Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Techniques related to the CBIEM have been analysed in [20], and used in [9, 8]. The CBIEM is also closely related to the ‘Complex Variable Boundary Element Method’ [15]. This report contains an application of it, using h-p quadrature to achieve high rates of convergence, even in the presence of corner singularities. This application owes its conception to my supervisor. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The ideas of graded meshes and h-p quadrature (numerical integration) are pre- 2, and are illustrated by experimental results. The CBIEM itself is de- sented in 4 details numerical implementation of the CBIEM, using the quadra- 3. scribed in § ture technique described in 2, and presents error results for some test problems. 5 concludes the report with suggestions for further development. matlab code § written for the implementation is listed in Appendix A. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | This section describes a high order numerical integration (quadrature) technique, that retains its high order in the case of end point singularities in the integrand. The method uses a graded mesh, with integration rules of high order used on larger inter- vals, and low order on smaller intervals. To achieve the ‘best’ possible convergence rates, whilst including the end points of each interval, the basic quadrature rules used are Gauß–Lobatto. The underlying mesh is graded in a geometric manner. As the method of using different quadrature rules on internal intervals is a generalisa- tion of earlier ‘h’ and ‘p’ methods, the resulting composite quadrature rule is called a ‘geometrically graded h-p’ method [1]. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The interval [a, b] is possibly infinite or semi-infinite, but this report considers only finite intervals; and without loss of generality, let [a, b] = [0, 1]. Similarly, the inte- grand could include a weighting factor ω (x), but this is not required here. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The remainder of this section describes some partial answers to these questions, and displays some experiments. The answers deal with the case 1 < α < 1, and the experiments demonstrate the case α = 1/2. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Gaußian quadrature rules are the best possible rules in the sense that they are of maximal degree.3 This is due to the exploitation of the maximal number of degrees of freedom in the choice of their nodes and weights. Gauß methods divide into categories depending on the associated weight function, and whether there are any prescribed quadrature points. For the applications in this report, it is preferable to use the end points of the interval as quadrature points, and a unit weight function is assumed. The appropriate set of rules are called the Gauß–Lobatto rules [6, pages 101–104]. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The ‘h-p’ nomenclature presented here originated in papers by Babuˇska et al. [1, 12, 13, 14], on finite element methods, based on previous work which did not explicitly use this schema. The following discussion of the three methods refers to their use with graded meshes. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | A particular basic rule is decided upon (e.g. Simpson’s rule), and desired accuracy is hopefully attained by simply increasing m (that is, N ). Whatever grading is chosen, the separation of the node points (h) decreases as m is increased, hence the name ‘h method’. For a uniform mesh (which works well for smooth integrands), h = (b 1) is constant. Alternatively, open rules, or Gauß rules can be used, the only important factor is that all the basic rules are of the same type and degree. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | In a p method, again a graded mesh is created. Integration is performed over each mesh interval using a basic quadrature rule on n points. Here, n instead of m is varied by the user. That is, for a given number of mesh subintervals, m, a set of rules of increasing degree (that is n, the number of points involved) is used, until desired accuracy is obtained. The same functional relationship N (m, n) exists. As n is increased, the rules used grow in their degree (p), hence the name ‘p method’ (see also Table 1). |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | To illustrate, consider the family of closed Newton–Cotes rules. Assume that the interval has been subdivided, possibly using an adaptive algorithm that chooses smaller subdivisions where there the integrand has greater derivative. Approximate the integral over each division using the trapezoidal rule (p = 1), and inspect the result. If it is unacceptable, repeat using Simpson’s rule (p = 3). Continue this process until results are acceptable. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The h-p method is the natural combination of the two previous methods. The user may vary both m and n. The idea behind this is to create a composite rule that minimises errors in the approximation, for a given number of node points N . (Experiment demonstrates that this is achievable.) The user chooses a family of basic quadrature rules, then decides how to vary n with mesh interval. As the singularities considered are always at end points, a good choice is to organise small mesh intervals and low degree rules (small n) near the end points, and larger mesh intervals and high degree rules away from them, where the integrand is expected to be smooth. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | A simple choice is to begin with a rule on n = 2 points on the smallest interval, then linearly increase n with the number of the mesh interval. Other discrete integer functions nj, j = 1 : m are easily designed. The only constraint on these functions is that if any error analysis is to be done, there should be some regularity in nj. (Choosing basic rules from the same family facilitates this.) This implementation uses the Gauß–Lobatto rule of degree 1 (n = 2) on the first interval, degree 3 (n = 3) on the second, etc. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Creating the composite quadrature rule is quite difficult. Each of the basic quadrature rules must be appropriately scaled and shifted, and then coincident mesh points must be combined. This is further complicated in the cases of closed meshes, closed quadrature rules, and contour integration, where the end points of various segments of the parameterisation must also be combined. (This is exacer- bated if the contour is closed.) The CBIEM requires all of these to be implemented. The (closed) contours involved have corners, and the integrand will usually have singularities at these corners. As it will be important to keep the collocation points (see 3.2.3) between, and not on, the corners, the underlying meshes must include end points. This means that the basic quadrature rules must be closed, so as to include the end points. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Consider an algebraic grading, using closed basic quadrature rules. For an h or h-p method, the integral on the jth interval is computed using a quadrature rule with nj > 2 points. For an h method, nj is constant; for instance, nj ≡ 2 means that the integral over each mesh interval is computed using the trapezoidal rule – the rule is a composite trapezoidal rule. The degrees of some common quadrature rules are presented in Table 1. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | When γ = (p + 1) / (α + 1), Em 6 C ln (m) /mp+1. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The integral converges if: γ (α + 1) (p + 2) < 1, and in this case, it converges − 1, independent of m. Absorbing this into the to the constant [(p + 1) γ (α + 1)]− main constant, the result for 1 6 γ < (p + 1) / (α + 1) is created. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | This is not surprising, as the integral itself becomes unbounded. A similar result to Theorem 2.2 exists for a geometric mesh. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | In summary, using an algebraic mesh of m subintervals, and an h method with α integrands, a degree p quadrature rule on each mesh interval, can achieve, for convergence rate. This is a significant improvement on the equiva- an (α+1) lent result for a uniform mesh, which is . The exponential convergence O rate for smooth integrands is not achieved, but can be, using an h-p method. The αf (x), for smooth functions methods may be extended to integrands of the form f . |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The error for an algebraic mesh is polynomial, whilst the error for a geometric m2/2 are plotted for a mesh is exponential. If the errors with increasing N geometric mesh, an error of the form Em 6 Cρ√N is observed (see Figure 2), for some ρ < 1. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Application of the result in (7) allows |e1| to be bounded by a constant multiple of this C1. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Babuˇska et al. [1, 12, 13, 14] describe an approximation theory for the h, p and h-p methods for the finite element method. This material may be able to be simplified and adapted (as the theory for integration should be easier than for approximation), and also used in error analysis. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Figure 1 compares convergence rates for various choices of the algebraic grading parameter γ, whilst holding constant the number of points in the quadrature rule used on each interval; that is p = 6 is fixed. Figure 2 shows similar data, but varies p. (Here, γ is allowed to vary, and is chosen to be equal to p for convenience.) Both plots are shown compared with the corresponding h-p result, using a geometric grading, and σ = 0.15. The linear increase in number of points used in the quadrature rule means that nj = j + 1, j = 1 : m, so N = m (m + 1) /2 + 1 (as rules are closed, but behaviour. ends are not), and thus m The h-p data does not show as a straight line, but this can be seen in Figure 3, where it is plotted versus √N . |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | matlab code used (cint.m and funcci.m) is contained in Appendix A. Error results for an h-p method using Gauß–Lobatto quadrature rules are presented in Figure 3. The mesh is geometrically graded, with parameter σ = 0.15. For a segment of a closed contour, with a corner at either end, let D be chosen as the number of mesh intervals between each corner and a wide, central interval, so m = 2D + 1 is the number of mesh intervals over that segment. Here, as the contour is the unit circle, 2 artificial corners are placed, and D is varied from 8 to 15. The grading of 2.6.3 – the number of points used the quadrature rules is similar to that used in in the quadrature rule increases linearly with the number of mesh intervals from the nearest corner, starting at 2 on the interval nearest the corner, and finishing at D + 2 on the central interval. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Convergence is plotted for the logarithm of the error with √N . Observe that . These superb results show that the plot is linear, that is, the error is O the method is excellent for the numerical approximation of closed complex con- tour integrals. This success motivates the use of the h-p method in the CBIEM, where quadrature rules for complex contour integrals are required in the numerical approximation of the solution to an integral equation. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The CBIEM is a technique which numerically approximates the solution of the Dirichlet problem.9 It reformulates the solution of the Dirichlet problem as the real part of a function which can be found as the solution of a complex boundary integral equation. The solution of a discretised version of this integral equation is then found using a collocation technique. Finally, a discretisation of Cauchy’s integral formula is used to approximate the solution to the original problem at interior points, based on the approximate boundary data. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | It is related to the ‘Complex Variable Boundary Element Method’ [15], which is a Galerkin version of the same technique, using ‘hat’ functions as a basis. (The collocation method creates an approximation to the boundary data by interpolating from known data, whilst the Galerkin constructs an approximation in terms of a series of basis functions defined on segments of the boundary.) As originally stated, the CVBEM only works on polygonal domains,10 whilst the CBIEM is more general in that it also works on non-polygonal domains. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Thus, the problem is to find the solution to Laplace’s equation over a region, given functional data around its perimeter. This has many applications in the so- lution of potential problems, such as electrostatics and fluid flow. The value of U at points interior to Ω is determined by the boundary data being ‘diffused’ from the boundary inwards, according to the Laplacian operator. It turns out that the problem has a unique solution for all cases of f . In all but the most trivial cases, this solution is not expressible in closed form, and a numerical approximation is re- quired. With sufficient (possibly enormous) computational effort, an approximation to any degree of accuracy can usually be obtained. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Problems with ‘corner singularities’ are of particular interest. In these problems, U is differentiable in the interior, but U becomes unbounded as the corner is ∇ approached. It is known that this behaviour is typical of solutions to the Dirichlet problem on domains with corners. Even if the boundary data is smooth, U still becomes singular near the corner. Numerical methods must be able to produce good approximations to U , in spite of the corner singularities. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The usual boundary integral methods based on Green’s functions lead to a kernel with a logarithmic singularity, even on a smooth domain. This is tedious to program, and computationally inefficient if high order methods are used. If the CBIEM is used with singularity subtraction, the integrands are smooth and can be done simply and accurately by direct quadrature. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The problems caused by the corners and corner singularities are dealt with using h-p quadrature methods, and would be difficult to implement with other types of integral equations [3]. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The solution to the Dirichlet problem, U , is harmonic, as it satisfies Laplace’s equa- C. Now U can be thought of as the real tion in Ω. Identify x component of an analytic function W (z) = U (z) + iV (z), where V is uniquely determined to within a constant. V can be made unique by requiring V (ζ0) = 0 for f (z) is immediately known. The Γ, U (z) some ζ0 ∈ CBIEM first approximates V (z) on Γ, and then uses Cauchy’s integral formula to approximate W , and hence U , at points within Ω. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Figure 4: CBIEM nomenclature. Corner, mesh, node and collocation points on contour Γ, about a region Ω. The lengths marked on segment 1 are the positions of geometric mesh points, in terms of a unit arc length on that segment. The node points correspond to a closed Newton–Cotes rule on each mesh interval. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | In summary, discretisation of Cauchy’s integral formula leads to a linear system, the solution to which is an approximation to W at the N node points on Γ. This approximation can be used to approximate W , and hence U , at points within Ω. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | More sophisticated interpolations to the Vk 1/2 could involve higher degree poly- nomials, splines, or trigonometric polynomials. Initially, the linear choice will be used to illustrate the process. In 3.5.1, the method is extended to higher degree polynomials. For a particular problem (Ω, f ), there is an optimal choice of degree for the interpolation, as errors incurred by interpolation increase with the degree, and eventually become of greater magnitude than those due to discretisation. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | CN,1 − d2 − d3 − ... dN dN − Solution of the order N and hence W (W at the N node points). |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Application of the CBIEM yields different quality results depending on the conti- nuity of the model problem and the contour. In the simplest case, both are smooth, there is no singularity, and standard quadrature gives good results. In fact, because of periodicity, even the trapezoidal rule on a uniform mesh gives very good results. There is no need to use h-p quadrature, but it will still work well. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | If W is smooth, but Γ is not (has corners), in general, the errors will be expected to increase with the sharpness of the corner. The most difficult cases are cusps or reentrant corners (e.g. the corner in a cardioid). Even for a model problem with a smooth solution U , a corner singularity in V (and hence W ) will occur. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Let r represent radial distance from a corner. It is known [23, pages 257–259], χ) that at a corner with interior angle (1 1, so the form is r1/2. A will be found. At worst, for a reentrant corner, χ = good model problem is thus a contour with a corner where the true solution has r1/2. This is obtained, for example, using W (z) = z1/2. If a local behaviour U uniform mesh were used, the greatest component of the error in V will come from the intervals adjacent to the corner. The use of a geometrically graded mesh reduces this component to a level comparable with that of other mesh intervals. Errors will not be of the very high order that is expected for smooth contours, but should still be acceptable. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | To extend the technique described in 3.2.4, the linear interpolation in (12) is re- placed by a higher degree interpolation. The net result of this is to change the definition of the matrix B in (15). Other possible techniques of improving the ac- curacy of the interpolation, such as splines, are not considered here, as they are difficult to implement. The principle involved is that increasing the order of the interpolatory polynomial should reduce the discretisation error, which is expected to be greatest on the largest intervals. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Let Fk be the kth row of a table F , and let the above λj be the jth element of the kth row of another table, L, of the associated weights. The notations F (k, j) and L (k, j) are used to describe the set of O nodal indices and weights associated with the interpolation at point ζk 1/2, where k = 1 : N and j = 1 : O. The structure of F becomes more complicated with increasing O and with increasing number of corners. Details of its construction are not provided here, but can be read from the program cbiem.m in Appendix A.1. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The matrix B is required in the construction of the linear system in (16), and is the only thing that changes when O is varied. For the case of linear interpolatory polynomials (O = 2), a formula involving the terms Hk is used to approximate the value of V at the collocation points. For larger O, this is replaced with a considerably more sophisticated formula. As O increases, the bandwidth of B increases. Naturally, this new B simplifies to the earlier definition if O = 2 is used, but is obtained at greater computational expense. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Construction of the real order N matrix with (k, j)th entry L (k, j) VF (k,j) is re- quired. Multiplying each row by the corresponding Hk converts this to B. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Additionally, in the first O/2 rows, the first of these O entries is shifted to the N th column, due to a ‘wraparound’ effect. In view of the banded structure of B, it appears that a method designed to exploit this structure would be appropriate in the solution of (18). Unfortunately A is neither banded, nor of a particularly simple structure, so this approach is nontrivial, and is a direction for further work. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The discussion and results presented in 2 prompt the use of a geometrically graded mesh, with the number of points in the (closed) quadrature rule on each mesh interval linearly increasing with interval number from the corner, beginning with 2 adjacent to the corner, and becoming D on the central (widest) interval. The use of a linear grading is found in the literature of the finite element method and the usual boundary element method [1, 19]. Changing from linear to quadratic or higher degree may reduce the errors, but its implementation is beyond the scope of this report. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | corners, with m = 2D + 1 intervals. Basic quadrature rules on nj = 2, 3, . . . , D − 1, D, D 1, . . . , 3, 2 points, are used over intervals j = 1 : m. After all of the common end points are considered, the total number of points in the final composite quadrature rule for that segment is S = (D + 1)2 +1. Slightly different results would apply if the quadrature rules were open. Recall that this choice is rejected, as it leads to node points that avoid the singularity. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | S imposes a limit on O, the number of points used in the interpolation rule. As O is usually small (for reasons of computational efficiency, typically O 6 6), this limitation is usually not significant. For example, if D = 2, O 6 8, and if D = 3, O 6 16. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Within cbiem.m is a description of its input parameters. For test problems, ˆV, and where the true solution is known, it plots and calculates norms of V also calculates some discretisation errors. Explicit computation of the approximate solution within Ω is not performed. Experiments with doing this demonstrate that the error results obtained are of the same order as those returned. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | An appropriately weighted discretisation of the L2 norm might seem more appro- priate, but would effectively only present the norm over the central interval, as the widths of the end point intervals are very small. The use of an infinity norm is also appealing, but the 2 norm allows the user to experiment with interpolation formula gradings, to independently reduce the error over different regions of the contour (see 4.2.6). Also, experimental data shows the behaviour of the infinity norm is very § similar to that of the 2 norm. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Tables 2 to 10 present error results for the teardrop contour, for three different model problems: W (z) = z2, z1/2 and z1/4; various choices of two mesh grading parameters (σ and D); and choices of O, the number of points used by the interpo- latory polynomial in the collocation process. In each table N = (D + 1)2 is the size of the linear system being solved, such that there are 2D + 1 mesh intervals between one corner and the next (see 3.5.2). The nine tables cover three illustrative choices of the mesh parameter σ for each of the three model problems. In each case, the results presented are for a choice of σ close to the optimal σ, and two nearby values of σ that demonstrate the increase in the error in each direction. Results have been selected from a much larger data set. Within each table, the minimum error result is emboldened. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | This test function does not have a singularity, and the results are good. Despite the corner, the error reduces with increasing either D or O, until a point is reached where roundoff error, caused by excessive order in the interpolatory polynomial, begins to encroach. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | This case has a corner singularity, and represents the ‘worst’ that singularities get in practice (that is, for zα singularities, in practice α > 1/2). Although the error decreases with increasing D or O, it does so more slowly than for z2, and is orders of magnitude larger. As for z2, there comes a point where increasing O causes the ˆV for the case error to increase, and indeed grow exponentially. The results for V σ = 0.10, D = 9, O = 6 are plotted in Figure 7. The abscissae are plotted uniformly, for if they were plotted versus parameter t, the geometric grading would bunch up most of the results at the ends (corner of teardrop). Observe that these results are, as expected, antisymmetric. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | A z1/4 singularity is beyond the range of singularities expected for smooth test functions. The errors are worse again than for z1/2, and they do not decrease as fast with increasing D or O. In fact, when the test problem is this pathological, the Dirichlet problem is fast becoming a boundary layer problem, which should be dealt with using more specialised methods. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The minimum error results obtained for each test problem are plotted versus σ in Figure 8, and demonstrate that there is an optimal choice of σ, which varies significantly with the test problem. The use of the CBIEM in applications, where the true solution is not known in advance, could falter on the setting of σ. If the computational cost is to be minimised, then it is important to find the optimal σ, however, it may be expensive to try many σ until the optimal one is found. The literature does not justify a choice of σ, but merely states it, e.g. [22] uses σ = 0.15 for a particular (finite element) application. The optimal choice of σ for the paradigm test problem z1/2 is σ = 0.10. As z1/2 is the worst singularity expected in practice (see 3.4), this should be a good guide as a starting guess for any problem with an unknown solution. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Consider Table 6, where the best error result is obtained using O = 6. The error may be able to be reduced by grading the order of the interpolatory polynomial over the mesh intervals. Near the corner, the use of high order interpolation may actually increase the component of the error, although this may be appropriate far away from the corner. It may be ideal to grade the order of the interpolatory polynomial from O = 2 near the corner, to O = 6 (or greater) farthest from the corner. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Direct implementation of this result requires extensive modification to the matrix B used by the CBIEM21 (see 3.5.1), and is beyond the scope of this report. Another way of achieving the same effect is to calculate B matrices B2, B4, . . . , B16 for O = 2, 4, . . . , 16, then construct a new B from appropriate rows of them, and insert this new B at the relevant point in the CBIEM. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Figure 9: The best error results for z1/2, D = 9, σ = 0.10 and a graded interpolation rule. c.f. Figure 7. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | However, intuition is misleading here. The minimum error in Table 6 is 2.4158 × 5, using σ = 0.10 and D = 9. This corresponds to O = 6 on each mesh inter- 10− val. Many experiments in variation of the order of the interpolation rule, hold- ing fixed σ and D, find that the very best error result that can be obtained 5 (a 28% reduction), using a grading with interpolation rules of is 1.7444 O = 12, 2, 2, 6, 6, 6, 10, 10, 10 over the 9 mesh intervals from the corner to the cen- tre. Surprisingly, the component of the error over the first interval decreases with increasing O. This is depicted in Figure 9, which shows that the error is uniformly distributed around the contour, except for the largest component, at the corner. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | It appears that what is happening is that the method has come up against a discretisation error barrier. For this problem, the discretisation error does not decrease particularly quickly, and is a maximum at the corner. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The technique appears to be computationally wasteful, but may be worth investigating, as it would be simpler to implement. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | The same collocation process previously used to approximate V can in this case be used to approximate U . |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Beyond this, the technique is particularly applicable to free boundary prob- lems [8], and may be able to be generalised to other elliptic (and possibly other second order) operators. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | avoid the branch cut on the negative real axis, and make it ACW in orientation. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Density of the geometric mesh, a positive integer. Choice of D forces N, the size of the linear system being solved, to be N = NC.(D+1)^2. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | from the nearest corner. 0 < sigma < 0.5. Try sigma = 0.25 as a starting guess. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Order of the interpolatory polynomial used to aproximate V at the collocation points. Actually the (even) number of nearest points used, thus 2 is linear, 4 is cubic. Must be kept 2 <= O <= NS+1; in practice keep O <= 20. This limits O <= 8 for D = 2, 16 for D = 3, etc. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Number of points on each side of the contour. N node points create NS = (N+1)^2 + 1 mesh points on each segment. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Generate quadrature rule for a geometrically graded mesh, on one segment in the t domain using hprmesh, then insert this into the grid with corners, using rmesh. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Compute the (complex) values of zn and zc (z at the node and collocation points specified by tn and tc, respectively). zc are found as the midpoints of zn in an arc-length sense, by mapping the midpoints of tc to the contour. Also calculate gdot, which is used to modify wn. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | zn = - sin(3*ptn) - 5 * i* tn .* (1 - tn) .* sin(2*ptn); zc = - sin(3*ptc) - 5 * i* tc .* (1 - tc) .* sin(2*ptc); gdot = - 3 * pi * cos(3*ptn) - 5 * i * ... |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | a = 7/24; zn = tn .* (tn - a) .* (tn - 1 + a) .* (tn - 1) + ... |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | zc = tc .* (tc - a) .* (tc - 1 + a) .* (tc - 1) + ... |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | gdot = (tn - a) .* (tn - 1 + a) .* (tn - 1) + ... |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Set up the complex order N matrices A and B. Establish F, a matrix of indices to be used in calculating B, using Ft, a submatrix of the pattern of indices for one edge. Also compute L, the matrix of the coefficients of the interpolatory polynomial, using the indices contained in F. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Compute the matrix B. J is a shift vector, used to create rows of B from rows of L. Multiply the temporary result through by H, then write out C. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Given C, establish the real, order N-1 matrix, Cstar, then calculate Vn, using Vn(N) = 0. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Error norms. Vne and Vce are the differences between the true and the computed values of V at the node and collocation points respectively. r is the discretisation error in the CBIE. p is the residual in the (above) computation for Vtn, when the true soln at the node points is substituted into the equations. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Get tables of nodes and weights for quadrature rules. User inputs maximum number of points required, and the type required. The default type is Gauss--Lobatto (QCase = 2), as it is of higher order than Newton--Cotes. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Create a new quadrature rule, based on a mesh G, where between points G(i) and G(i+1) is a (closed) quadrature rule of Gauss--Lobatto type on S(i) points, including the 2 end points G(i) and G(i+1); for i = 1:length(G)-1. If IClosed is 1, then the contour is closed, and the ends are tied together. function is a generalisation of rmesh. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Return the weights w and points x of the n-point Gauss--Lobatto quadrature rule on the interval [a, b]. See G. H. Golub, SIAM Review 1973 p 318. |
Processed_Geometrically_Graded_h-p_Quadrature_Applied_to_the.txt | Create a new quadrature rule, based on a mesh G, where a (closed) quadrature rule (t1, w1) is inserted over each interval of G. and the ends are tied together. into hprmesh. Originally conceived by Graeme Chandler. |
Processed_A_constraint_on_a_varying_proton--electron_mass_ra.txt | A molecular hydrogen absorber at a lookback time of 12.4 billion years, corresponding to 10% of the age of the universe today, is analyzed to put a constraint on a varying proton–electron mass ratio, µ. A high resolution spectrum of the J1443+2724 quasar, which was observed with the Very Large Telescope, is used to create an accurate model of 89 Lyman and Werner band transitions whose relative frequencies are sensitive to µ, yielding a limit on the relative deviation from the current laboratory value of ∆µ/µ = (−9.5 ± 5.4stat ± 5.3sys) × 10−6. |
Processed_A_constraint_on_a_varying_proton--electron_mass_ra.txt | The accelerated expansion of the universe is ascribed to an elusive form of gravitational repulsive action re- ferred to as dark energy [1]. Whether it is a cosmological constant, inherent to the fabric of space-time, or whether it may be ascribed to some dynamical action in the form of a scalar field φ [2], is an open issue. In the latter case it has been shown that the interaction of the postulated quintessence fields φ to matter cannot be ignored, giving rise to a variation of the fundamental coupling constants and a breakdown of the equivalence principle [3, 4]. In this context it is particularly interesting to probe possible variations of the fundamental constants in the cosmolog- ical epoch of the phase transition, going from a matter- dominated universe to a dark energy-dominated universe, covering redshift ranges z = 0.5 5 [5]. While models of Big Bang nuclear synthesis probe fundamental con- stants at extremely high redshifts (z = 108) [6], the Oklo phenomenon (z = 0.14) [7] and laboratory atomic clock experiments (z = 0) [8] probe low redshifts. Absorbing galaxies in the line-of-sight of quasars are particularly suitable for investigating the range of medium–high red- shifts, for a varying fine-structure constant, α, based on metal absorption [9] and for a varying proton–electron mass ratio, µ = mp/me, based on molecular absorption [10]. Furthermore, unification scenarios predict that vari- ations of α and µ are connected, while in most schemes µ is a more sensitive target for varying constants [11]. |
Processed_A_constraint_on_a_varying_proton--electron_mass_ra.txt | Here we constrain variations in µ at substantially higher redshift by analyzing an H2 absorber at z = 4.22 along the sightline towards the background (z = 4.42) quasar PSS J1443+2724 [22]. This step to higher redshift is challenging for several reasons. Firstly, more distant quasars are typically fainter, making initial discovery of the H2 absorption more difficult and requiring longer in- tegration times for a high-quality spectrum with which to constrain ∆µ/µ. Secondly, absorption lines from neutral hydrogen (Hi) in the intergalactic medium are more nu- merous at higher redshifts, complicating analysis of the H2 transitions. The Lyman-series transitions from the many unrelated Hi clouds in the intergalactic medium form a characteristic ‘forest’ of broader spectral features against which the H2 have to be identified and analyzed (see Fig. 1). However, the ‘comprehensive fitting’ method of simultaneously treating this Hı forest and the H2 ab- sorption, developed previously and documented exten- sively [18, 20, 23, 24], is employed here to reliably meet this challenge. |
Processed_A_constraint_on_a_varying_proton--electron_mass_ra.txt | FIG. 1. Part of the final, fitted spectrum of J1443+2724. The sticks indicate the 3 velocity components of each H2 transition from the rotational levels J = (0 − 4). The residuals are shown above the spectrum for the regions fitted. The broad features that surround and overlap the fitted H2 transitions are multiple, unrelated Hi Lyman-series absorption lines arising in other absorbers along the line of sight. |
Processed_A_constraint_on_a_varying_proton--electron_mass_ra.txt | TABLE I. Results of fitting models of increasing complexity to the spectrum. The two velocity features (see Fig. 2) re- quire fitting at least 2 velocity components (VC) but as it can be seen from the goodness-of-fit measure χ2 ν (where ν is the number of degrees of freedom; here ν ∼ 4600) models with 3, 4, and 5 VC replicate the data better. For each model, a resulting ∆µ/µ value is shown in the last column. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Abstract. We report the first observation of a ~27 day period component in daily maximum temperature recorded at widely spaced locations in Australia. The ~27 day component, extracted by band pass filtering, is correlated with the variation of daily solar radio flux during years close to solar minimum. We demonstrate that the correlation is related to the emergence of regions of solar activity on the Sun separated, temporally, from the emergence of other active regions. In this situation, which occurs only near solar minimum, the observed ~27 day variation of temperature can be in-phase or out-of-phase with the ~27 day variation of solar activity. During solar maximum correlation of temperature and solar activity is much less defined. The amplitude of the ~27 day temperature response to solar activity is large, at times as high as 6oC, and much larger than the well documented temperature response to the ~11 year cycle of solar activity. We demonstrate that the ~27 day temperature response is localised to the Australian continent and increases towards the centre of the continent. Several aspects of the ~27 day temperature response are consistent with a ~27 day period solar input to the Earth’s continents generating large scale planetary waves. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | the Arctic (opposite phase) during 1995-2005 and interpreted this as due to the effect of solar activity on the global electric circuit. Takahashi et al (2010), compared FFT of F10.7 index and OLR over the interval 1980 to 2002 to demonstrate a ~27 day period variation in cloud amount in the Western Pacific but did not propose a mechanism. In this paper we provide evidence of a significant component in Australian surface temperature that is In Section 2 of the paper we correlated with the ~27 day variation of solar activity. describe a band pass filter method of comparing solar activity and daily maximum temperature, TMAX, and explain why the method is superior to other methods of detecting association of climate with short term solar activity. In the results section we (1), demonstrate correlation of the filtered components of F10.7 flux, sunspot area (SSA) and daily maximum temperature, TMAX; the ~27 day component of TMAX to solar activity; (3), find the spatial distribution of the ~27 day temperature response to solar activity, (4) compare the variation in TMAX directly to images of sunspot activity on the Sun and, (5) use SSA data to extend the study of association of TMAX and solar activity back to 1870. In the discussion section the observations of this paper are related to the Sun induced standing planetary wave mechanism, King et al (1977), Volland (1979). |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Figure 1. Approximately one ~11 year cycle of F10.7 solar radio flux indicating the relative amplitudes of the ~11 year cycle and the ~27 day cycle. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | One might expect, therefore, that when a frequency spectrum is obtained over a F10.7 index record that peaks of comparable amplitude, about 50 sfu at ~11 years and about 30 sfu at ~27 days, would be obtained. The FFT for the interval 1987 to 2012 is shown in Figure 2. The ~11 year peak is, as expected, about 50 sfu but the ~27 day peak is about 3 sfu. It is clear that the efficiency of detecting the ~27 day variation is low. The reason for this is now examined. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Figure 2. FFT of the daily F10.7 solar radio flux in the years 1987 to 2012. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | emergence and evolution of individual sunspot groups or active regions on the Sun; in 1975 three sunspot groups evolved; in 2010-2011 five sunspot groups evolved. During solar minimum sunspot group evolutions (SGE) are relatively infrequent and the modest overlapping of one evolution with another is evident in Figure 3. At solar maximum, when several sunspot groups are evolving simultaneously, the situation is more complex. The time axis in each graph of Figure 3 is spaced at 27 day intervals indicative of solar rotation. It is evident that the phase lag relative to solar rotation is different for each SGE. For example, in 1975, the phase of the variation during the first evolution is shifted by about radians relative to the phase of the variation during the third evolution. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Figure 3. The emergence and evolution of sunspot groups (numbered) can be followed by the variations in F10.7 solar radio flux. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | N is the number of SGE, but as N1/2 as in a random walk or Markov process. As the length of the record over which the spectrum is obtained increases the efficiency of detecting a component at ~27 days period decreases as 1/ N1/2. For example, the average duration of a SGE is about three solar rotations or 81 days. Therefore, during the 9125 day interval of 1987 to 2012 about N = 113 SGE are expected to occur. The efficiency is 1/ N1/2 = 0.094 and the expected peak height at ~27 day period is about 30 x 0.094 = 2.8 sfu, about the level observed in Figure 3. This decrease in the efficiency of spectral analysis obviously applies to any record of solar activity that depends directly on SGE such as F10.7, SSN, and SSA records. However, it also applies to any variable that is connected with the solar activity. Therefore the efficient use of spectral analysis in detecting ~27 day period association between solar activity and climate variables is limited to record lengths encompassing one, or at the most a few SGE. For example, wavelet analysis would be efficient provided the wavelet duration is comparable with the average duration of a SGE, typically three solar rotations or 81 days. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Correlation and epoch superposition studies that compare the solar activity variation itself with climate variation, e.g. King et al (1977), Hood (2003), Burns et al (2008), would, at first sight, seem to be unaffected by the above considerations. However, in this paper we will show that there is a variation in phase lag between solar activity and temperature for different SGE. The cause of this variable phase lag must be different from the effect discussed above and is discussed later. However, the effect on the efficiency of averaging methods such as correlation and epoch superposition is essentially the same as the effect on frequency analysis i.e. the efficiency of detecting association decreases as the length of the record or the number of SGE encountered increases. For this reason, in this paper, we detect association between solar activity and temperature by direct comparison, over time intervals of a few SGE, of band pass filtered or unfiltered records of solar activity and surface temperature. Unfiltered records of solar activity are used where appropriate to: (1), minimise the processing of data; (2), retain detail useful in confirming association, and (3), to avoid confusing the band pass filter response to a spike in activity (a sinc function) with the band pass filter response to the evolution of a sunspot group over several solar rotations (similar to a sinc function). |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | variation clearly associated with the intervals of stronger F10.7 variation. It is evident that, during 1975 and 1986, the intervals of stronger F10.7 activity correspond to intervals of SGE and that the average SGE lifetime is about three or four solar rotations. The change from mainly in-phase variation (positive correlation) of F10.7 and 27TMAX to mainly out-of-phase variation (negative correlation) of F10.7 and 27TMAX is also evident in other years, e.g. 1977 and 1985, provided one SGE does not overlap with the previous or the next SGE. Other examples, shown later in this paper, occur in 1927, 1928 and 2011. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Figure 5. Band pass filtered components of F10.7 solar radio flux and daily maximum temperature in Melbourne for 1969, 1975 and 1986. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | between F10.7 and 27TMAX is less defined. Comparison of the band pass filtered components of both variables, Figure 5, allows better definition of the phase lag associated with each SGE. Variation near in-phase or near out-of-phase predominates, and near quadrature phase can occasionally be discerned e.g. the first part of 1975. In the transition interval between one SGE and the next the phase lag is changing and indefinite. Observations of F10.7 and 27TMAX that illustrate the situation, near solar maximum, when SGE overlap are shown in Figure 6. A noticeable feature in Figure 6 is that both the amplitude of F10.7 and amplitude of 27TMAX are relatively constant during both 1974 and 1980. This can be compared with the situation near solar minimum, 1975 and 1986 in Figures 4 and 5, when the amplitudes of F10.7 and 27TMAX increase and decrease in step over the duration of each SGE. Near solar maximum, the variation of phase lag, except in short intervals when the amplitude of F10.7 is high, is complex. As a result, during 1974 and 1980, it is difficult to identify consistent amplitude or phase connection between the two variables or with any SGE. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | The product S33Y*Y regenerates the actual amplitude of the Y variation and the addition of S33Y restores the actual baseline. Any separate contribution to Y can be similarly regenerated so that the ~ 27 day period contribution to TMAX can be readily compared with the overall TMAX variation by use of (2) as in Figure 7. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Figure 7 shows the regenerated ~ 27 day variation of TMAX, known from Figure 5 to be correlated with daily F10.7 values, superimposed on the daily TMAX variation in Melbourne for the years 1969, 1975 and 1986. The graphs also show that the change of the average level of TMAX between summer and winter in Melbourne is about 12oC. The variation in the ~27 day period component of TMAX can be as high as 7oC peak to peak, see for example the last days of 1969 or 1986. Thus the ~ 27 day period solar UV related component can have a very significant effect on day to day temperature in Melbourne to the extent that peak or minimum daily values of TMAX in any given period can often be associated with peaks or minima in 27TMAX. Figures 4 and 5 show that in the second half of 1969 and the first part of 1986 peaks in 27TMAX are consistently associated with peaks in F10.7. However, around day 220 in 1975 and at the end of 1986 peaks in 27Tmax coincide with minima in F10.7. So we have the interesting observation that peaks in solar activity may result in peaks or in minima oftemperature. The reason why the ~27 day component of in-phase or out-of-phase with solar activityduring a SGE is one of the outstanding problems arising from this work and is discussed later. However this observation goes some way to explaining why this very significant ~27 day period variation in surface temperature has not been previously reported. While the correlation may be strongly positive or strongly negative during a SGE, e.g. in the second half of 1969 and in the middle of 1975, the correlation obtained over a long record encompassing many SGE will, as discussed in section 2, tend to zero. For example, the correlation coefficient of F10.7 and 27TMAX obtained over the entire 1969 to 1986 record is 0.07. |
Processed_The_correlation_of_27_day_period_solar_activity_an.txt | Figure 8. Comparison of the variation of the F10.7 solar radio flux, reduced by a factor of 10, (dotted line), and the ~27 day component of TMAX (full line) during the second half of 1969. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.