a bimaterial interface [6], and bimaterial interfacial cracks [7] are well-known examples. In the numerical treatment of such problems with the boundary element and enriched finite element methods, the numerical integration of weakly singular integrands of O 1/rα arises where 0 < α < n in Rn . The development of meshfree [8] and partition of unity finite element methods [9] for modeling singularities—cracks in isotropic media (λ = 1/2) [10–13], strong and weak singularities for a crack perpendicular to a bimaterial interface [14], complete sliding contact [15], HRR cracktip fields with λ = 1/(n + 1) (n is the hardening exponent) [16], hydraulic fracture (λ = 1/2, 2/3) [17, 18], and parametric enrichment for singular problems (λ is itself a parameter, which is obtained through optimization in the solution procedure) [19]—underscores the need to develop accurate numerical integration schemes to compute weak form integrals, which provides the impetus for pursuing this contribution. The use of standard integration techniques such as Gauss-Legendre quadrature rules to evaluate singular integrands has its limitations, both, from the accuracy and cost perspectives. Extrapolation techniques [20–23] construct more accurate integration formulae based on asymptotic error expansions of standard quadratures, whereas in adaptive subdivision scheme [24,25] the integration domain is subdivided into uniform/nonuniform subdomains and well-known integration rules are used over the subdomains. Klees [26] shows that for weakly singular integrands, extrapolation and adaptive subdivision techniques behave poorly in terms of both accuracy and efficiency. Variable transformation methods (also referred as cancelation schemes) to numerically integrate weakly singular integrands are well-established in the litera-

The mapped kernel over the square (cube) is smooth enough and can be integrated within a desired accuracy using a tensor-product Gauss quadrature rule. v. In Section 3. z): x = u. The Duffy transformation [28] from a triangle (pyramid) to a square (cube) is: (u. When the singularity falls inside an element. Park et al. the accuracy and efficiency of this technique is significantly reduced [53].2 y v x (a) 0 ≤ x ≤ 1. which transforms a master triangular element to a square element and as a result the 1/r singularity is removed. As noted by Monegato and Scuderi [34]. If f is non-homogeneous. B´chet et e al. the integrand in the stiffness matrix may consist of terms with singularities α < 1 and α ≥ 1. y. which eliminates singularities of the type 1/r (see Fig. For partition of unity finite element applications with corners or cracks. . v. though attributed to Duffy. application to the Laplace equation with a re-entrant corner is presented. We close with the main findings and a few final remarks in Section 4. it is not as efficient for 1/rα when α = 1. Laborde et al. 2 Formulation A generalization of the Duffy transformation is proposed: (u. but as they point out. it is more convenient to adopt standard Gauss quadrature in all directions. [50] present a series of transformations in two dimensions.1–3. y): x = ρ cos2 θ.and three-dimensional domains. the above transformation in two dimensions was proposed earlier by Fairweather et al. y. Other types of mappings have also been employed for different classes of functions and singularities. For the two-dimensional integral considered by Duffy [28].3. This permits integration with respect to r to be carried out algebraically and the multiple integral in two (three) dimensions is reduced to a line (surface) integral. The aim of this paper is to present a generalization of the Duffy transformation that can provide improved accuracy for integrating vertex singularities within two. [49] triangulate the element with the singularity inside it so that the source point lies at a vertex of a triangle and then the Duffy transformation (though not mentioned in Reference [49]) is applied to integrate singular functions over the triangle. For example. Among the transformation techniques. The Duffy transformation has also been applied within boundary element and finite element methods: applications in Stokes flow [41]. θ) → (x. and in Section 3. where the rate of convergence in strain energy is studied for finite element (FE) and higher-order partition of unity finite element (PUFE) methods. which permits machine-precision accuracy to be realized [32]. The main idea in this approach is to map the physical domain to a parent domain so that the singularity is removed through the introduction of the Jacobian. in most boundary element and finite element applications. the integrand had a u−1/2 term and therefore Gauss-Jacobi quadrature rule was used in the u-direction. Nagarajan and Mukherjee [32] use a polar mapping (ρ. w) → (x. z = xw = uw.4. z): x = uβ . In Section 2. y = xv = uv. wave scattering and Helmholtz equation [42–44]. we introduce the generalized transformation. The Duffy transformation has been used for the integration of stiffness matrix entries with singular finite element shape function derivatives [36–40]. 1 Duffy transformation from (a) the standard triangle to (b) the unit square. This technique proves to be more accurate and has a better convergence rate than standard Gauss quadrature. we compare its performance with the Duffy transformation. the so-called Duffy transformation has found wide appeal. 1). [53] generalize this transformation to threedimensional tetrahedral elements and use it to integrate singular enrichment bases for crack problems within the generalized finite element method. y = xv γ = uβ v γ . The main advantage of this mapping is in the integration of terms such as f (x)/r. Numerical studies on the accuracy and convergence rate of the new transformation are presented in Sections 3. their mapping does not readily extend to three dimensional domains. w) → (x. v ≤ 1 u Fig. 0 ≤ y ≤ x (b) 0 ≤ u. and quantum mechanical density-functional calculations [45–48] to name a few. However. [35] to numerically integrate a 1/r vertex singularity over a triangle. y = ρ sin2 θ. then the element can be divided into triangles in R2 (pyramids in R3 ) with the singularity lying at a vertex of the subdivisions and the transformation can be applied to each subdomain separately. ture [27–33]. and Gauss quadrature in the v-direction. Many recent studies have tackled the issue of numerical integration of singularities within partition of unity finite element methods [49–53]. where f is a homogeneous function. which eventually cancels 1/rα singularities. Even though the Duffy transformation works very well for 1/r singularity.

uβ v γ . We choose α = {1.4). different choices of β are tried. For all values of α and β. the term u3/2 is present which leads to loss in accuracy. 1) to a unit cube [28].g. 1/2. 0). and α = 1/3. Therefore. whereas in the Duffy transformation (β = 1). (1b) From the expression for the Jacobian and the denominator in (1a). This transformation maps the standard pyramid with vertices at (0. 4/3}.1 Vertex singularity in two dimensions To validate the criterion for the selection of β. where p(x. Similar arguments apply in the two-dimensional case (Fig. For some of the values of α. (1. which results in the term u4 . This requirement will marginally increase the polynomial order of f . uβ v. 1) and (1. 1 through a shift of the coordinates and an affine mapping (see Appendix A). Each triangle is mapped to the standard triangle of Fig. Different values of the singularity-exponent α are considered. 0). . uβ w) and u3β−1−αβ have the simplest possible forms that can be easily integrated. For each α. The reported relative errors are the norm of the relative error in the integration of p(x. Finally. we apply the Duffy transformation and the generalized transformation for the integration of functions p(x. which appear in various applications: α = 1/2. 0. The reference values are calculated using very high-order quadrature rules so that sufficient number of digits are converged. (1. the maximum error in the integration as nsp increases corresponds to (1/rα ) dx dy. we pick β so that f remains a polynomial in the transformed space. for a 1/ r (α = 1/2) singularity. y. 1 for cracks in isotropic media [10]. e. where β. we I= = dx 0 1 0 0 1 0 0 1 dy 0 dz f (x. a fractional exponent in the term u3β−1−αβ needs a much higher-order quadrature rule to capture the non-linear behavior of the integrand. This is due to the fact that bivariate polynomials xi y j for i + j > 0 have a radial dependence and therefore (xi y j /rα ) dx dy has a milder radial singularity than (1/rα ) dx dy.3 z = xwζ = uβ wζ . where p includes all bivariate polynomials up to order three (ten functions in two dimensions). uβ w) βu3β−1−αβ dudvdw.e. uβ v. 1. γ and ζ are selected so that the transformed kernel is as smooth as possible and can be integrated with the fewest number of evaluation points. β = 2 is chosen. tensorproduct quadrature rules over the triangulated domain and the unit square are tested (indicated as tensor 1 and tensor 2. 3. 0. respectively). 0). Figures 2d to 2g show similar results for other values of α. However. (1a) [u2β (1 + v 2γ + w2ζ )]α/2 where J is the Jacobian of the transformation: J = βγζu3β−1 v γ−1 wζ−1 . y)/rα . 2/3. uβ wζ ) J dudvdw. The pera formance of the proposed transformation in the numerical integration of a vertex singularity in a square and a cube are presented. for a constant in the numerator of the integrand. the generalized transformation in two dimensions is provided in Appendix A. 4/3 in the solution of Laplace equation in an L-shaped domain with a re-entrant corner (see Section 3. Equation (1) shows the calculation of the integral in three dimensions after the transformation: 1 x x 3 Numerical Examples We present numerical tests to affirm the improvements in accuracy that are realized by the generalized transformation vis-`-vis the Duffy transformation. (1) reduces to 1 1 0 0 1 I= 0 f (uβ . and therefore require a slight increase in the order of the quadrature rule to exactly integrate the polynomial in the u-direction. For example. (1. For instance. 0. On using γ = ζ = 1. In accordance with our expectations. 2c shows the convergence curves for 1/r singularity with β varying from 1 to 5. i. it is evident that the choice γ = 1 and ζ = 1 provides the lowest exponents of the variables v and w in the transformed space. 2/3 in hydraulic fracture [18]. 1. y) ∈ Pd (x. y)/rα .. the reference integral is evaluated using symbolic packages such as Mapleç and MATLABç either exactly or with very high accuracy. the Duffy transformation does not remove the singularity since the exponent of u in the transformed kernel is negative. α = 1/2. y) are bivariate polynomials up to degree d with respect to x and y and r is the distance from the origin. We point out that for singularities α > 2 in three dimensions (α > 1 in two dimensions). we select the minimum β ∈ Z+ so that the exponent 3β − 1 − αβ is a √ positive integer. 1). The integration is carried out over a unit square after dividing it into two triangles (Figures 2a and 2b).. the Laplace problem on an L-shaped domain with a corner singularity is considered and the accuracy and rate of convergence of higher-order FE and PUFE solutions are studied. 2/3. Since in boundary element and finite element applications. z) (x2 + y 2 + z 2 )α/2 f (uβ . Fig. α = 1 and α = 1/2. (1 + v 2 + w2 )α/2 where β is now selected so that both f (uβ . Also. 1/3. polynomial bases are used and the integration of polynomials is required.

(f) α = 2/3.4 10 0 α=1 Relative error 10 −4 10 −8 10 −12 tensor 1 tensor 2 β=1 β=2 β=3 β=4 β=5 10 10 Total number of integration points 1 2 10 −16 (a) (b) (c) 10 0 α = 1/2 10 0 α = 1/3 Relative error Relative error 10 −4 10 −4 10 −8 10 −12 tensor 1 tensor 2 β=1 β=2 β=3 β=4 β=5 10 10 Total number of integration points 1 2 10 −8 10 −12 tensor 1 tensor 2 β=1 β=2 β=3 β=4 β=5 1 2 10 −16 10 −16 10 10 Total number of integration points (d) (e) 10 0 α = 2/3 10 0 α = 4/3 Relative error 10 −8 10 −12 tensor 1 tensor 2 β=1 β=2 β=3 β=4 β=5 Relative error 10 −4 10 −4 10 −8 10 −12 tensor 1 tensor 2 β=1 β=2 β=3 β=4 β=5 1 2 10 −16 10 10 Total number of integration points 1 2 10 −16 10 10 Total number of integration points (f) (g) Fig. and (g) α = 4/3. (d) α = 1/2. 2 Convergence curves for integration of p(x)/r α over a unit square. (e) α = 1/3. Relative errors for (c) α = 1. (a) Unit square and its (b) triangulation. .

1/3. 2d). 2) and (1. It should be noted that increasing β has a dual effect: (1) Improves the accuracy due to the higher exponent of u in the transformed kernel (the transformed domain of integration is the unit square and therefore 0 ≤ u ≤ 1). with p(x. Fig. z)/rα over a unit cube that have a singularity at the origin. 4. then the optimal β delivers markedly better accuracy than the Duffy transformation (β = 1).3 Optimization Careful observation of the behavior of our generalized transformation shows that by increasing the number of integration points in each direction. the kernel K(u.5 note from Figure 2 that the choice β = {1. and z (twenty functions in three dimensions). As an example. the optimal β requires about 1. and each of these pyramids is mapped to the standard one and then the generalized transformation is applied. 2/3. it is evident from the plots that if α = 1. and in partition-ofunity enriched finite element methods with parametric enrichment functions [19]. after which increasing the number of integration points in the u-direction does not have any effect on the accuracy. Hence. y. Moreover. 3 shows that an accuracy of O(10−8 ) is realized on using the generalized Duffy transformation with β = 4 and a 8 × 8 tensor product over the unit square. and (2) More integration points are needed in the unit square due to the increase in the polynomial order of the transformed kernel. 1). 3. numerical tests can guide the choice for β. Such values of α arise in re-entrant corner problems in orthotropic media [54]. different singularities α = {1. Further inspection reveals that the kernel after transformation remains a polynomial with respect to u. 311. Similar to the the two-dimensional case. 3. The results are plotted in Figures 4c to 4g. Based on this finding. for α = 150/311. Consider an integrand of the form p(x. 3 Convergence curves for integration of p(x)/r α over the triangle with vertices (1. Figures 2c to 2g also reveal that Gauss-Legendre quadrature has a very low convergence rate and the errors are O(10−5 ) for 500 integration points. the best choices of β are identical to those obtained in the two-dimensional case. 1) and α = 150/311. 2. 311/150}. 3} gives us the most accurate results for the singularities α = {1. 2. For a relative error of 10−8 . As indicated in Appendix A. The optimal β outperforms all other values of β and is able to reduce the relative error to close to machine precision in all cases. 000 integration points.2 Vertex singularity in three dimensions We examine the three-dimensional case by integrating functions of the form p(x. This idea is made more precise through the following example. we propose to use the minimum number of Gauss points in the u-direction to exactly evaluate the integral with respect to u. . 2. A similar behavior is seen in almost all the convergence curves. 2/3. A MATLAB code that performs the generalized Duffy transformation is provided in Appendix B. the rate of convergence has a sudden decrease at a point and after that it converges with a much lower rate (for example. we consider β = {1. we apply it to the integration of functions with singularities α = p/q when p and q are moderately large positive integers with no common divisor. 5. The rationale provided in Section 2 for selecting β would lead to the choice β = q. 1/3. and once again. see the curve for β = 2 in Fig.5. Similar to the two-dimensional case. the tensor-product Gauss rule can at best deliver an accuracy of 10−7 with 10. respectively. The unit cube is divided into three pyramids with planar bases. one can determine the number of Gauss points in the udirection to obtain exact integration with respect to u. but is an irrational function with respect to v and w. Singularity is at (1. v) after the 10 0 10 Relative error −4 α = 150/311 tensor 1 β=1 β=2 β=3 β=4 β=5 β = 311 β = 311/150 1 2 10 −8 10 −12 10 −16 10 10 Total number of integration points Fig. To further illustrate the performance of the generalized transformation. 4a and one of its partitions (standard pyramid) is depicted in Fig. y. The function p consists of trivariate polynomials up to degree three with respect to x. 3. 3. 3. For such α. 000 integration points. 4/3} are tested with β varying from 1 to 5. 1/2. the constant term in the numerator of the integrand dominates the error in the numerical integration depicted in Fig. y)/rα .3). The unit cube is shown in Fig. 4. (3. 4b. 1/2. 4/3}. y) consisting of polynomials up to order d. and a higher-order quadrature rule in the other directions.

4 Convergence curves for integration of p(x)/r α over a unit cube. (e) α = 1/3. (d) α = 1/2. (f) α = 2/3. and (g) α = 4/3. (b) the standard pyramid. . (a) Unit cube and one of its partitions.6 10 0 α=1 Relative error 10 −4 10 −8 10 −12 tensor β=1 β=2 β=3 β=4 β=5 1 2 3 4 10 −16 10 10 10 Total number of integration points 10 (a) (b) (c) 10 0 α = 1/2 10 0 α = 1/3 10 −4 Relative error 10 −8 10 −12 tensor β=1 β=2 β=3 β=4 β=5 Relative error 10 −4 10 −8 10 −12 tensor β=1 β=2 β=3 β=4 β=5 1 2 3 4 10 −16 10 1 10 10 Total number of integration points 2 3 10 4 10 −16 10 10 10 Total number of integration points 10 (d) (e) 10 0 α = 2/3 10 0 α = 4/3 Relative error 10 −4 Relative error 10 −4 10 −8 10 −12 tensor β=1 β=2 β=3 β=4 β=5 1 2 3 4 10 −8 10 −12 tensor β=1 β=2 β=3 β=4 β=5 10 −16 10 −16 10 10 10 Total number of integration points 10 10 1 10 10 Total number of integration points 2 3 10 4 (f) (g) Fig. Relative errors for (c) α = 1.

When β = 3 (optimized) is used. The PUFE approximation for the trial function u is [9]: uh (x) = i∈I Ni (x)ui + j∈J PU Nj (x)ψ(x)aj ≡ Φk (x)dk . The boundaryvalue problem is posed as: −∇2 u = 0 in Ω. As seen in these figures. 8. 7. On substituting the above trial function and using Φk as test functions in the weak form (4). which corresponds to an enrichment support radius re = ǫ (ǫ is a small number). the PUFE solution is proximal to the exact solution and is able to capture . where H 1 (Ω) is the Sobolev space that consists of functions and their derivatives that are square integrable in Ω. respectively. [12]. a sample mesh of the domain with eight divisions along each coordinate direction is presented. The partition of unity enriched basis is always constructed as the product of the bilinear finite element basis function (NiPU = NiQ4 ) and the enrichment function. we use 4. ∇u = 3 3 3 3 which has a derivative singularity at r = 0. In this paper. 64 and 128 number of divisions along each coordinate direction. In the PUFE plots. uβ v) 2β−1−αβ βu . w = 0 on ΓD }. The DECUHR adaptive algorithm [57] (restricted to hyper-rectangular regions) was adopted by Strouboulis et al. the contributions due to the FE basis and the enriched basis are depicted separately. Kij = Ω d = [u a]T . quadratic (p = 2. the high convergence rate is maintained and in comparison to β = 3 (without optimization). k∈K where Φi = Ni for a classical degree of freedom and Φi = NiPU ψ for an enriched degree of freedom. Q4). In Fig. The support of the enriched basis function is the shaded region in Fig. Following Strouboulis et al. 8. we only enrich the node at the origin (one additional degree of freedom). whereas the generalized Duffy transformation is used in this study. 8). 16. respectively. we choose boundary conditions that are consistent with the exact solution: 2θ r−2/3 2θ θ − sin i + cos j . Q8) and cubic (p = 3. Nj (x) are the FE basis functions used to form the enriched basis function. Q12) serendipity finite elements. [12] to compute the enriched contributions in the stiffness matrix. which has been considered in previous studies [12. 8. and ui and aj are nodal coefficients associated with the finite element and enriched bases. we use a 10×10 generalized Duffy quadrature rule (total of 600 evaluation points over the three elements) in the PUFE computations. it is sufficient to use nspu = (2β−αβ+dβ)/2 Gauss points in the u-direction to exactly evaluate the integral with respect to u. ψ(x) is the enrichment function. and 32 divisions are chosen. We apply Dirichlet boundary condition u = 0 on ΓD and Neumann boundary condition ∇u · n = g on ΓN and all other edges. (4a) (4b) U = {w : w ∈ H 1 (Ω). Fig. For FE computations. (5b) 3. 6. Figures 5a and 5b show the convergence curves for α = 1/2 over the unit square and α = 4/3 over the unit cube. 56]. 9a shows a plot of the solution along the diagonal AB (see Fig. ∂n where Ω is the L-shaped domain shown in Fig. Table 1 shows the number of degrees of freedom for the different meshes and element types. (3) u = r1/3 sin . Thus. PU where Ni (x) are FE basis functions. 55. For FE stiffness matrix calculations and for elements that do not contain an enriched node in PUFE calculations. and for PUFE computations. 8.7 transformation in two dimensions is K(u. fi = ΓN (5a) Φi g dS. meshes with 4. First. we obtain the following discrete system of equations: Kd = f . which are shown in Fig. To obtain higher precision one only needs to increase the number of evaluation points in the other directions. 16. ∇Φi · ∇Φj dV.4 L-shaped domain with a re-entrant corner The L-shaped domain with a corner singularity is a well-known benchmark problem. The weak form of the problem in (2) is: Find u ∈ U such that Ω ∇w · ∇u dV = wg dS ΓN ∀w ∈ U. For stiffness matrix calculation of the three elements containing the singularity (shaded elements in Fig. we use ψ(x) = r1/3 sin θ/3 as the enrichment function in the PUFE method. 4 × 4 tensor-product Gauss quadrature rule is used. 32. (2a) (2b) (2c) u = 0 on ΓD ∂u = g on ΓN . fewer number of integration points are needed to attain higher accuracies. 8) for FE with different pth order elements and Figures 9b and 9c show the results for PUFE. (1 + v 2 )α/2 Collecting like-terms reveals that the highest exponent of u is 2β − 1 − αβ + dβ. we use linear (p = 1. v) = p(uβ . As in Reference [12].

9r4/3 the sharp gradients of the solution in the vicinity of the singular point. The expression for k11 contains terms with 1/rα singularities. u)/2 is the exact strain energy. u ) . 9d for pth order FE and PUFE are in agreement with theory. using Duffy transformation. opt. 2λ) = 2λ = 2/3 [58]. however. u) h h 2 sin(θ/3)N1 ∇N1 · q 3r1/3 E= a(u. This sub-optimal convergence was also noted in previous studies where enrichment for crack problems (λ = 1/2) is used [49. 9d shows the convergence for FE and PUFE with h and p-refinements. u) − a(u . and by the generalized Duffy transformation that is proposed in this paper. On letting q = − sin(2θ/3) i + cos(2θ/3) j and using (3). 9e. 9f. 2d).423569003301483. α = 1/2 (compare to Fig. whereas even on coarse meshes the accuracy of PUFE (one extra degree of freedom in comparison to the corresponding FE problem) is superior and reaches relative errors of 10−4 on a cubic mesh with 32 divisions along each coordinate direction. Fig.5) Q4 Q8 Q12 26 58 90 77 189 301 266 682 1098 989 2589 4189 3815 10087 16359 — — — 10 0 α = 1/2 Relative error 10 −4 10 −8 β=1 β=2 β = 2. see Fig. where ψ = r1/3 sin(θ/3). 8) is plotted for a tensorproduct Gauss rule. the relative error in the computation of k12 (corresponding to the enriched degree of freedom of node 1 and the classical degree of freedom of node 2 . 4g). 4/3: k11 = ∇(N1 ψ) · ∇(N1 ψ) dV ≡ K(x) dV. Due to the singularity (λ = 1/3) at the corner. 8). α = 4/3 (compare to Fig. a(u. convergence in the strain energy is studied. (a) unit square. We return to a potential remedy for this issue later on.50]. The rates for all the curves in Fig. opt. In Fig. The exact strain energy for the problem under consideration is: Eex = 0. 10 −12 10 −16 10 10 Total number of integration points 1 2 (a) 10 Relative error −2 α = 4/3 10 −6 10 −10 β=1 β=3 β = 3. Ω1 10 −14 Ω1 10 1 10 10 Total number of integration points 2 3 (b) Fig. Divisions 4 8 16 32 64 128 Q4 21 65 225 833 3201 12545 FE Q8 53 177 641 2433 9473 37377 Q12 85 289 1057 4033 15745 62209 PUFE (re = ǫ) Q4 Q8 Q12 22 54 86 66 178 290 226 642 1058 834 2434 4034 — — — — — — PUFE (re = 0. the relative error in the computation of k11 (the entry in the global stiffness matrix corresponding to the enriched degree of freedom assigned to the node at the origin. the theoretical asymptotic rate of convergence in strain energy for finite elements is min(2p. The relative error in the strain energy is defined as: a(u. where a(u. where α = 1/3. As a measure of accuracy of the methods. For all FE computations. and (b) unit cube. u) = Ω ∇u · ∇u dV. 5 Effect of optimization. the relative error remains greater than O(10−2 ). In Fig. N1 is the bilinear finite element basis function associated with node 1 and Ω1 is its support (shaded region in Fig. we can write the integrand K(x) as K(x) = r2/3 sin2 (θ/3)||∇N1 ||2 + + 2 N1 .8 Table 1 Number of degrees of freedom for meshes used in FE and PUFE computations.

and 64 divisions) for the linear. v). and cubic PUFE solutions. 6 Laplace problem on an L-shaped domain. whereas the Duffy transformation (β = 1) minimally improves the tensor product rule. 600 function evaluations (10 × 10 over each triangle after the square elements are triangulated) are required. 9d). B 1 2 h = 0. Laborde et al. sin(θ/3)∇N1 · ∇N2 + 3r2/3 θ Γ D x A benefit that accrues with the generalized Duffy transformation is that β = 3 can be used to integrate all the terms in the above integrals. To ensure convergence. with the generalized Duffy transformation the same accuracy is attained with just over 200 evaluation points and more importantly. we obtain rates of convergence in strain energy of 1.85. the quadrature error must be at least an order smaller than the PUFE approximation error. [49] and B´chet et al. the Γ N (−1. 2) are bilinear finite element basis functions. whereas the Duffy quadrature fails (non-monotonicity and incorrect rate of convergence). respectively.25 A Fig. The expression for k12 contains a term with r−2/3 singularity: k12 = Ω2 y (1. 10b shows the convergence curves for the strain energy using linear. 8 Sample mesh (8×8 mesh divisions) for L-shaped domain. a consequence that bears out due to the presence of a rational exponent in the radial term of the kernel K(u. 8). .9 (see Fig. but is not so when the Duffy transformation (see Figures 9e and 9f) or a tensor-product Gauss rule is applied. As can be discerned from Figures 9e and 9f. Fig. To address the issue of sub-optimality in the rate of convergence of PUFE with topological enrichment (see Fig. [50] suge gested the notion of geometric enrichment—all nodes that are inside a fixed region of the singularity are enriched. 3. which indicates that close to the optimal 2p rate of convergence is recovered. Duffy transformation (β = 1) with a 16 × 16 tensorproduct rule is compared to the generalized transformation (β = 3) with a 10 × 10 tensor-product rule. quadratic.5 and all nodes within a distance of re or less from the vertex are enriched (see Fig. We choose an enrichment radius re = 0. and K(x) in this case is given by K(x) = r 1/3 r O N1 ∇N2 · q . The relative error for k11 in Reference [12] is 10−5 for about 400 evaluation points. Ω2 where Ni (i = 1. which is the case when the generalized Duffy transformation is used. The number of degrees of freedom in the PUFE computations are listed in Table 1. 32. quadratic and cubic elements.92 (rates are computed for PUFE solutions using 16. the generalized transformation integrates both k11 and k12 to almost machine precision with about 200 function evaluations per element (k11 is computed over three elements whereas k12 is evaluated on one element).92 and 5. the error can be further reduced without significant addition in the number of evaluation points. 10a). Ω2 is the element that contains nodes 1 and 2 in its connectivity.1) ∇(N1 ψ) · ∇N2 dV ≡ K(x) dV. It is noteworthy to point out that the generalized Duffy transformation outperforms the adaptive algorithm used in Reference [12]. For computing k11 . Hence.−1) Fig. is presented. generalized Duffy quadrature is accurate and yields the correct rate of convergence. With the generalized Duffy quadrature.

2. with β as an additional parameter. y = uβ v transforms the integral from the standard triangle (Fig. The choice of β was guided by the observation that the transformed kernel not have a fractional exponent. 7 Serendipity elements. the generalized Duffy transformation can be combined with the node elimination algorithm presented in Mousavi et al. We also showed that the number of evaluation points can be further reduced by using the minimum number of Gauss points in u-direction (both in two. p = 3).5 was enriched. use of the Duffy transformation or a tensor-product Gauss rule led to inaccuracies since the quadrature error dominated the approximation error. y) + y 2 )α/2 f (uβ . when there is a need for frequent integration of the singular kernel over the same domain. This was possible due to the highly accurate integration that the generalized Duffy transformation afforded. Furthermore. respectively. 3} was the optimal choice. We introduced the following generalized Duffy transformation: (u. The L-shaped domain with a corner singularity was considered. 1/2. and (c) Cubic (Q12. [59] to construct a very efficient quadrature rule with far fewer number of evaluation points. and the numerical results clearly demonstrated the superior accuracy and efficiency of the generalized transformation over the standard Duffy transformation. 3. uβ v γ ) J dudv. 1a) to the unit square (Fig. v. 3. The convergence in strain energy of PUFE was studied: when only the vertex node was enriched. which has been widely used for integration of kernels having 1/r singularity. y) : x = uβ . Appendix A Generalized Duffy Transformation in Two Dimensions The mapping (u. z = uβ w. we presented a generalization to the wellknown Duffy transformation. near-optimal 2p rate of convergence was recovered. 2/3. for α = {1. the merits of the generalized Duffy transformation were revealed in the computation of the enriched stiffness matrix entries. 1/3. z): x = uβ . v) → (x. (a) Linear (Q4. w) → (x. (b) Quadratic (Q8. the stringent demands on accuracy in nonsingular PUFE applications such as acoustics [61] and Schr¨dinger and Poisson solutions in quantum mechano ics [62] reinforces the need and importance of developing accurate and efficient quadrature rules for enriched finite element methods. . + v 2γ )]α/2 0 0 where J is defined as = [u2β (1 J = (βuβ−1 )(γuβ v γ−1 ) = βγu2β−1 v γ−1 . β = {1. y. Even though this study targeted integrands with vertex singularities in PUFE methods. 4/3}. For instance.66 was realized. and higher-order FE and PUFE computations were performed. Once again. 1b): 1 x I= 1 dx 0 1 0 dy (x2 f (x. p = 2).and three-dimensional applications) so that integration is carried out exactly with respect to u and higher-order quadrature rules are used in the other directions. The term 1/rα containing the singularity is used as the weight function in these quadrature rules similar to the quadratures presented by Haegemans [60].10 (a) (b) (c) Fig. p = 1). a convergence rate of 0. The generalized Duffy transformation can be implemented in boundary element and enriched finite element methods for the integration of singular functions without adding to the complexity of the programming and at the same time reducing the number of evaluation points with respect to the Duffy transformation. y = uβ v. 4 Conclusions In this paper. All cases were tested in two and three dimensions. but when a fixed region of radius 0.

5 1 1.2 1 Exact and PUFE solutions 0.5 3 10 −4 Distance along the diagonal AB Number of elements in each direction (c) 0 (d) 10 10 0 Relative error in k11 Tensor product Duffy (β = 1) Generalized Duffy (β = 3) 10 −10 Relative error in k 10 −5 12 10 −4 10 −8 Tensor product Duffy (β = 1) Generalized Duffy (β = 3) 10 −12 10 −15 10 2 10 3 10 4 10 −16 10 2 10 3 Total number of evaluation points Total number of evaluation points (e) (f) Fig.6 0. In the PUFE computations. only the vertex node is enriched (re = ǫ).5 2 2.(c) Solution along the diagonal AB for higher-order PUFE. (b). and (e).4 0.2 0 0 Exact FE (Q12) Enrichment Total (PUFE) 10 Relative error in strain energy 10 −1 10 −2 10 −3 FE−Q4 FE−Q8 FE−Q12 PUFE (Q4) PUFE (Q8) PUFE (Q12) 4 8 16 32 64 128 0.5 2 2.5 2 2.(f) Convergence of k11 and k12 for PUFE (8 × 8 mesh divisions) using different integration schemes.2 1 Exact and FE solutions 0.11 1.2 1 Exact and PUFE solutions 0.8 0.4 0.2 0 0 Exact FE (Q4) Enrichment Total (PUFE) 0.5 1 1.8 0.2 0 0 Exact FE (Q4) FE (Q8) FE (Q12) 1.5 Distance along the diagonal AB 3 0. .4 0.6 0. (d) Convergence of strain energy for FE and PUFE using the generalized Duffy transformation with β = 3.5 3 Distance along the diagonal AB (a) 0 (b) 1. 9 Convergence study for FE and PUFE solutions. (a) Solution along the diagonal AB for FE with different element types.6 0.5 1 1.8 0.

enriched nodes (re = 0. a22]. 2) . The transformation A can be used to map the points of a rule over the standard triangle to R. % get a tensor product over unit square % sq = [0. w0 = gauss_weights(nsp). 1. a22 = sh_coord(3. a12 = sh_coord(3. % shift sh_coord = [coord(:. 0. 0.sh_coord(2. 10 Geometric enrichment for Laplace problem on an L-shaped domain.wg] = getGeneralizedDuffyQuad(coord.5) are shown by open circles.beta.x = x0(i).nsp) % External Dependencies (m-files) % gauss_points(nsp) : 1D Gauss points % gauss_weights(nsp): 1D Gauss weights % Input Parameters % coord : 3 x 2 matrix (triangle coordinates) % beta : Duffy parameter (positive integer) . (a) Sample mesh. coord(:. a21.a. 0. end end Appendix B MATLAB Code for Construction of Generalized Duffy Quadrature in Two Dimensions function [xg. X(ind). there exists an affine transformation A that takes the standard triangle to R [21]. ’y’. Y ) is the physical coordinate system containing the arbitrary triangle and (x. 2) . detA = det(A). A = [a11. 1). 1]. we obtain 1 1 0 % nsp : Number of quadrature points I= 0 f (uβ . for i = 1:nsp for j = 1:nsp ind = (i-1)*nsp+j. 1). 1) . and (b) Convergence in strain energy for PUFE using Duffy transformation (β = 1. % location of vertex singularity a = coord(1. 1) . On setting γ = 1. ’w’. a12. where (X. 1. 15 × 15 quadrature rule) and generalized Duffy transformation (β = 3. a21 = sh_coord(2. X(ind).y = x0(j). 2). X(ind).b]. x0 = gauss_points(nsp). y) is the plane of the standard triangle. Equivalently. 0. (1 + v 2 )α/2 Any arbitrary triangle with a vertex singularity is first translated so that the singularity is moved to the origin and then an affine map to the standard triangle is used: x = aX + bY and y = cX + dY . 2).12 10 Relative error in strain energy −2 10 −4 10 −6 re 10 −8 10 (a) −10 Linear PUFE (β = 1) Quadratic PUFE (β = 1) Cubic PUFE (β = 1) Linear PUFE (β = 3) Quadratic PUFE (β = 3) Cubic PUFE (β = 3) 4 8 16 (b) 32 64 Number of elements in each direction Fig. w0 = w0/2. b = coord(1. X = struct(’x’.w = w0(i)*w0(j). % affine mapping a11 = sh_coord(2.sh_coord(2. uβ v) 2β−1−αβ βu dudv. 2).yg. x0 = (1+x0)/2. X(1:nsp*nsp) = X. 0. 10 × 10 quadrature rule). given an arbitrary triangle R having one vertex at the origin. 0). 1. 1).

Quadrature formulas obtained by variable transformation. Takahasi and M. I. I. Shimizu. Numerical Integration. Recent Developments. 1952. J. 25. 46(1):131– 150. P. 19:526–528. Variable order composite quadrature of singular and nearly singular integrals. C. Mo¨s. 2006. O. V.-H. Boston. C. and K. Department of Computer Science. 53(2):173–194. Stress singularities resulting from various boundary conditions in angular corners of plates in extension. Lecampion. Z. 56(11):3101–3115. The partition of unity method. An extended finite element method for hydraulic fracture problems. Mori. Fuenmayor. R. Elastic crack growth in finite elements with minimal remeshing. P. Melenk. 12. T. 70:2513–2526. 2003. Computer Methods in Applied Mechanics and Engineering. 1994.y = y. Hansbo. Copps. Y. R. International Journal for Numerical Methods in Engineering. for i = 1:length(X) % generalized transformation u = X(i). R. Strouboulis. Y. The design and s analysis of the generalized finite element method. Belgium. Channele cracking of thin films with the extended finite element method. N. s International Journal for Numerical Methods in Engineering. 21. Belytschko. 1985. Belytschko. L. Bogy.x = u^beta. Zak and M. Journal of Applied Mechanics. Lu. Elliptic Problems in Nonsmooth Domains. Huang. Crack point stress singularities at a bi-material interface.x. Computer Methods in Applied Mechanics and Engineering. B. 1963. N. Z1(i). Belytschko and T. Elastic fracture mechanics concepts for interfacial cracks. NATO ASI Series C: Mathematical and Physical Sciences.w = w. Stress singularities in a two material wedge. Singularity enrichment for complete sliding contact using the partition of unity finite element method. Z2(i). 1999. 19. On handling singularities in finite elements. and T.13 9. and Z. J. 20:346–364. de Doncker. Klees. Inc. A. 1). A finite element e method for crack growth without remeshing. Journal of Computational Physics. % reverse affine mapping and shift u = Z1(i). U. 16. An adaptive numerical cubature algorithm for simplices. K. 27. J. yg(i) = Z2(i). Suo. 76(9):1402–1418. 1999. Software and Applications. Waisman and T. Experimental validation of the tip asymptotics for a fluid-driven crack.x. In T. 34:62–79. Fujimoto. Journal of Applied Mechanics. Black. Bunger and E. N. Helpful comments and suggestions of the anonymous reviewers are also greatly appreciated. Two edge-bonded elastic wedges of different materials and wedge angles under surface tractions. 193(33–35):3523–3540. w = detA*w.y. M. A. Lyness. Genz and R. References 1. 73:1671–1692. Detournay. Gravouil. L. Giner. Williams. Grisvard. v = Z1(i). Espelid. 40:727–758. 30:142–143. A. A. 2004.y. editors. Communications in Numerical Methods in Engineering. Technical Report TW 273. Computing. and T. 70(11):781–797. 7. end Acknowledgements The authors acknowledge the research support of the National Science Foundation through contract grants CMMI-0626481 and DMS-0811025 to the University of California at Davis. Fleming. Engineering Fracture Mechanics.w. Numerical calculation of weakly singular surface integrals. Z2(i). Computation of loop integrals using extrapolation. Kluwer Academic Publishers. wg = xg. Li and T.O. International Journal for Numerical Methods in Engineering. Espelid and A. Vercher. Applications of extrapolation techniques to multidimensional quadrature of some integrand functions with a singularity. w = Z1(i). R.x. Numerische Mathematik. 5. 22.y. Hansbo and P. and A. 1971. Computer Methods in Applied Mechanics and Engineering. 1997. Erdogan. 1992. 18. Z2(i). F. for i = 1:nsp*nsp xg(i) = Z2(i). Lyness. 4. w = X(i). Moran. Cools. 31:97–145. Journal of Applied Mechanics. Journal of the Mechanics and Physics of Solids. J. Genz. 11. M. Journal of Geodesy. 195:501–515. Babuˇka and J. Dolbow. Schwab. J.w = w * beta*u^(2*beta-1). 25:121–133. end % write output xg = zeros(nsp^2. J. volume 357. 1971. Leuven. x = a11*u+a12*v+a. 55:98–103. International Journal for Numerical Methods in Engineering. We thank John Pask for pointing us to the work of Batcho [47]. 14. Computer Physics Communications. A.y = u^beta * v. Z. 24. yg = xg. and for his many suggestions that led to significant improvements in this paper. y = a21*u+a22*v+b. wg(i) = Z2(i). 1997. 40:1483–1504. 2008. Yuasa. Pr`vost. Z1(i). 181(1– 3):43–69. v = X(i). E. Enriched element-free Galerkin methods for crack tip fields. 2000. 6. Singularities and treatments of elliptic boundary value problems. T. M. 38:377. and A. B. 21(3):206–219. Belytschko. J. 2008. B. N. D. International Journal of Fracture Mechanics. Parametric enrichment adaptivity by the extended finite element method. 1996. Combescure.w.x = x. 2. Journal of Applied Mechanics. The Netherlands. 2004. Mathematical and Computer Modelling. Z1(i). L. 13. 10. Hein and F. Elguedj. 159:145–156. Pitman Publishing. 17. Chu. 8. Babuˇka. International Journal for Numerical Methods in Engineering. Rice. 15. pages 219–233.w. 2000. 1997. E. 26. 45(5):601–620. On integrating vertex singularities using extrapolation. Sukumar. 2008. T. A. 1976. A finite element method for the simulation of strong and weak discontinuities in solid mechanics. 1994. 3. Appropriate extended functions for X-FEM simulation of plastic fracture mechanics. BIT Numerical Mathematics. . 20. Huang. Dordrecht. H. 1973. T. International Journal for Numerical Methods in Engineering. T. Z2 = X. Williams. 23. H. Y. and F. % transformation Z1 = X. 2009. 1988. 7:317–330.

102(3):1257–1265. 59. L. H. 1994. P. 2002. J. IEEE Transactions on Antennas and Propagation. Bandrauk. P. and D. 38. Schiff. Generalized Gaussian quadrature rules on arbitrary polygons. Yosibash and B. Ling. Cook. Efficient quadrature for a boundary element method to compute three-dimensional Stokes flow. Sukumar and M. Nonequilibrium electron transport in two-dimensional nanostructures modeled using Green’s functions and the finiteelement method. International Journal for Numerical Methods in Engineering. P. Ventura. 37. 53. Quadrature over a pyramid or cube of integrands with a singularity at a vertex. 69(11):115325. E. H. Rizzo. 39. 2009. 2003. 95:331–342. On using enriched cover function in the Partition-of-unity method for singular boundary-value problems. 33. IEEE Transactions on Microwave Theory and Techniques. Batcho. Numerical Algorithms. 1971. A. P. Stepleman. 1982. Efficient electromagnetic modeling of three-dimensional multilayer microstrip antennas and circuits. Genz. 222(4):387–393. Analytical removal of singularities and one-dimensional integration of three-dimensional boundary element method kernels. 1992. Varnhorn. M. 77(1):1–29. On numerical cubatures of singular surface integrals in boundary element methods. N. Lee. Schwab and W. 64(8):1033–1056. S. 20:395–408.. Tracey. D. Journal of Chemical Physics. 2005. 44. 45. P. Computational Mechanics. Bandrauk. F. A conforming crack tip element with quadratic variation in the singular fields. Brussels. M. M.14 28. and S. International Journal for Numerical Methods in Engineering. Belytschko. J. H2 . 31. DOI: 10. Higher order hierarchical Legendre basis functions for electromagnetic modeling. 18:959–966. 12:279–288. H. IEEE Transactions on Antennas and Propagation. D. M. 1987. M. C. Khayat and D. Sukumar and J. K. Analysis of power type singularities using finite elemetns. 2005. G. 52(11):2985–2995. Some special interpolation formulas for triangular and quadrilateral elements. V. S. 62(3):343–369. 35. Journal of Computational and Applied Mathematics. Bettess. Finite elements for determination of crack tip elastic stress intensity factors. Pask. Laborde. Yu and A. N. Q. 62. Laghrouche. J. 60. 1997. J. 52. 29:212–225. U. E. B. 12:19–26. 11:1225–1233. L. 3:255–265. Physical Review B. SIAM Journal on Numerical Analysis. 1977. Liu. J. International Journal for Numerical Methods in Engineering. G. 57. Superelements for the finite element solution of two-dimensional elliptic problems with boundary singularities. N. Computational Mechanics. H. Belgium. A. and T. 64:354–381. C. Finite Elements in Analysis and Design. Wendland. 2005. Computational method for general multicenter electronic structure calculations. International Journal for Numerical Methods in Fluids. Numerical integration of functions with boundary singularities. Xiao. R. F. B. 1982. Classical and enriched finite element formulations for Bloch-periodic boundary conditions.J. Gracie. Mukherjee. H. On quadrature and singular finite-elements. 112(1–2):201–214. 1992. Technical Report TW 192. International Journal for Numerical Methods in Engineering. 56(4):531–552. L. International Journal of Fracture. 47. 30.2759. Breinbjerg. A mapping method for numerical evaluation of two-dimensional integrals with 1/r singularity. Kumosa. M. Yu. Stern and E. 1993. Numerische Mathematik. Cubature formulas for triangles and squares with a 1/r singularity. 43. and R. editors. Duarte. 36. C. Three-dimensional Cartesian finite element method for the time dependent Schr¨dinger o equation of molecules in laser fields. Fix. 58(2):177–192. 77(8):1121–1138. A numerical integration scheme for special finite elements for the Helmholtz equation. K. 54. Volakis. Puska. Computation of double integrals in the boundary integral equation method. 19(6):1260–1262. H. 48. Y. Shirron. 1995. Ying. Renard. International Journal for Numerical Methods in Engineering. International Journal for Numerical Methods in Engineering. R. N. Shippy. Becker. International Journal for Numerical Methods in Engineering. Sugimoto. 1999. International Journal for Numerical Methods in Engineering. T. 2009. Z. Fairweather. 2009. Application of the finite element iterative method to cracks and sharp notches in orthotropic media. Improving the accuracy of XFEM crack tip fields using higher order quadrature and statically admissible stress recovery. M. Engineering Fracture Mechanics. 46. 41. Mo¨s. Burgardt. Integration of singular enrichment functions in the generalized/extended finite element method for three-dimensional problems. 1978. 1993. Pereira. Trevelyan. 1985. Sonnad. 50. 61. 49. Ime e proved implementation and robustness study of the X-FEM for stress analysis around cracks. and J. Solecki and J. A. and M. Liu. Chemical Physics Letters. Computer Methods in Applied Mechanics and Engineering. 61(6):7169–7183. G. 55. F. Fast integration and weight function blending in the extended finite element method. and N. International Journal for Numerical Methods in Engineering. Paulino. M. Espelid and A.1002/nme. Havu. J. 2006. Three-dimensional finite element method for electronic properties of small poly+ 2+ + atomic molecules: H2 . 1973. 66(9):1378–1410. Salaun. J. 40. Swedlow. 26(4):315–335. Karihaloo. 2004. P. P. International Journal for Numerical Methods in Engineering. 4:21–24. 1992. D. Scuderi. K. E. 1979. International Journal for Numerical Methods in Engineering. S. DECUHR: an algorithm for automatic integration of singular functions over a hyperrectangular region. J. 9(2):185– 191. 56. Engineering Analysis. S. Physical Review E. 21:211–228. 34. Sukumar. Harris. Vichnevetsky and R. Numerical evaluation of singular and near-singular potential integrals. Xiao and B. Lean and A. 1994. Wilton. and V. Nieminen. Strang and G. Pommier. 32. Prentice-Hall. 2000. In R. 1984. J. Accurate numerical integration of singular boundary element kernels over boundaries with curvature. Wexler. G. Z. IMACS Publ. O. D. L. Advances in Computer Methods for Partial Differential Equations III. 51. Park. E. International Journal for Numerical Methods in Engineering. C. Meincke. An Analysis of the Finite Element Method. A. and B. Peseux. 58. Englewood Cliffs. 42. Duffy. Mousavi. 2004. J. and J. W. H. Aliabadi and W. M. Highorder extended finite element method for cracked domains. R. 8:201–220. Haegemans. B´chet. A boundary element method for the Helmholtz equation using finite part integration. H3 and H3 . A. X. Monegato and L. Hall. International Journal for Numerical Methods in Engineering. . 1989. Tracey and T. Fan. Jin. Leuven. 78(10):1220–1257. and O. Minnebo. S. Jørgensen. 50(6):1628–1635. Havu. G. 2009. O. Belgium. Department of Computer Science. pages 331–334. 53(10):3180–3190. Nagarajan and S. 29. 2002. and G. L.