A Survey of Numerical Methods in Fractional Calculus

Transcription

1 A Survey of Numerical Methods in Fractional Calculus Fractional Derivatives in Mechanics: State of the Art CNAM Paris, 17 November 2006 Kai Diethelm Gesellschaft für numerische Simulation mbh, Braunschweig Institut Computational Mathematics, TU Braunschweig K. Diethelm, Numerical Methods in Fractional Calculus p. 1/21

2 Outline The basic problem Grünwald-Letnikov methods Lubich s fractional linear multistep methods Methods based on quadrature theory Examples Extensions K. Diethelm, Numerical Methods in Fractional Calculus p. 2/21

3 The Basic Problem Find a numerical solution for the fractional differential equation D α y(x) = f(x, y(x)) with appropriate initial condition(s), where D α is either the Riemann-Liouville derivative or the Caputo derivative. K. Diethelm, Numerical Methods in Fractional Calculus p. 3/21

4 Fundamental Definitions Riemann-Liouville integral: J α f(x) = 1 Γ(α) x 0 (x t) α 1 f(t)dt K. Diethelm, Numerical Methods in Fractional Calculus p. 4/21

5 Fundamental Definitions Riemann-Liouville integral: J α f(x) = 1 Γ(α) x 0 (x t) α 1 f(t)dt Riemann-Liouville derivative: D α f(x) = D n J n α f(x), n= α K. Diethelm, Numerical Methods in Fractional Calculus p. 4/21

6 Fundamental Definitions Riemann-Liouville integral: J α f(x) = 1 Γ(α) x 0 (x t) α 1 f(t)dt Riemann-Liouville derivative: D α f(x) = D n J n α f(x), n= α Caputo derivative: D α f(x) = J n α D n f(x), n= α K. Diethelm, Numerical Methods in Fractional Calculus p. 4/21

7 Fundamental Definitions Riemann-Liouville integral: J α f(x) = 1 Γ(α) x 0 (x t) α 1 f(t)dt Riemann-Liouville derivative: D α f(x) = D n J n α f(x), n= α natural initial condition for DE: D α k y(0) = y (k) 0, k =1, 2,..., n Caputo derivative: D α f(x) = J n α D n f(x), n= α K. Diethelm, Numerical Methods in Fractional Calculus p. 4/21

8 Fundamental Definitions Riemann-Liouville integral: J α f(x) = 1 Γ(α) x 0 (x t) α 1 f(t)dt Riemann-Liouville derivative: D α f(x) = D n J n α f(x), n= α natural initial condition for DE: D α k y(0) = y (k) 0, k =1, 2,..., n Caputo derivative: D α f(x) = J n α D n f(x), n= α natural initial condition for DE: D k y(0) = y (k) 0, k =0, 1,..., n 1 K. Diethelm, Numerical Methods in Fractional Calculus p. 4/21

9 Grünwald-Letnikov Methods Integer-order derivatives: D n 1 y(x) = lim h 0 h n k=0 ( 1) k Γ(n + 1) y(x kh) (n N) Γ(k + 1)Γ(n k + 1) K. Diethelm, Numerical Methods in Fractional Calculus p. 5/21

10 Grünwald-Letnikov Methods Integer-order derivatives: D n 1 y(x) = lim h 0 h n k=0 ( 1) k Γ(n + 1) y(x kh) (n N) Γ(k + 1)Γ(n k + 1) Fractional extension (Liouville 1832, Grünwald 1867, Letnikov 1868): DGLy(x) α 1 ( 1) k Γ(α + 1) = lim y(x kh) (α > 0) h 0 h α Γ(k + 1)Γ(α k + 1) k=0 K. Diethelm, Numerical Methods in Fractional Calculus p. 5/21

11 Grünwald-Letnikov Methods Integer-order derivatives: D n 1 y(x) = lim h 0 h n k=0 ( 1) k Γ(n + 1) y(x kh) (n N) Γ(k + 1)Γ(n k + 1) Fractional extension (Liouville 1832, Grünwald 1867, Letnikov 1868): x/h DGLy(x) α 1 ( 1) k Γ(α + 1) = lim y(x kh) (α > 0) h 0 h α Γ(k + 1)Γ(α k + 1) k=0 y defined only on [0, X]; thus assume y(x) = 0 for x < 0. K. Diethelm, Numerical Methods in Fractional Calculus p. 5/21

12 Grünwald-Letnikov Methods Integer-order derivatives: D n 1 y(x) = lim h 0 h n k=0 ( 1) k Γ(n + 1) y(x kh) (n N) Γ(k + 1)Γ(n k + 1) Fractional extension (Liouville 1832, Grünwald 1867, Letnikov 1868): x/h DGLy(x) α 1 ( 1) k Γ(α + 1) = lim y(x kh) (α > 0) h 0 h α Γ(k + 1)Γ(α k + 1) k=0 y defined only on [0, X]; thus assume y(x) = 0 for x < 0. D α GL y(x) = Dα y(x) if y is smooth. K. Diethelm, Numerical Methods in Fractional Calculus p. 5/21

13 Grünwald-Letnikov Methods Exact analytical expression: DGLy(x) α 1 = lim h 0 h α x/h k=0 ( 1) k Γ(α + 1) y(x kh) Γ(k + 1)Γ(α k + 1) K. Diethelm, Numerical Methods in Fractional Calculus p. 6/21

14 Grünwald-Letnikov Methods Numerical approximation: hd α GLy(x) = 1 h α x/h k=0 ( 1) k Γ(α + 1) y(x kh) Γ(k + 1)Γ(α k + 1) for some h > 0. K. Diethelm, Numerical Methods in Fractional Calculus p. 6/21

15 Grünwald-Letnikov Methods Numerical approximation: hd α GLy(x) = 1 h α x/h k=0 ( 1) k Γ(α + 1) y(x kh) Γ(k + 1)Γ(α k + 1) for some h > 0. Error = O(h) + O(f(0)) if f is smooth, i.e. slow convergence if f(0) = 0, completely unfeasible if f(0) 0. K. Diethelm, Numerical Methods in Fractional Calculus p. 6/21

16 Fractional Linear Multistep Methods Generalization of classical linear multistep methods for first-order ODEs (Lubich ) General form: y m = α 1 j=0 y (j) 0 x j m j! + h α m j=0 ω m j f(x j, y j ) + h α s j=0 w m,j f(x j, y j ) K. Diethelm, Numerical Methods in Fractional Calculus p. 7/21

17 Fractional Linear Multistep Methods Generalization of classical linear multistep methods for first-order ODEs (Lubich ) General form: y m = α 1 j=0 y (j) 0 x j m j! + h α m j=0 ω m j f(x j, y j ) + h α s j=0 w m,j f(x j, y j ) initial condition convolution weights starting weights K. Diethelm, Numerical Methods in Fractional Calculus p. 7/21

18 Fractional Linear Multistep Methods Generalization of classical linear multistep methods for first-order ODEs (Lubich ) General form: y m = α 1 j=0 y (j) 0 x j m j! + h α m j=0 ω m j f(x j, y j ) + h α s j=0 w m,j f(x j, y j ) initial condition convolution weights starting weights Computation of convolution weights directly from coefficients of underlying classical linear multistep method K. Diethelm, Numerical Methods in Fractional Calculus p. 7/21

19 Fractional Linear Multistep Methods Computation of starting weights: Linear system of equations K. Diethelm, Numerical Methods in Fractional Calculus p. 8/21

20 Fractional Linear Multistep Methods Computation of starting weights: Linear system of equations Properties of linear system depend strongly on α: K. Diethelm, Numerical Methods in Fractional Calculus p. 8/21

21 Fractional Linear Multistep Methods Computation of starting weights: Linear system of equations Properties of linear system depend strongly on α: Assume underlying classical method to be of order p, and let A = {γ = j + kα : j, k = 0, 1,..., γ p 1}. Then, error of resulting fractional method = O(h p ), dimension of linear system = cardinality of A, condition of linear system depends on distribution of elements of A. K. Diethelm, Numerical Methods in Fractional Calculus p. 8/21

22 Fractional Linear Multistep Methods Further comments on linear system for starting weights: needs to be solved at every grid point, coefficient matrix always identical, coefficient matrix has generalized Vandermonde structure, right-hand sides change from one grid point to another, accuracy of numerical computation of right-hand side suffers from cancellation of digits. K. Diethelm, Numerical Methods in Fractional Calculus p. 9/21

23 Fractional Linear Multistep Methods A well-behaved example: p = 4, α = 1/3 A = { 0, 1, 2, 1, 4, 5,..., 3} card(a) = 10, highly regular spacing of elements of A (equispaced) Matrix of coefficients can be transformed to usual Vandermonde form mildly ill conditioned system special algorithms can be used for sufficiently accurate solution (Björck & Pereyra 1970) K. Diethelm, Numerical Methods in Fractional Calculus p. 10/21

24 Fractional Linear Multistep Methods A badly behaved example: p = 4, α = 0.33 A = {0, 0.33, 0.66, 0.99, 1, 1.32, 1.33,..., 2.97, 2.98, 2.99, 3} card(a) = 22, spacing of the elements of A is highly irregular (in particular, some elements are very close together) severely ill conditioned system no algorithms for sufficiently accurate solution known (Diethelm, Ford, Ford & Weilbeer 2006) K. Diethelm, Numerical Methods in Fractional Calculus p. 11/21

25 Quadrature-Based Methods Direct approach: Use representation D α 1 x y(x) = (x t) α 1 y(t)dt Γ( α) 0 (finite-part integral) in differential equation Discretize integral by classical quadrature techniques Example: Replace y in integrand by piecewise linear interpolant with step size h (Diethelm 1997) product trapezoidal method; Error = O(h 2 α ) Similar concept applicable to Caputo derivatives K. Diethelm, Numerical Methods in Fractional Calculus p. 12/21

26 Quadrature-Based Methods Indirect approach: (Diethelm, Ford & Freed ) Rewrite Caputo initial value problem as integral equation α 1 y(x) = y (k) x k 0 k! + J α [f(, y( ))](x) k=0 Apply product integration idea to J α Example 1: Product rectangle quadrature fractional Adams-Bashforth method explicit method Error = O(h) K. Diethelm, Numerical Methods in Fractional Calculus p. 13/21

27 Quadrature-Based Methods Indirect approach (continued): Example 2: Product trapezoidal quadrature fractional Adams-Moulton method implicit method Error = O(h 2 ) P(EC) m E scheme: Combine Adams-Bashforth predictor and m Adams-Moulton correctors Error = O(h p ), p = min{2, 1 + mα} Similar concept applicable to Riemann-Liouville derivatives K. Diethelm, Numerical Methods in Fractional Calculus p. 14/21

28 Quadrature-Based Methods Common properties: Coefficients can be computed in accurate and stable way without problems Method is applicable for any type of differential equation Reasonable rate of convergence can be achieved K. Diethelm, Numerical Methods in Fractional Calculus p. 15/21

29 Example 1 D α y(x) = x Γ(5 α) x4 α y(x), y(0) = 1 exact solution on [0, 2]: y(x) = x Error at x=2 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 1e-10 α = K. Diethelm, Numerical Methods in Fractional Calculus p. 16/21

30 Example 1 D α y(x) = x Γ(5 α) x4 α y(x), y(0) = 1 exact solution on [0, 2]: y(x) = x Error at x=2 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 1e-10 α = K. Diethelm, Numerical Methods in Fractional Calculus p. 16/21

31 Example 1 D α y(x) = x Γ(5 α) x4 α y(x), y(0) = 1 exact solution on [0, 2]: y(x) = x Error at x=2 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 1e-10 α = K. Diethelm, Numerical Methods in Fractional Calculus p. 16/21

32 Example 2 D α y(x) = Γ(5 + α/2) Γ(9 α) x8 α 3 Γ(5 α/2) x4 α/2 + 9 Γ(α + 1) 4 + ( 1.5x α/2 x 4) 3 y(x) 3/2, y(0) = 0 exact solution on [0, 1]: y(x) = x 8 3x 4+α/ x α Error at x=1 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 α = e K. Diethelm, Numerical Methods in Fractional Calculus p. 17/21

33 Example 2 D α y(x) = Γ(5 + α/2) Γ(9 α) x8 α 3 Γ(5 α/2) x4 α/2 + 9 Γ(α + 1) 4 + ( 1.5x α/2 x 4) 3 y(x) 3/2, y(0) = 0 exact solution on [0, 1]: y(x) = x 8 3x 4+α/ x α Error at x=1 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 α = e K. Diethelm, Numerical Methods in Fractional Calculus p. 17/21

34 Example 2 D α y(x) = Γ(5 + α/2) Γ(9 α) x8 α 3 Γ(5 α/2) x4 α/2 + 9 Γ(α + 1) 4 + ( 1.5x α/2 x 4) 3 y(x) 3/2, y(0) = 0 exact solution on [0, 1]: y(x) = x 8 3x 4+α/ x α Error at x=1 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 α = e K. Diethelm, Numerical Methods in Fractional Calculus p. 17/21

35 Example 3 D α y(x) = x 1 α E 1,2 α ( x) + exp( 2x) y(x) 2, y(0) = 1 exact solution on [0, 5]: y(x) = exp( x) Error at x=5 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 α = e K. Diethelm, Numerical Methods in Fractional Calculus p. 18/21

36 Example 3 D α y(x) = x 1 α E 1,2 α ( x) + exp( 2x) y(x) 2, y(0) = 1 exact solution on [0, 5]: y(x) = exp( x) Error at x=5 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 α = e K. Diethelm, Numerical Methods in Fractional Calculus p. 18/21

37 Example 3 D α y(x) = x 1 α E 1,2 α ( x) + exp( 2x) y(x) 2, y(0) = 1 exact solution on [0, 5]: y(x) = exp( x) Error at x=5 vs. no. of nodes Grünwald-Letnikov direct quadrature indirect quadrature multistep (BDF4) 0,01 0,0001 1e-06 1e-08 α = e K. Diethelm, Numerical Methods in Fractional Calculus p. 18/21

38 Summary Grünwald-Letnikov: reliable but slowly convergent Linear multistep methods: very bad for some values of α, very good for others Quadrature-based methods: useful compromise, in particular because of simple speed-up possibility All types of basic routines available in GNS Numerical Fractional Calculus Library (presently FORTRAN77) K. Diethelm, Numerical Methods in Fractional Calculus p. 19/21

39 Extensions Application of methods to time-/space-fractional partial differential equations, fractional optimal control problems. Common problem: High arithmetic complexity due to non-locality of fractional operators. Solution for quadrature methods: Nested mesh concept (Ford & Simpson 2001). K. Diethelm, Numerical Methods in Fractional Calculus p. 20/21

40 Merci pour votre attention! contact: K. Diethelm, Numerical Methods in Fractional Calculus p. 21/21