1. Introduction to Numerical Methods Andreas Gürtler Hans Maassen (maassen@math.ru.nl) tel.: 52991

2. Plan for today Introduction Errors Representation of real numbers in a computer When can we solve a problem? A first numerical method: Bisection method for root-finding (solving f(x)=0) Practical part: Implementation of the bisection method

3. What are numerical methods for? Many problems can be formulated in simple mathematical equations This does not mean, that they are solved easily! For any application, you need numbers  Numerical Methods! Needed for most basic things: exp(3),sqrt(7),sin(42°), log(5),  Often, modeling and numerical calculations can help in design, construction, safety Note: many everyday’s problems are so complicated that they cannot be solved yet  Efficiency is crucial

4. Numerics is about numbers Numerical methods: Numerical approximation of solutions to understood problems Numerical representation of real numbers has far-reaching consequences Two main objectives –quantify errors Approximation without error estimation is useless –increase efficiency Solutions which take years or need more resources that you have are useless Nowadays, many fields depend on numerics

5. Errors

6. Numerical errors Roundoff error: finite precision numerical calculations are almost always approximations Truncation error: a calculation has to stop Examples: –Approximation (e.g. finite Taylor series) –Discretization It is crucial to know when to stop (i.e. when a calculations is converged!). To check this, change parameters (e.g. step size, number of basis states) and check result. modeling error

7. Truncation errors Truncation errors are problem specific Often, every step involves an approximation, e.g. a finite Taylor series The truncation errors accumulate Often, truncation errors can be calculated

8. Roundoff errors Example: Logistic Map x i+1 =r * x i * (1-x i ) x 0 = 0.7 ; r=4 single and double precision Precision of representation of numbers is finite - errors accumulate a real number x can be represented as fl(x) = x·(1+  ) : floating point computer representation | fl(x)-x| =  x absolute error (often also  x) | fl(x)-x|/x =  relative error

9. Computer representation of numbers

10. back to Matlab: data types Real numbers: floating point representation double (64 bit) standard! single (32 bit) less precision, half memory Integer: signed and unsigned (8,16,32 bit) a=uint8(255), b=int16(32767), c=int32(2 31 -1) Integers use less space, calculations are precise! Complex: w=2+3i ; v=complex(x,y) x,y can be matrices! Boolean: true (=1) or false (=0) Strings: s='Hello World' special values: +inf, -inf, NaN Infinity, Not-a-Number check with isinf(x), isnan(x)

11. floating point standard IEEE 754 Normalized: N=1.f · 2 p Denormalized: N=0.f · 2 p Limits of double precision (64bit) FP numbers realmax (10 308 ), realmin (10 -308 ) largest, smallest real number eps (10 -16 ) accuracy strictly speaking, FP numbers are not associative and not distributive, but they are to a very good approximation if used reasonably mathematically floating point numbers are not a field! Many fundamental theorems of analysis and algebra have to be used with care! Significant f51 – 0 Exponent p62 – 52 sign63 usebit (double)

12. Choose your unit system keep numbers in similar order of magnitude (choose natural units) e.g. atomic units hydrogen energy levels: E n =-1/(2n 2 ) rather than -2.18E-18/n 2 J Note: Calculation of Bohr radius in SI units exceeds range of single precision FP: a 0 =4  0 h 2 /m e e 2  5.3E-11m (involves 10 -78 /2*10 -68 ) 15.3E-11 mdistance a 0 1/1372.99E8 m/svelocity c 14.36E-18 Jenergy E h 19.1E-31 kgmass m e 11.62E-19 Ccharge e a.u.SI -

13. floating point caveats not all real numbers can be represented as FP 0.1+0.1+0.1+0.1+0.1+0.1+0.1+0.1+0.1+0.1-1 -1E-16 Floating point calculations are not precise! sin(pi) is not 0, 1E20+1-1E20 is not 1 never compare two floats (a==b) directly try 100*(1/3)/100 == 1/3 FALSE!!! use abs(a-b)<eps (eps=2.2E-16) be careful with mixing integer and FP i=int32(1); i/2 produces 1 as i/2 stays integer! i/3 produces 0 ! This is much more dangerous in C and FORTRAN etc. Solution: explicit type conversion double(i)/2 produces 0.5

14. What can we solve numerically?

15. What can we solve Suppose we want to evaluate f(x) with perfect algorithm we have FP number x+  x with error  x  f(x) = f(x+  x)-f(x)  f’(x)  x (if f differentiable) relative error: Definition: condition number  >>1 : problem ill-conditioned  small: problem well-conditioned

16. well- and ill-conditioned methods Let’s try a simple calculation: 99-70*sqrt(2) (  0.00505) suppose we have 1.4 as approximation for  2 We have 2 mathematically equivalent methods: f 1 : 99-70*  2 f 1 (1.4) = 1 f 2 : 1/(99+70*  2) f 2 (1.4)  0.0051 Condition numbers: f 1 (x)= 99-70 x    2   20000 f 1 (x)= 1/(99+70 x)    2   0.5

17. what happened? f 1 : 99-70*  2 f 2 : 1/(99+70*  2) Condition number of subtraction, addition: f(x)=x-a   = |-x/(x-a)| ill-conditioned for x-a≈0 f(x)=x+a   = |x/(x+a)| ill-conditioned for x+a≈0 Condition number for multiplication, division: f(x)=ax  = |xa/(ax)| =1 f(x)=1/x  = |xx -2 /(x -1 )| =1 well- conditioned

18. numerical disasters Patriot system hit by SCUD missile –position predicted from time and velocity –the system up-time in 1/10 of a second was converted to seconds using 24bit precision (by multiplying with 1/10) –1/10 has non-terminating binary expansion –after 100h, the error accumulated to 0.34s –the SCUD travels 1600 m/s so it travels >500m in this time Ariane 5 –a 64bit FP number containing the horizontal velocity was converted to 16bit signed integer –range overflow followed from http://ta.twi.tudelft.nl/nw/users/vuik/wi211/disasters.html

19. Root finding: a standard numerical problem A common task is to solve an equation f(x)=0 task: find the zero-crossing (root) of f(x) Pitfalls: f(x) might have –no roots –extrema –many roots –singularities –roots which are not FP numbers –roots close to 0 or at very large x There are different methods with advantages and disadvantages

20. Root finding: bisection method The bisection method finds a root in [a,b] if f(x)  C(a,b), f(a)*f(b)<0 wanted: x 0 : f(x 0 )=0 (There is a root because of the intermediate value theorem) Bisection is slow, but always works! ab a1a1 b1b1 a2a2 b2b2 a 3 p 3 b 3 p1p1 p2p2 f(x)

21. Where does bisection converge to? If the function changes sign asound a root, bisection converges to the root If the function does not change sign, bisection cannot find the root! If the function changes sign around a singularity, bisection converges to the singularity! Important: stop criterion -2012 -0.500.51 -2 0 2 4 6 -100 -50 0 50 100 a p b f(x) problematic for singularities, small f’(x) large a,b p≈0 p≈0, small f’(x)

22. convergence Bisection always converges, but maybe slow f  C[a,b] and f(a)*f(b)<0 then {p n } converges to a root p of f and |p n -p|  (b-a)/2 n Proof:  n  1: b n -a n = (b-a)/2 n-1 and p  (a n,b n ) convergence is linear

23. order of convergence Let converge to p with if positive  exist with then p n converges with order   =1 : linear convergence  =2 : quadratic convergence

24. Implementation of the bisection method f(a)*f(b)<0   p  (a,b): f(p)=0 ab a1a1 b1b1 a2a2 b2b2 a 3 p 3 b 3 p1p1 p 2 Bisection in words: 1.choose an interval (a,b) around the root 2.calculate the center point p=(b+a)/2 3.if p approximates the root well enough, STOP 4.if the root is in [a,p] set b=p 5.if the root is in [p,b] 6. set a=p 7.goto step 2 initialize a,b p=(a+b)/2 while abs(b-a)>  do f(p)*f(a) b=p f(p)*f(b) a=p repeat

25. Exercise Implement the Bisection method in Matlab use it to solve x 3 = ln(1+x) (x>0)

26. solve x 3 = ln(1+x) (x>0) using bisection % bisection.m a=0.1; b=2; p=(a+b)/2; epsilon=1E-10; n=1; while abs(b-a)>epsilon if (f(p)*f(a)<0) b=p; end if (f(p)*f(b)<0) a=p; end p=(a+b)/2; n=n+1; end fprintf ('Root: %f, Steps: %d.',p,n); function ret=f(x) % function of which root should be found ret=x^3-log(1+x); end ab a1a1 b1b1 a2a2 b2b2 a 3 p 3 b 3 p1p1 p 2 Bisection in words: 1.choose an interval (a,b) around the root 2.calculate the center point p=(b+a)/2 3.if p approximates the root well enough, STOP 4.if the root is in (a,p) set b=p 5.if the root is in (p,b) 6. set a=p 7.goto step 2