Title: Summary week 7
1Summary week 7
- Fast Fourier transform
- Example with gravitational waves, and exercise
- Huge enhancement signal/noise for periodic
signals - Wavelets
- Linear operation like FFT
- Daubechy set
- Smoothing operator running weighted average
- Cancels constant/linear terms
- Difference operator
- Enhances the local structures
- Wavelets are located in both domains
- Wavelets have discontinuous derivatives
- Reduction of information
2wavelet transform
- What do these wavelets look like?
- perform DAUB4 and DAUB20 on unit vectors
3wavelets
- cusps non-smooth derivatives
- use of wavelets
- 1) to get strength of located events (e.g. a
signal in a measurement stream) - 2) to render matrices sparse.
4Traveling salesman problem
- Energy function total distance of path
- Steps
- Moving 1 city (6 distances change)
-
5Traveling salesman problem
- Energy function total distance of path
- Steps
- Moving 1 city (6 distances change)
- in principle all paths can be built from this
step - Swapping 2 cities (8 distance changes)
6Traveling salesman problem
- Moving a stretch of k1 cities
- 6 distances to calculate total path lenght, 2
additional distances if reversal of order is
included
7Traveling salesman
- Make choice of design
- I chose Class to read in cities
- IO, distance between 2 cities, index
- Distance I went to polar coordinates and use
- Structures to keep track of different paths
between cities - I implemented Vector of integers, indices of
cities - Array containing distances between cities i and j
- Method to calculate distance for a given path
- Methods to calculate Delta-E for a proposed
pathchange (move, swap, move/reverse stretch) - Store current distance, calculate change
8Traveling salesman
- Make choice of design
- Methods to execute path changes
- Methods for the random-number generation
- I used NRran1 and NRexpdev, that is expensive
- (methods to keep track of statistics)
- In main
- Control value of Delta-E
- Control which methods are used with what
parameters (lenght of the stretch of cities,
reversal, start configurations, number of trials) - Keep track of shortest path encountered
- Verify end result (millions of changes
recalculation of total path is the same within
micrometers). - Next analyze system path lenghts, parameters
9Traveling Salesman
- Start configuration
- Random average distance about 1.2-1.3 Gm
- Short take random city, take closest neighbour
- Average distance 165 Mm
How to calculate amount Of possible
configurations?
Overflows on 64-bit machine Take logarithms of i,
add them, take exponent of remainder
10Traveling Salesman
- Explore strategies values of sigma, number of
trials, methods, path lenghts - I started with 10 start configurations
- Ntrial 20000, sigma 5000 km
- If after 20000 tries no shorter path is found,
- Reduce sigma with factor 3
- Increase amount of trials with factor 2
- Until sigma lt 1 km
- Around 1 million different configurations, takes
a fraction of a second - How many configurations to study?
- Keep track of improvement as a function of amount
of trials
11Parameter determination
method. sigma mean av move1 37 150.6 6.37 mov
e1 1-12 150.2 6.18 swap 37 202.7 8.92 swap 1-12
202.3 8.754 stretch 37 155.0 4.23 stretch 12 15
1.3 4.39 stretch 4 146.4 2.48 stretch 1 143.6 1.
93 all 37 134.1 1.672 all 12 133.6 1.592 all 4
133.0 1.571 all 1 132.5 1.803
All methods together shortest paths Gain from
sigma 12-1 1000 km Computation time loss of
factor 3
12Search strategy
- Shortest path is found when all 3 methods are
used together - Sigma 5000 too large (after 20000 trials paths
still around 700,000 km) - Start with sigma 1000, path lenghts up to 60
- For sigmas smaller than 12 not much improvement
- Trials went up to 640000 failures
- 3 times a shorter path found, average lt1000 km
shorter - Path lenghts reductions for swapping up to 60
cities in initial stages (sigma gt100) - Reductions up to 8 cities sometimes at small
sigma (10 km) - My chosen strategy
- start with sigma of 1000, divide by 2.5 each
refinement - Start with 20000 trials, multiply by 1.5 each
refinement - Stop at sigma lt1, if distance smaller than
129.500 km, try more trials at sigma of 10, 4,
1.5 (triples computation time for shortest 5 of
paths)
13Final results
- 1500 start configurations in 12 minutes
- 20 short start paths pick closest city
- Shortest distance 127096 km
- (better strategy lt126000, worse strategy
gt129000) - Reduction of pathlenghts
method short start path sigma mean RMS 1.6 132.2
1819 4 132.2 1802 10 132.5 1827 25 133.2 1937 6
4 134.8 1876 160 145.8 4098 400 164.3 10580 1000 1
64.3 10580
method all sigma mean RMS 1.6 131.8 1693 4 131.9
1714 10 132.2 1776 25 132.8 1866 64 134.4 2147 160
143.7 2827 1000 310.4 8014
14Final Results
White sigma1 Red sigma4 Blue
sigma10 Turquoise sigma25 Yellow
sigma64 Pink sigma160
15Final results
- 3 distances lt 128000 km in 800 seconds
- 79 configurations lt 129500 km
16Integration of functions (Quadrature)
- intuitive expectation integral is much harder to
calculate than differential - basic task calculate
- chapter 16 integration of differential
equations. - if possible, use this differential form when the
function is concentrated in sharp peaks or when
the shape is strongly scale-dependent. - Already discussed Chebyshev integration
- Fourier integration also methods available
17Classical formulas
18Trapezoidal Rule
- zeroth- order function replaced by sum
19Trapezoidal rule, Simpsons rule
- The trapezoidal rule is correct for first-order
polynomials. It considers 2 points, ½(f0f1),
so it is called a two-point function - With more than 2 points, one can calculate an
integral between these points to higher order.
The three-point formula is correct to order 2
(and due to a cancellation, also to order 3)
Simpsons rule - The four-point formula is also correct up to
order three.
20Classical formulas
- The five-point formula is correct to fifth order
Bodes formula - factors derive from writing down the equations
with arbitrary factors p,q,r,s,... calculate the
integral for functions 1, x, x2, x3, ... and
solving the coupled equations.
21Extended formulas (Closed)
- calculate N times for intervals 0-1, 1-2,...
- Example 5 points calculation of sin(x) between
0,pi and ex and sqrt(x) between 0,2, with
trapezoidal, Simpsons, and Bodes rule
22Extended formulas (Closed)
- calculate N times for intervals 0-1, 1-2,...
- Example 5 points calculation of sin(x) between
0,pi and ex and sqrt(x) between 0,2, with
trapezoidal, Simpsons, and Bodes rule
23Extended formulas (Closed)
- calculate N times for intervals 0-1, 1-2,...
- Example 5 points calculation of sin(x) between
0,pi and ex and sqrt(x) between 0,2, with
trapezoidal, Simpsons, and Bodes rule
24Extended formulas
- Precision is gained, just by weighting the
function values differently! - 1 order of precision obtained, just from relative
weight of end point ( both for Simpson and for
trapezoidal rule). - Open formulas
25Extended formulas (open)
- construct them by using the closed formula for
the interior part and adding the extrapolated
open step for the ends. - end steps done only once. Use 1 order lower
extrapolation than for interior steps. - open formulas 1 order lower precision overall.
26Elementary algorithms
- Starting point extended trapezoidal rule
- one can add new points in between without loosing
the previous work. - error extrapolation is entirely even in powers of
1/N
27Euler-MacLaurin Summation formula
- Only even terms enter in this series.
- calculate integral with N steps (SN), then with
2N steps (S2N)
28Romberg integration
- The leading error in the second step will be ¼ of
the previous step. It can be cancelled by - The next order error is fourth order in 1/N. This
procedure, applied once, is exactly equivalent to
Simpsons rule. - Romberg integration use k successive refinements
to eliminate errors up to O(1/N2k). - Much more precise than trapezoid rule or Simpsons
rule.
29Improper integrals
- Improper
- finite value, but cannot be calculated at
boundary (e.g.) (sinx/x at 0) - one of the limits is infinity
- integrable singularity (e.g. 1/sqrt(x) at 0)
- at boundary
- at known place in interval a,b
- at unknown place
- How to solve we need an algorithm like the
Romberg integration, but for open formula.
Extended midpoint rule also only has even errors -
30Open integration
- It is not possible to double the number of points
by adding intermediate points, but you can triple
them. - the new sum is S(9S3N-SN)/8.
- routine qromo does Romberg-like integration for
open intervals. - infinite boundary-gt change of variables
- if the range a-b contains 0, split the integral
in two parts. (a-1, 1-inf)
31Gaussian quadrature
- Freedom to choose coefficients (Romberg
integration) leads to much faster convergence. - Freedom to choose abscissas twice as much
freedom - you can arrive at higher-order convergence
faster, when integrand is smooth - you can make the integral exact for a class of
functions of polynomialsweight function W
32Gauss-Chebyshev
33Gaussian quadrature
- Chebyshev W(x) 1/sqrt(1-x2)
- W1 tabulated weights qgaus (10 points)
- How to calculate weights and abscissa?
- orthonormal set ltfggt 0 if f not equal g,
ltffgt1 - orthonormal sets can be found by recurrence
34Gaussian quadrature
- The polynomials pj have exactly j distinct roots.
The roots of pj interleave the roots of pj1. - The abscissa in Gaussian quadrature are formed by
the roots (Like in Chebyshev) - The weights can be calculated by e.g.
35Gaussian quadrature
- examples
- routines are provided in the book.
36Multi-dimensional integrals
- number of function evaluations grow like the
power of the dimension (e.g. typically 30 for
1-dim, 30,000 for 3 dim) - boundary may be complicated
- Try always to reduce it to lower dimension.
- (e.g. spherical symmetry polar coordinates)
- complicated boundary resort to Monte Carlo
techniques. - care when function is strongly peaked
- if boundary is simple, Gaussian quadratures may
give a relatively fast answer.
37Multidimensional integration