I'm trying to compute Fractal Dimension of very specific time series array.
I've found implementations of Higuchi FD algorithm:
def hFD(a, k_max): #Higuchi FD
L = []
x = []
N = len(a)
for k in range(1,k_max):
Lk = 0
for m in range(0,k):
#we pregenerate all idxs
idxs = np.arange(1,int(np.floor((N-m)/k)),dtype=np.int32)
Lmk = np.sum(np.abs(a[m+idxs*k] - a[m+k*(idxs-1)]))
Lmk = (Lmk*(N - 1)/(((N - m)/ k)* k)) / k
Lk += Lmk
L.append(np.log(Lk/(m+1)))
x.append([np.log(1.0/ k), 1])
(p, r1, r2, s)=np.linalg.lstsq(x, L)
return p[0]
from https://github.com/gilestrolab/pyrem/blob/master/src/pyrem/univariate.py
and Katz FD algorithm:
def katz(data):
n = len(data)-1
L = np.hypot(np.diff(data), 1).sum() # Sum of distances
d = np.hypot(data - data[0], np.arange(len(data))).max() # furthest distance from first point
return np.log10(n) / (np.log10(d/L) + np.log10(n))
from https://github.com/ProjectBrain/brainbits/blob/master/katz.py
I expect results of ~1,5 in both cases however get 2,2 and 4 instead...
hFD(x,4) = 2.23965648024 (k value of here is chosen as an example, however result won't change much in range 4-12 edit: I was able to get result of ~1,9 with k=22, however this still does not make any sense);
katz(x) = 4.03911343057
Which in theory should not be possible for 1D time-series array.
Questions here are: are Higuchi and Katz algorithms not suitable for time-series analysis in general, or am I doing something wrong on my side? Also are there any other python libraries with already implemented and error-less algorithms to verify my results?
My array of interest (each element represents point in time t, t+1, t+2,..., t+N)
x = np.array([373.4413096546802, 418.58026161917803,
395.7387698762124, 416.21163042783206,
407.9812265426947, 430.2355284504048,
389.66095393296763, 442.18969320408166,
383.7448638776275, 452.8931822090381,
413.5696828065546, 434.45932712853585
,429.95212301648996, 436.67612861616215,
431.10235365546964, 418.86935850068545,
410.84902747247423, 444.4188867775925,
397.1576881118471, 451.6129904245434,
440.9181246439599, 438.9857353268666,
437.1800408012741, 460.6251405281339,
404.3208481355302, 500.0432305427639,
380.49579242696177, 467.72953450552893,
333.11328535523967, 444.1171938340972,
303.3024198243042, 453.16332062153276,
356.9697406524534, 520.0720647379901,
402.7949987727925, 536.0721418821788,
448.21609036718445, 521.9137447208354,
470.5822486372967, 534.0572029633416,
480.03741443274765, 549.2104258193126,
460.0853321729541, 561.2705350421926,
444.52689144575794, 560.0835589548401,
462.2154563472787, 559.7166600213686,
453.42374550322353, 559.0591804941763,
421.4899935529862, 540.7970410737004,
454.34364779193913, 531.6018122709779,
437.1545739076901, 522.4262260216169,
444.6017030695873, 533.3991716674865,
458.3492761150962, 513.1735160522104])
The array you are trying to estimate hDF is too short. You need to get longer sample or oversample the current one to have at least 128 points for hDF and more then 4000 points for Katz
import scipy.signal as signal
...
x_res=signal.resample(x,128)
hfd(x_res,4) will be 1.74383694265
Related
Given n pairs of integers. Split into two subsets A and B to minimize sum(maximum difference among first values of A, maximum difference among second values of B).
Example : n = 4
{0, 0}; {5;5}; {1; 1}; {3; 4}
A = {{0; 0}; {1; 1}}
B = {{5; 5}; {3; 4}}
(maximum difference among first values of A, maximum difference among second values of B).
(maximum difference among first values of A) = fA_max - fA_min = 1 - 0 = 1
(maximum difference among second values of B) = sB_max - sB_min = 5 - 4 = 1
Therefore, the answer if 1 + 1 = 2. And this is the best way.
Obviously, maximum difference among the values equals to (maximum value - minimum value). Hence, what we need to do is find the minimum of (fA_max - fA_min) + (sB_max - sB_min)
Suppose the given array is arr[], first value if arr[].first and second value is arr[].second.
I think it is quite easy to solve this in quadratic complexity. You just need to sort the array by the first value. Then all the elements in subset A should be picked consecutively in the sorted array. So, you can loop for all ranges [L;R] of the sorted. Each range, try to add all elements in that range into subset A and add all the remains into subset B.
For more detail, this is my C++ code
int calc(pair<int, int> a[], int n){
int m = 1e9, M = -1e9, res = 2e9; //m and M are min and max of all the first values in subset A
for (int l = 1; l <= n; l++){
int g = m, G = M; //g and G are min and max of all the second values in subset B
for(int r = n; r >= l; r--) {
if (r - l + 1 < n){
res = min(res, a[r].first - a[l].first + G - g);
}
g = min(g, a[r].second);
G = max(G, a[r].second);
}
m = min(m, a[l].second);
M = max(M, a[l].second);
}
return res;
}
Now, I want to improve my algorithm down to loglinear complexity. Of course, sort the array by the first value. After that, if I fixed fA_min = a[i].first, then if the index i increase, the fA_max will increase while the (sB_max - sB_min) decrease.
But now I am still stuck here, is there any ways to solve this problem in loglinear complexity?
The following approach is an attempt to escape the n^2, using an argmin list for the second element of the tuples (lets say the y-part). Where the points are sorted regarding x.
One Observation is that there is an optimum solution where A includes index argmin[0] or argmin[n-1] or both.
in get_best_interval_min_max we focus once on including argmin[0] and the next smallest element on y and so one. The we do the same from the max element.
We get two dictionaries {(i,j):(profit, idx)}, telling us how much we gain in y when including points[i:j+1] in A, towards min or max on y. idx is the idx in the argmin array.
calculate the objective for each dict assuming max/min or y is not in A.
combine the results of both dictionaries, : (i1,j1): (v1, idx1) and (i2,j2): (v2, idx2). result : j2 - i1 + max_y - min_y - v1 - v2.
Constraint: idx1 < idx2. Because the indices in the argmin array can not intersect, otherwise some profit in y might be counted twice.
On average the dictionaries (dmin,dmax) are smaller than n, but in the worst case when x and y correlate [(i,i) for i in range(n)] they are exactly n, and we do not win any time. Anyhow on random instances this approach is much faster. Maybe someone can improve upon this.
import numpy as np
from random import randrange
import time
def get_best_interval_min_max(points):# sorted input according to x dim
L = len(points)
argmin_b = np.argsort([p[1] for p in points])
b_min,b_max = points[argmin_b[0]][1], points[argmin_b[L-1]][1]
arg = [argmin_b[0],argmin_b[0]]
res_min = dict()
for i in range(1,L):
res_min[tuple(arg)] = points[argmin_b[i]][1] - points[argmin_b[0]][1],i # the profit in b towards min
if arg[0] > argmin_b[i]: arg[0]=argmin_b[i]
elif arg[1] < argmin_b[i]: arg[1]=argmin_b[i]
arg = [argmin_b[L-1],argmin_b[L-1]]
res_max = dict()
for i in range(L-2,-1,-1):
res_max[tuple(arg)] = points[argmin_b[L-1]][1]-points[argmin_b[i]][1],i # the profit in b towards max
if arg[0]>argmin_b[i]: arg[0]=argmin_b[i]
elif arg[1]<argmin_b[i]: arg[1]=argmin_b[i]
# return the two dicts, difference along y,
return res_min, res_max, b_max-b_min
def argmin_algo(points):
# return the objective value, sets A and B, and the interval for A in points.
points.sort()
# get the profits for different intervals on the sorted array for max and min
dmin, dmax, y_diff = get_best_interval_min_max(points)
key = [None,None]
res_min = 2e9
# the best result when only the min/max b value is includes in A
for d in [dmin,dmax]:
for k,(v,i) in d.items():
res = points[k[1]][0]-points[k[0]][0] + y_diff - v
if res < res_min:
key = k
res_min = res
# combine the results for max and min.
for k1,(v1,i) in dmin.items():
for k2,(v2,j) in dmax.items():
if i > j: break # their argmin_b indices can not intersect!
idx_l, idx_h = min(k1[0], k2[0]), max(k1[1],k2[1]) # get index low and idx hight for combination
res = points[idx_h][0]-points[idx_l][0] -v1 -v2 + y_diff
if res < res_min:
key = (idx_l, idx_h) # new merged interval
res_min = res
return res_min, points[key[0]:key[1]+1], points[:key[0]]+points[key[1]+1:], key
def quadratic_algorithm(points):
points.sort()
m, M, res = 1e9, -1e9, 2e9
idx = (0,0)
for l in range(len(points)):
g, G = m, M
for r in range(len(points)-1,l-1,-1):
if r-l+1 < len(points):
res_n = points[r][0] - points[l][0] + G - g
if res_n < res:
res = res_n
idx = (l,r)
g = min(g, points[r][1])
G = max(G, points[r][1])
m = min(m, points[l][1])
M = max(M, points[l][1])
return res, points[idx[0]:idx[1]+1], points[:idx[0]]+points[idx[1]+1:], idx
# let's try it and compare running times to the quadratic_algorithm
# get some "random" points
c1=0
c2=0
for i in range(100):
points = [(randrange(100), randrange(100)) for i in range(1,200)]
points.sort() # sorted for x dimention
s = time.time()
r1 = argmin_algo(points)
e1 = time.time()
r2 = quadratic_algorithm(points)
e2 = time.time()
c1 += (e1-s)
c2 += (e2-e1)
if not r1[0] == r2[0]:
print(r1,r2)
raise Exception("Error, results are not equal")
print("time of argmin_algo", c1, "time of quadratic_algorithm",c2)
UPDATE: #Luka proved the algorithm described in this answer is not exact. But I will keep it here because it's a good performance heuristics and opens the way to many probabilistic methods.
I will describe a loglinear algorithm. I couldn't find a counter example. But I also couldn't find a proof :/
Let set A be ordered by first element and set B be ordered by second element. They are initially empty. Take floor(n/2) random points of your set of points and put in set A. Put the remaining points in set B. Define this as a partition.
Let's call a partition stable if you can't take an element of set A, put it in B and decrease the objective function and if you can't take an element of set B, put it in A and decrease the objective function. Otherwise, let's call the partition unstable.
For an unstable partition, the only moves that are interesting are the ones that take the first or the last element of A and move to B or take the first or the last element of B and move to A. So, we can find all interesting moves for a given unstable partition in O(1). If an interesting move decreases the objective function, do it. Go like that until the partition becomes stable. I conjecture that it takes at most O(n) moves for the partition to become stable. I also conjecture that at the moment the partition becomes stable, you will have a solution.
Problem statement:
We are given three arrays A1,A2,A3 of lengths n1,n2,n3. Each array contains some (or no) natural numbers (i.e > 0). These numbers denote the program execution times.
The task is to choose the first element from any array and then you can execute that program and remove it from that array.
For example:
if A1=[3,2] (n1=2),
A2=[7] (n2=1),
A3=[1] (n3=1)
then we can execute programs in various orders like [1,7,3,2] or [7,1,3,2] or [3,7,1,2] or [3,1,7,2] or [3,2,1,7] etc.
Now if we take S=[1,3,2,7] as the order of execution the waiting time of various programs would be
for S[0] waiting time = 0, since executed immediately,
for S[1] waiting time = 0+1 = 1, taking previous time into account, similarly,
for S[2] waiting time = 0+1+3 = 4
for S[3] waiting time = 0+1+3+2 = 6
Now the score of array is defined as sum of all wait times = 0 + 1 + 4 + 6 = 11, This is the minimum score we can get from any order of execution.
Our task is to find this minimum score.
How can we solve this problem? I tried with approach trying to pick minimum of three elements each time, but it is not correct because it gets stuck when two or three same elements are encountered.
One more example:
if A1=[23,10,18,43], A2=[7], A3=[13,42] minimum score would be 307.
The simplest way to solve this is with dynamic programming (which runs in cubic time).
For each array A: Suppose you take the first element from array A, i.e. A[0], as the next process. Your total cost is the wait-time contribution of A[0] (i.e., A[0] * (total_remaining_elements - 1)), plus the minimal wait time sum from A[1:] and the rest of the arrays.
Take the minimum cost over each possible first array A, and you'll get the minimum score.
Here's a Python implementation of that idea. It works with any number of arrays, not just three.
def dp_solve(arrays: List[List[int]]) -> int:
"""Given list of arrays representing dependent processing times,
return the smallest sum of wait_time_before_start for all job orders"""
arrays = [x for x in arrays if len(x) > 0] # Remove empty
#functools.lru_cache(100000)
def dp(remaining_elements: Tuple[int],
total_remaining: int) -> int:
"""Returns minimum wait time sum when suffixes of each array
have lengths in 'remaining_elements' """
if total_remaining == 0:
return 0
rem_elements_copy = list(remaining_elements)
best = 10 ** 20
for i, x in enumerate(remaining_elements):
if x == 0:
continue
cost_here = arrays[i][-x] * (total_remaining - 1)
if cost_here >= best:
continue
rem_elements_copy[i] -= 1
best = min(best,
dp(tuple(rem_elements_copy), total_remaining - 1)
+ cost_here)
rem_elements_copy[i] += 1
return best
return dp(tuple(map(len, arrays)), sum(map(len, arrays)))
Better solutions
The naive greedy strategy of 'smallest first element' doesn't work, because it can be worth it to do a longer job to get a much shorter job in the same list done, as the example of
A1 = [100, 1, 2, 3], A2 = [38], A3 = [34],
best solution = [100, 1, 2, 3, 34, 38]
by user3386109 in the comments demonstrates.
A more refined greedy strategy does work. Instead of the smallest first element, consider each possible prefix of the array. We want to pick the array with the smallest prefix, where prefixes are compared by average process time, and perform all the processes in that prefix in order.
A1 = [ 100, 1, 2, 3]
Prefix averages = [(100)/1, (100+1)/2, (100+1+2)/3, (100+1+2+3)/4]
= [ 100.0, 50.5, 34.333, 26.5]
A2=[38]
A3=[34]
Smallest prefix average in any array is 26.5, so pick
the prefix [100, 1, 2, 3] to complete first.
Then [34] is the next prefix, and [38] is the final prefix.
And here's a rough Python implementation of the greedy algorithm. This code computes subarray averages in a completely naive/brute-force way, so the algorithm is still quadratic (but an improvement over the dynamic programming method). Also, it computes 'maximum suffixes' instead of 'minimum prefixes' for ease of coding, but the two strategies are equivalent.
def greedy_solve(arrays: List[List[int]]) -> int:
"""Given list of arrays representing dependent processing times,
return the smallest sum of wait_time_before_start for all job orders"""
def max_suffix_avg(arr: List[int]):
"""Given arr, return value and length of max-average suffix"""
if len(arr) == 0:
return (-math.inf, 0)
best_len = 1
best = -math.inf
curr_sum = 0.0
for i, x in enumerate(reversed(arr), 1):
curr_sum += x
new_avg = curr_sum / i
if new_avg >= best:
best = new_avg
best_len = i
return (best, best_len)
arrays = [x for x in arrays if len(x) > 0] # Remove empty
total_time_sum = sum(sum(x) for x in arrays)
my_averages = [max_suffix_avg(arr) for arr in arrays]
total_cost = 0
while True:
largest_avg_idx = max(range(len(arrays)),
key=lambda y: my_averages[y][0])
_, n_to_remove = my_averages[largest_avg_idx]
if n_to_remove == 0:
break
for _ in range(n_to_remove):
total_time_sum -= arrays[largest_avg_idx].pop()
total_cost += total_time_sum
# Recompute the changed array's avg
my_averages[largest_avg_idx] = max_suffix_avg(arrays[largest_avg_idx])
return total_cost
With the Julia Language, I defined a function to sample points uniformly inside the sphere of radius 3.14 using rejection sampling as follows:
function spherical_sample(N::Int64)
# generate N points uniformly distributed inside sphere
# using rejection sampling:
points = pi*(2*rand(5*N,3).-1.0)
ind = sum(points.^2,dims=2) .<= pi^2
## ideally I wouldn't have to do this:
ind_ = dropdims(ind,dims=2)
return points[ind_,:][1:N,:]
end
I found a hack for subsetting arrays:
ind = sum(points.^2,dims=2) .<= pi^2
## ideally I wouldn't have to do this:
ind_ = dropdims(ind,dims=2)
But, in principle array indexing should be a one-liner. How could I do this better in Julia?
The problem is that you are creating a 2-dimensional index vector. You can avoid it by using eachrow:
ind = sum.(eachrow(points.^2)) .<= pi^2
So that your full answer would be:
function spherical_sample(N::Int64)
points = pi*(2*rand(5*N,3).-1.0)
ind = sum.(eachrow(points.^2)) .<= pi^2
return points[ind,:][1:N,:]
end
Here is a one-liner:
points[(sum(points.^2,dims=2) .<= pi^2)[:],:][1:N, :]
Note that [:] is dropping a dimension so the BitArray can be used for indexing.
This does not answer your question directly (as you already got two suggestions), but I rather thought to hint how you could implement the whole procedure differently if you want it to be efficient.
The first point is to avoid generating 5*N rows of data - the problem is that it is very likely that it will be not enough to generate N valid samples. The point is that the probability of a valid sample in your model is ~50%, so it is possible that there will not be enough points to choose from and [1:N, :] selection will throw an error.
Below is the code I would use that avoids this problem:
function spherical_sample(N::Integer) # no need to require Int64 only here
points = 2 .* pi .* rand(N, 3) .- 1.0 # note that all operations are vectorized to avoid excessive allocations
while N > 0 # we will run the code until we have N valid rows
v = #view points[N, :] # use view to avoid allocating
if sum(x -> x^2, v) <= pi^2 # sum accepts a transformation function as a first argument
N -= 1 # row is valid - move to the previous one
else
rand!(v) # row is invalid - resample it in place
#. v = 2 * pi * v - 1.0 # again - do the computation in place via broadcasting
end
end
return points
end
This one is pretty fast, and uses StaticArrays. You can probably also implement something similar with ordinary tuples:
using StaticArrays
function sphsample(N)
T = SVector{3, Float64}
v = Vector{T}(undef, N)
n = 1
while n <= N
p = rand(T) .- 0.5
#inbounds v[n] = p .* 2π
n += (sum(abs2, p) <= 0.25)
end
return v
end
On my laptop it is ~9x faster than the solution with views.
I have a relativelly small vector in Matlab
R = randn(1,1000);
Now I would like to create a much bigger vector by selecting a specified set of elements like so
Q = R([1 5 8 5 8 1 3 4 19 1, etc]);
The number of the selected elements numel(Q) is 1,000,000+, very big. Is it possible to do this step such that the resulting vector Q is automatically a distributed array, ready for parallel processing on a multicore machine?
Thanks!
The approaches mentioned here assumes that you want to have at least R and Q as the distributed arrays.
Approach #1
This solution would be based on this very smart solution -
N = 3;
R = randn(1,N,'distributed');
[~,ind] = sort(rand(numel(R)));
Q = R(ind(:));
Note that for the above code, ind would be on the client side. If you would like to have it as a distributed array too, use this -
N = 3;
R = randn(1,N,'distributed');
ind = ones(N,'distributed');
[~,ind(:,:)] = sort(rand(numel(R)));
Q = R(ind(:))
Output -
R =
0.3080 0.8227 0.4248
Q =
0.8227 0.3080 0.4248 0.4248 0.8227 0.3080 0.3080 0.4248 0.8227
In your case, N = 1000.
Approach #2
If you don't care about how many times an element from R is repeated in Q, then you may use this -
R = randn(1,N,'distributed');
Q = R(reshape(ceil(N*rand(N)),1,[]));
Sorry for the vague question, but I hope for an experienced Haskeller this is a no-brainer.
I have to represent and manipulate symmetric matrices, so there are basically three different choices for the data type:
Complete matrix storing both the (i,j) and (j,i) element, although m(i,j) = m(j,i)
Data.Array (Int, Int) Int
A map, storing only elements (i,j) with i <= j (upper triangular matrix)
Data.Map (Int, Int) Int
A vector indexed by k, storing the upper triangular matrix given some vector order f(i,j) = k
Data.Array Int Int
Many operations are going to be necessary on the matrices, updating a single element, querying for rows and columns etc. However, they will mainly act as containers, no linear algebra operations (inversion, det, etc) will be required.
Which one of the options would be the fastest one in general if the dimensionality of the matrices is going to be at around 20x20? When I understand correctly, every update (with (//) in the case of array) requires full copies, so going from 20x20=400 elements to 20*21/2 = 210 elements in the cases 2. or 3. would make a lot of sense, but access is slower for case 2. and 3. needs conversion at some point.
Are there any guidelines?
Btw: The 3rd option is not a really good one, as computing f^-1 requires square roots.
You could try using Data.Array using a specialized Ix class that only generates the upper half of the matrix:
newtype Symmetric = Symmetric { pair :: (Int, Int) } deriving (Ord, Eq)
instance Ix Symmetric where
range ((Symmetric (x1,y1)), (Symmetric (x2,y2))) =
map Symmetric [(x,y) | x <- range (x1,x2), y <- range (y1,y2), x >= y]
inRange (lo,hi) i = x <= hix && x >= lox && y <= hiy && y >= loy && x >= y
where
(lox,loy) = pair lo
(hix,hiy) = pair hi
(x,y) = pair i
index (lo,hi) i
| inRange (lo,hi) i = (x-loy)+(sum$take(y-loy)[hix-lox, hix-lox-1..])
| otherwise = error "Error in array index"
where
(lox,loy) = pair lo
(hix,hiy) = pair hi
(x,y) = pair i
sym x y
| x < y = Symmetric (y,x)
| otherwise = Symmetric (x,y)
*Main Data.Ix> let a = listArray (sym 0 0, sym 6 6) [0..]
*Main Data.Ix> a ! sym 3 2
14
*Main Data.Ix> a ! sym 2 3
14
*Main Data.Ix> a ! sym 2 2
13
*Main Data.Ix> length $ elems a
28
*Main Data.Ix> let b = listArray (sym 0 0, sym 19 19) [0..]
*Main Data.Ix> length $ elems b
210
There is a fourth option: use an array of decreasingly-large arrays. I would go with either option 1 (using a full array and just storing every element twice) or this last one. If you intend to be updating a lot of elements, I strongly recommend using a mutable array; IOArray and STArray are popular choices.
Unless this is for homework or something, you should also take a peek at Hackage. A quick look suggests the problem of manipulating matrices has been solved several times already.