Reduce time /space complexity of simple loop - python

So basically if i have an iteration like this in python
Ive editted the question to include my full code
class Solution:
def myPow(self, x: float, n: int) -> float:
temp = [];
span = range(1,abs(n))
if n ==0:
return 1
if abs(n)==1:
temp.append(x)
else:
for y in span:
if y == 1:
temp = []
temp.append(x*x)
else:
temp.append(temp[-1] * x)
if(n < 0):
return 1/temp[-1]
else:
return temp[-1]
The problem link is : Pow(x,n)-leetcode
How can I modify this to conserve memory and time. Is there another data structure i can use. Im just learning python....
------------EDIT------------
ive modified the code to use a variable instead of a list for the temp data
class Solution:
def myPow(self, x: float, n: int) -> float:
span = range(1,abs(n))
if n ==0:
return 1
if abs(n)==1:
temp = x
else:
for y in span:
if y == 1:
temp = x*x
else:
temp = temp * x
if(n < 0):
return 1/temp
else:
return temp
I still have a problem with my time complexity.
Its working for many testcases, however when it trys to run with x = 0.00001 and n = 2147483647. The time limit issue arises

To reduce the time complexity you can divide the work each time by taking x to the power of 2 and dividing the exponent by two. This makes a logarithmic time algorithm since the exponent is halved at each step.
Consider the following examples:
10^8 = 10^(2*4) = (10^2)^4 = (10*10)^4
Now, there is one edge case. When the exponent is an odd number you can't integer divide it by 2. So in that case you need to multiply the results by the base one additional time.
The following is a direct recursive implementation of the above idea:
class Solution:
def myPow(self, x: float, n: int) -> float:
sign = -1 if n < 0 else 1
n = abs(n)
def helper(x, n):
if n == 1: return x
if n == 0: return 1
if n % 2 == 1:
return helper(x*x, n // 2) * x
else:
return helper(x*x, n // 2)
res = helper(x, n)
if sign == -1:
return 1/res
else:
return res
Note that we have taken abs of the exponent and stored the sign and deal with it at the end.

Instead of iterating from 1 to n, use divide-and-conquer: divide the exponent by 2 and use recursion to get that power, and then square that result. If n was odd, multiply one time more with x:
class Solution:
def myPow(self, x: float, n: int) -> float:
if n == 0:
return 1
if n == 1:
return x
if n < 0:
return self.myPow(1/x, -n)
temp = self.myPow(x, n // 2)
temp *= temp
if n % 2:
temp *= x
return temp

A simple naive solution might be:
def myPow(x: float, n: int) -> float:
## -----------------------
## if we have a negative n then invert x and take the absolute value of n
## -----------------------
if n < 0:
x = 1/x
n = -n
## -----------------------
retval = 1
for _ in range(n):
retval *= x
return retval
While this technically works, you will wait until the cows come home to get a result for:
x = 0.00001 and n = 2147483647
So we need to find a shortcut. Lets' consider 2^5. Our naïve method would calculate that as:
(((2 * 2) * 2) * 2) * 2 == 32
However, what might we observe about the problem if we group some stuff together in a different way:
(2 * 2) * (2 * 2) * 2 == 32
similarly:
((2 * 2) * (2 * 2) * 2) * ((2 * 2) * (2 * 2) * 2) == 32 * 32 = 1024
We might observe that we only technically need to calculate
(2 * 2) * (2 * 2) * 2 == 32
once and use it twice to get 2^10.
Similarly we only need to calcuate:
2 * 2 = 4
once and use it twice to get 2^5....
This suggests a recursion to me.
Let's modify our first attempt to use this divide and concur method.
def myPow2(x: float, n: int) -> float:
## -----------------------
## if we have a negative n then invert x and take the absolute value of n
## -----------------------
if n < 0:
x = 1/x
n = -n
## -----------------------
## -----------------------
## We only need to calculate approximately half the work and use it twice
## at any step.
## -----------------------
def _recurse(x, n):
if n == 0:
return 1
res = _recurse(x, n//2) # calculate it once
res = res * res # use it twice
return res * x if n % 2 else res # if n is odd, multiple by x one more time (see 2^5 above)
## -----------------------
return _recurse(x, n)
Now let's try:
print(myPow2(2.0, 0))
print(myPow2(2.0, 1))
print(myPow2(2.0, 5))
print(myPow2(2.1, 3))
print(myPow2(2.0, -2))
print(myPow2(0.00001, 2147483647))
That gives me:
1
2.0
32.0
9.261000000000001
0.25
0.0

If you have to loop, you have to lope and there is nothing that can be done. Loops in python are slow. That said you may not have to loop and if you do have to loop, it may be possible to push this loop to a highly optimised internal function. Tell us what you are trying to do (not how you think you have to do it, appending elements to a lis may or may not be needed). Always recall the two rules of program optimisation General Rule: Don't do it. Rule for experts: Don't do it yet. Make it work before you make it fast, who knows, it may be fast enough.

Related

How do I go about coding a custom function (without using any library) which returns the log of a number in python [duplicate]

I need to generate the result of the log.
I know that:
Then I made my code:
def log(x, base):
log_b = 2
while x != int(round(base ** log_b)):
log_b += 0.01
print(log_b)
return int(round(log_b))
But it works very slowly. Can I use other method?
One other thing you might want to consider is using the Taylor series of the natural logarithm:
Once you've approximated the natural log using a number of terms from this series, it is easy to change base:
EDIT: Here's another useful identity:
Using this, we could write something along the lines of
def ln(x):
n = 1000.0
return n * ((x ** (1/n)) - 1)
Testing it out, we have:
print ln(math.e), math.log(math.e)
print ln(0.5), math.log(0.5)
print ln(100.0), math.log(100.0)
Output:
1.00050016671 1.0
-0.692907009547 -0.69314718056
4.6157902784 4.60517018599
This shows our value compared to the math.log value (separated by a space) and, as you can see, we're pretty accurate. You'll probably start to lose some accuracy as you get very large (e.g. ln(10000) will be about 0.4 greater than it should), but you can always increase n if you need to.
I used recursion:
def myLog(x, b):
if x < b:
return 0
return 1 + myLog(x/b, b)
You can use binary search for that.
You can get more information on binary search on Wikipedia:
Binary search;
Doubling search.
# search for log_base(x) in range [mn, mx] using binary search
def log_in_range(x, base, mn, mx):
if (mn <= mx):
med = (mn + mx) / 2.0
y = base ** med
if abs(y - x) < 0.00001: # or if math.isclose(x, y): https://docs.python.org/3/library/math.html#math.isclose
return med
elif x > y:
return log_in_range(x, base, med, mx)
elif x < y:
return log_in_range(x, base, mn, med)
return 0
# determine range using doubling search, then call log_in_range
def log(x, base):
if base <= 0 or base == 1 or x <= 0:
raise ValueError('math domain error')
elif 0 < base < 1:
return -log(x, 1/base)
elif 1 <= x and 1 < base:
mx = 1
y = base
while y < x:
y *= y
mx *= 2
return log_in_range(x, base, 0, mx)
elif 0 <= x < 1 and 1 < base:
mn = -1
y = 1/base
while y > x:
y = y ** 0.5
mn *= 2
return log_in_range(x, base, mn, 0)
import math
try :
number_and_base = input() ##input the values for number and base
##assigning those values for the variables
number = int(number_and_base.split()[0])
base = int(number_and_base.split()[1])
##exception handling
except ValueError :
print ("Invalid input...!")
##program
else:
n1 = 1 ##taking an initial value to iterate
while(number >= int(round(base**(n1),0))) : ##finding the most similer value to the number given, varying the vlaue of the power
n1 += 0.000001 ##increasing the initial value bit by bit
n2 = n1-0.0001
if abs(number-base**(n2)) < abs(base**(n1)-number) :
n = n2
else :
n = n1
print(math.floor(n)) ##output value
Comparison:-
This is how your log works:-
def your_log(x, base):
log_b = 2
while x != int(round(base ** log_b)):
log_b += 0.01
#print log_b
return int(round(log_b))
print your_log(16, 2)
# %timeit %run your_log.py
# 1000 loops, best of 3: 579 us per loop
This is my proposed improvement:-
def my_log(x, base):
count = -1
while x > 0:
x /= base
count += 1
if x == 0:
return count
print my_log(16, 2)
# %timeit %run my_log.py
# 1000 loops, best of 3: 321 us per loop
which is faster, using the %timeit magic function in iPython to time the execution for comparison.
It will be long process since it goes in a loop. Therefore,
def log(x,base):
result = ln(x)/ln(base)
return result
def ln(x):
val = x
return 99999999*(x**(1/99999999)-1)
log(8,3)
Values are nearly equal but not exact.

How to write 2**n - 1 as a recursive function?

I need a function that takes n and returns 2n - 1 . It sounds simple enough, but the function has to be recursive. So far I have just 2n:
def required_steps(n):
if n == 0:
return 1
return 2 * req_steps(n-1)
The exercise states: "You can assume that the parameter n is always a positive integer and greater than 0"
2**n -1 is also 1+2+4+...+2n-1 which can made into a single recursive function (without the second one to subtract 1 from the power of 2).
Hint: 1+2*(1+2*(...))
Solution below, don't look if you want to try the hint first.
This works if n is guaranteed to be greater than zero (as was actually promised in the problem statement):
def required_steps(n):
if n == 1: # changed because we need one less going down
return 1
return 1 + 2 * required_steps(n-1)
A more robust version would handle zero and negative values too:
def required_steps(n):
if n < 0:
raise ValueError("n must be non-negative")
if n == 0:
return 0
return 1 + 2 * required_steps(n-1)
(Adding a check for non-integers is left as an exercise.)
To solve a problem with a recursive approach you would have to find out how you can define the function with a given input in terms of the same function with a different input. In this case, since f(n) = 2 * f(n - 1) + 1, you can do:
def required_steps(n):
return n and 2 * required_steps(n - 1) + 1
so that:
for i in range(5):
print(required_steps(i))
outputs:
0
1
3
7
15
You can extract the really recursive part to another function
def f(n):
return required_steps(n) - 1
Or you can set a flag and define just when to subtract
def required_steps(n, sub=True):
if n == 0: return 1
return 2 * required_steps(n-1, False) - sub
>>> print(required_steps(10))
1023
Using an additional parameter for the result, r -
def required_steps (n = 0, r = 1):
if n == 0:
return r - 1
else:
return required_steps(n - 1, r * 2)
for x in range(6):
print(f"f({x}) = {required_steps(x)}")
# f(0) = 0
# f(1) = 1
# f(2) = 3
# f(3) = 7
# f(4) = 15
# f(5) = 31
You can also write it using bitwise left shift, << -
def required_steps (n = 0, r = 1):
if n == 0:
return r - 1
else:
return required_steps(n - 1, r << 1)
The output is the same
Have a placeholder to remember original value of n and then for the very first step i.e. n == N, return 2^n-1
n = 10
# constant to hold initial value of n
N = n
def required_steps(n, N):
if n == 0:
return 1
elif n == N:
return 2 * required_steps(n-1, N) - 1
return 2 * required_steps(n-1, N)
required_steps(n, N)
One way to get the offset of "-1" is to apply it in the return from the first function call using an argument with a default value, then explicitly set the offset argument to zero during the recursive calls.
def required_steps(n, offset = -1):
if n == 0:
return 1
return offset + 2 * required_steps(n-1,0)
On top of all the awesome answers given earlier, below will show its implementation with inner functions.
def outer(n):
k=n
def p(n):
if n==1:
return 2
if n==k:
return 2*p(n-1)-1
return 2*p(n-1)
return p(n)
n=5
print(outer(n))
Basically, it is assigning a global value of n to k and recursing through it with appropriate comparisons.

How to calculate a power of a number with O(log n) complexity?

This is for pure understanding, but i have a code thats basically O(n), but i am unable to decipher how to change it to Olog(n), and each time i use recursion i get nlog(n) complexity.
def power(n,p):
val = []
for i in range(p):
val.append(n)
res = val
n = 1
for x in res:
n*= x
return n
print(power(2,8)) # returns 256
what i need is a code that does the exact same as this above code but it basically does it in Olog(n) as opposed to O(n)
This code implements power() with complexity O(log(n)):
def power(x, y):
if y==0:
return 1
n = power(x, y // 2)
if y % 2 == 0:
return n * n
else:
if y > 0:
return x * n * n
else:
return (n * n) / x

Efficiently generating Stern's Diatomic Sequence

Stern's Diatomic Sequence can be read about in more details over here; however, for my purpose I will define it now.
Definition of Stern's Diatomic Sequence
Let n be a number to generate the fusc function out of. Denoted fusc(n).
If n is 0 then the returned value is 0.
If n is 1 then the returned value is 1.
If n is even then the returned value is fusc(n / 2).
If n is odd then the returned value is fusc((n - 1) / 2) + fusc((n + 1) / 2).
Currently, my Python code brute forces through most of the generation, other than the dividing by two part since it will always yield no change.
def fusc (n):
if n <= 1:
return n
while n > 2 and n % 2 == 0:
n /= 2
return fusc((n - 1) / 2) + fusc((n + 1) / 2)
However, my code must be able to handle digits in the magnitude of 1000s millions of bits, and recursively running through the function thousands millions of times does not seem very efficient or practical.
Is there any way I could algorithmically improve my code such that massive numbers can be passed through without having to recursively call the function so many times?
With memoization for a million bits, the recursion stack would be extremely large. We can first try to look at a sufficiently large number which we can work by hand, fusc(71) in this case:
fusc(71) = fusc(35) + fusc(36)
fusc(35) = fusc(17) + fusc(18)
fusc(36) = fusc(18)
fusc(71) = 1 * fusc(17) + 2 * fusc(18)
fusc(17) = fusc(8) + fusc(9)
fusc(18) = fusc(9)
fusc(71) = 1 * fusc(8) + 3 * fusc(9)
fusc(8) = fusc(4)
fusc(9) = fusc(4) + fusc(5)
fusc(71) = 4 * fusc(4) + 3 * fusc(5)
fusc(4) = fusc(2)
fusc(3) = fusc(1) + fusc(2)
fusc(71) = 7 * fusc(2) + 3 * fusc(3)
fusc(2) = fusc(1)
fusc(3) = fusc(1) + fusc(2)
fusc(71) = 11 * fusc(1) + 3 * fusc(2)
fusc(2) = fusc(1)
fusc(71) = 14 * fusc(1) = 14
We realize that we can avoid recursion completely in this case as we can always express fusc(n) in the form a * fusc(m) + b * fusc(m+1) while reducing the value of m to 0. From the example above, you may find the following pattern:
if m is odd:
a * fusc(m) + b * fusc(m+1) = a * fusc((m-1)/2) + (b+a) * fusc((m+1)/2)
if m is even:
a * fusc(m) + b * fusc(m+1) = (a+b) * fusc(m/2) + b * fusc((m/2)+1)
Therefore, you may use a simple loop function to solve the problem in O(lg(n)) time
def fusc(n):
if n == 0: return 0
a = 1
b = 0
while n > 0:
if n%2:
b = b + a
n = (n-1)/2
else:
a = a + b
n = n/2
return b
lru_cache works wonders in your case. make sure maxsize is a power of 2. may need to fiddle a bit with that size for your application. cache_info() will help with that.
also use // instead of / for integer division.
from functools import lru_cache
#lru_cache(maxsize=512, typed=False)
def fusc(n):
if n <= 1:
return n
while n > 2 and n % 2 == 0:
n //= 2
return fusc((n - 1) // 2) + fusc((n + 1) // 2)
print(fusc(1000000000078093254329870980000043298))
print(fusc.cache_info())
and yes, this is just meomization as proposed by Filip Malczak.
you might gain an additional tiny speedup using bit-operations in the while loop:
while not n & 1: # as long as the lowest bit is not 1
n >>= 1 # shift n right by one
UPDATE:
here is a simple way of doing meomzation 'by hand':
def fusc(n, _mem={}): # _mem will be the cache of the values
# that have been calculated before
if n in _mem: # if we know that one: just return the value
return _mem[n]
if n <= 1:
return n
while not n & 1:
n >>= 1
if n == 1:
return 1
ret = fusc((n - 1) // 2) + fusc((n + 1) // 2)
_mem[n] = ret # store the value for next time
return ret
UPDATE
after reading a short article by dijkstra himself a minor update.
the article states, that f(n) = f(m) if the fist and last bit of m are the same as those of n and the bits in between are inverted. the idea is to get n as small as possible.
that is what the bitmask (1<<n.bit_length()-1)-2 is for (first and last bits are 0; those in the middle 1; xoring n with that gives m as described above).
i was only able to do small benchmarks; i'm interested if this is any help at all for the magitude of your input... this will reduce the memory for the cache and hopefully bring some speedup.
def fusc_ed(n, _mem={}):
if n <= 1:
return n
while not n & 1:
n >>= 1
if n == 1:
return 1
# https://www.cs.utexas.edu/users/EWD/transcriptions/EWD05xx/EWD578.html
# bit invert the middle bits and check if this is smaller than n
m = n ^ (1<<n.bit_length()-1)-2
n = m if m < n else n
if n in _mem:
return _mem[n]
ret = fusc(n >> 1) + fusc((n >> 1) + 1)
_mem[n] = ret
return ret
i had to increase the recursion limit:
import sys
sys.setrecursionlimit(10000) # default limit was 1000
benchmarking gave strange results; using the code below and making sure that i always started a fresh interperter (having an empty _mem) i sometimes got significantly better runtimes; on other occasions the new code was slower...
benchmarking code:
print(n.bit_length())
ti = timeit('fusc(n)', setup='from __main__ import fusc, n', number=1)
print(ti)
ti = timeit('fusc_ed(n)', setup='from __main__ import fusc_ed, n', number=1)
print(ti)
and these are three random results i got:
6959
24.117448464001427
0.013900151001507766
6989
23.92404893300045
0.013844672999766772
7038
24.33894686200074
24.685758719999285
that is where i stopped...

How to compute the nth root of a very big integer

I need a way to compute the nth root of a long integer in Python.
I tried pow(m, 1.0/n), but it doesn't work:
OverflowError: long int too large to convert to float
Any ideas?
By long integer I mean REALLY long integers like:
11968003966030964356885611480383408833172346450467339251
196093144141045683463085291115677488411620264826942334897996389
485046262847265769280883237649461122479734279424416861834396522
819159219215308460065265520143082728303864638821979329804885526
557893649662037092457130509980883789368448042961108430809620626
059287437887495827369474189818588006905358793385574832590121472
680866521970802708379837148646191567765584039175249171110593159
305029014037881475265618958103073425958633163441030267478942720
703134493880117805010891574606323700178176718412858948243785754
898788359757528163558061136758276299059029113119763557411729353
915848889261125855717014320045292143759177464380434854573300054
940683350937992500211758727939459249163046465047204851616590276
724564411037216844005877918224201569391107769029955591465502737
961776799311859881060956465198859727495735498887960494256488224
613682478900505821893815926193600121890632
If it's a REALLY big number. You could use a binary search.
def find_invpow(x,n):
"""Finds the integer component of the n'th root of x,
an integer such that y ** n <= x < (y + 1) ** n.
"""
high = 1
while high ** n <= x:
high *= 2
low = high/2
while low < high:
mid = (low + high) // 2
if low < mid and mid**n < x:
low = mid
elif high > mid and mid**n > x:
high = mid
else:
return mid
return mid + 1
For example:
>>> x = 237734537465873465
>>> n = 5
>>> y = find_invpow(x,n)
>>> y
2986
>>> y**n <= x <= (y+1)**n
True
>>>
>>> x = 119680039660309643568856114803834088331723464504673392511960931441>
>>> n = 45
>>> y = find_invpow(x,n)
>>> y
227661383982863143360L
>>> y**n <= x < (y+1)**n
True
>>> find_invpow(y**n,n) == y
True
>>>
Gmpy is a C-coded Python extension module that wraps the GMP library to provide to Python code fast multiprecision arithmetic (integer, rational, and float), random number generation, advanced number-theoretical functions, and more.
Includes a root function:
x.root(n): returns a 2-element tuple (y,m), such that y is the
(possibly truncated) n-th root of x; m, an ordinary Python int,
is 1 if the root is exact (x==y**n), else 0. n must be an ordinary
Python int, >=0.
For example, 20th root:
>>> import gmpy
>>> i0=11968003966030964356885611480383408833172346450467339251
>>> m0=gmpy.mpz(i0)
>>> m0
mpz(11968003966030964356885611480383408833172346450467339251L)
>>> m0.root(20)
(mpz(567), 0)
You can make it run slightly faster by avoiding the while loops in favor of setting low to 10 ** (len(str(x)) / n) and high to low * 10. Probably better is to replace the len(str(x)) with the bitwise length and using a bit shift. Based on my tests, I estimate a 5% speedup from the first and a 25% speedup from the second. If the ints are big enough, this might matter (and the speedups may vary). Don't trust my code without testing it carefully. I did some basic testing but may have missed an edge case. Also, these speedups vary with the number chosen.
If the actual data you're using is much bigger than what you posted here, this change may be worthwhile.
from timeit import Timer
def find_invpow(x,n):
"""Finds the integer component of the n'th root of x,
an integer such that y ** n <= x < (y + 1) ** n.
"""
high = 1
while high ** n < x:
high *= 2
low = high/2
while low < high:
mid = (low + high) // 2
if low < mid and mid**n < x:
low = mid
elif high > mid and mid**n > x:
high = mid
else:
return mid
return mid + 1
def find_invpowAlt(x,n):
"""Finds the integer component of the n'th root of x,
an integer such that y ** n <= x < (y + 1) ** n.
"""
low = 10 ** (len(str(x)) / n)
high = low * 10
while low < high:
mid = (low + high) // 2
if low < mid and mid**n < x:
low = mid
elif high > mid and mid**n > x:
high = mid
else:
return mid
return mid + 1
x = 237734537465873465
n = 5
tests = 10000
print "Norm", Timer('find_invpow(x,n)', 'from __main__ import find_invpow, x,n').timeit(number=tests)
print "Alt", Timer('find_invpowAlt(x,n)', 'from __main__ import find_invpowAlt, x,n').timeit(number=tests)
Norm 0.626754999161
Alt 0.566340923309
If you are looking for something standard, fast to write with high precision. I would use decimal and adjust the precision (getcontext().prec) to at least the length of x.
Code (Python 3.0)
from decimal import *
x = '11968003966030964356885611480383408833172346450467339251\
196093144141045683463085291115677488411620264826942334897996389\
485046262847265769280883237649461122479734279424416861834396522\
819159219215308460065265520143082728303864638821979329804885526\
557893649662037092457130509980883789368448042961108430809620626\
059287437887495827369474189818588006905358793385574832590121472\
680866521970802708379837148646191567765584039175249171110593159\
305029014037881475265618958103073425958633163441030267478942720\
703134493880117805010891574606323700178176718412858948243785754\
898788359757528163558061136758276299059029113119763557411729353\
915848889261125855717014320045292143759177464380434854573300054\
940683350937992500211758727939459249163046465047204851616590276\
724564411037216844005877918224201569391107769029955591465502737\
961776799311859881060956465198859727495735498887960494256488224\
613682478900505821893815926193600121890632'
minprec = 27
if len(x) > minprec: getcontext().prec = len(x)
else: getcontext().prec = minprec
x = Decimal(x)
power = Decimal(1)/Decimal(3)
answer = x**power
ranswer = answer.quantize(Decimal('1.'), rounding=ROUND_UP)
diff = x - ranswer**Decimal(3)
if diff == Decimal(0):
print("x is the cubic number of", ranswer)
else:
print("x has a cubic root of ", answer)
Answer
x is the cubic number of 22873918786185635329056863961725521583023133411
451452349318109627653540670761962215971994403670045614485973722724603798
107719978813658857014190047742680490088532895666963698551709978502745901
704433723567548799463129652706705873694274209728785041817619032774248488
2965377218610139128882473918261696612098418
Oh, for numbers that big, you would use the decimal module.
ns: your number as a string
ns = "11968003966030964356885611480383408833172346450467339251196093144141045683463085291115677488411620264826942334897996389485046262847265769280883237649461122479734279424416861834396522819159219215308460065265520143082728303864638821979329804885526557893649662037092457130509980883789368448042961108430809620626059287437887495827369474189818588006905358793385574832590121472680866521970802708379837148646191567765584039175249171110593159305029014037881475265618958103073425958633163441030267478942720703134493880117805010891574606323700178176718412858948243785754898788359757528163558061136758276299059029113119763557411729353915848889261125855717014320045292143759177464380434854573300054940683350937992500211758727939459249163046465047204851616590276724564411037216844005877918224201569391107769029955591465502737961776799311859881060956465198859727495735498887960494256488224613682478900505821893815926193600121890632"
from decimal import Decimal
d = Decimal(ns)
one_third = Decimal("0.3333333333333333")
print d ** one_third
and the answer is: 2.287391878618402702753613056E+305
TZ pointed out that this isn't accurate... and he's right. Here's my test.
from decimal import Decimal
def nth_root(num_decimal, n_integer):
exponent = Decimal("1.0") / Decimal(n_integer)
return num_decimal ** exponent
def test():
ns = "11968003966030964356885611480383408833172346450467339251196093144141045683463085291115677488411620264826942334897996389485046262847265769280883237649461122479734279424416861834396522819159219215308460065265520143082728303864638821979329804885526557893649662037092457130509980883789368448042961108430809620626059287437887495827369474189818588006905358793385574832590121472680866521970802708379837148646191567765584039175249171110593159305029014037881475265618958103073425958633163441030267478942720703134493880117805010891574606323700178176718412858948243785754898788359757528163558061136758276299059029113119763557411729353915848889261125855717014320045292143759177464380434854573300054940683350937992500211758727939459249163046465047204851616590276724564411037216844005877918224201569391107769029955591465502737961776799311859881060956465198859727495735498887960494256488224613682478900505821893815926193600121890632"
nd = Decimal(ns)
cube_root = nth_root(nd, 3)
print (cube_root ** Decimal("3.0")) - nd
if __name__ == "__main__":
test()
It's off by about 10**891
Possibly for your curiosity:
http://en.wikipedia.org/wiki/Hensel_Lifting
This could be the technique that Maple would use to actually find the nth root of large numbers.
Pose the fact that x^n - 11968003.... = 0 mod p, and go from there...
I may suggest four methods for solving your task. First is based on Binary Search. Second is based on Newton's Method. Third is based on Shifting n-th Root Algorithm. Fourth is called by me Chord-Tangent method described by me in picture here.
Binary Search was already implemented in many answers above. I just introduce here my own vision of it and its implementation.
As alternative I also implement Optimized Binary Search method (marked Opt). This method just starts from range [hi / 2, hi) where hi is equal to 2^(num_bit_length / k) if we're computing k-th root.
Newton's Method is new here, as I see it wasn't implemented in other answers. It is usually considered to be faster than Binary Search, although my own timings in code below don't show any speedup. Hence this method here is just for reference/interest.
Shifting Method is 30-50% faster than optimized binary search method, and should be even faster if implemented in C++, because C++ has fast 64 bit arithemtics which is partially used in this method.
Chord-Tangent Method:
Chord-Tangent Method is invented by me on piece of paper (see image above), it is inspired and is an improvement of Newton method. Basically I draw a Chord and a Tangent Line and find intersection with horizontal line y = n, these two intersections form lower and upper bound approximations of location of root solution (x0, n) where n = x0 ^ k. This method appeared to be fastest of all, while all other methods do more than 2000 iterations, this method does just 8 iterations, for the case of 8192-bit numbers. So this method is 200-300x times faster than previous (by speed) Shifting Method.
As an example I generate really huge random integer of 8192 bits in size. And measure timings of finding cubic root with both methods.
In test() function you can see that I passed k = 3 as root's power (cubic root), you can pass any power instead of 3.
Try it online!
def binary_search(begin, end, f, *, niter = [0]):
while begin < end:
niter[0] += 1
mid = (begin + end) >> 1
if f(mid):
begin = mid + 1
else:
end = mid
return begin
def binary_search_kth_root(n, k, *, verbose = False):
# https://en.wikipedia.org/wiki/Binary_search_algorithm
niter = [0]
res = binary_search(0, n + 1, lambda root: root ** k < n, niter = niter)
if verbose:
print('Binary Search iterations:', niter[0])
return res
def binary_search_opt_kth_root(n, k, *, verbose = False):
# https://en.wikipedia.org/wiki/Binary_search_algorithm
niter = [0]
hi = 1 << (n.bit_length() // k - 1)
while hi ** k <= n:
niter[0] += 1
hi <<= 1
res = binary_search(hi >> 1, hi, lambda root: root ** k < n, niter = niter)
if verbose:
print('Binary Search Opt iterations:', niter[0])
return res
def newton_kth_root(n, k, *, verbose = False):
# https://en.wikipedia.org/wiki/Newton%27s_method
f = lambda x: x ** k - n
df = lambda x: k * x ** (k - 1)
x, px, niter = n, 2 * n, [0]
while abs(px - x) > 1:
niter[0] += 1
px = x
x -= f(x) // df(x)
if verbose:
print('Newton Method iterations:', niter[0])
mini, minv = None, None
for i in range(-2, 3):
v = abs(f(x + i))
if minv is None or v < minv:
mini, minv = i, v
return x + mini
def shifting_kth_root(n, k, *, verbose = False):
# https://en.wikipedia.org/wiki/Shifting_nth_root_algorithm
B_bits = 64
r, y = 0, 0
B = 1 << B_bits
Bk_bits = B_bits * k
Bk_mask = (1 << Bk_bits) - 1
niter = [0]
for i in range((n.bit_length() + Bk_bits - 1) // Bk_bits - 1, -1, -1):
alpha = (n >> (i * Bk_bits)) & Bk_mask
B_y = y << B_bits
Bk_yk = (y ** k) << Bk_bits
Bk_r_alpha = (r << Bk_bits) + alpha
Bk_yk_Bk_r_alpha = Bk_yk + Bk_r_alpha
beta = binary_search(1, B, lambda beta: (B_y + beta) ** k <= Bk_yk_Bk_r_alpha, niter = niter) - 1
y, r = B_y + beta, Bk_r_alpha - ((B_y + beta) ** k - Bk_yk)
if verbose:
print('Shifting Method iterations:', niter[0])
return y
def chord_tangent_kth_root(n, k, *, verbose = False):
niter = [0]
hi = 1 << (n.bit_length() // k - 1)
while hi ** k <= n:
niter[0] += 1
hi <<= 1
f = lambda x: x ** k
df = lambda x: k * x ** (k - 1)
# https://i.stack.imgur.com/et9O0.jpg
x_begin, x_end = hi >> 1, hi
y_begin, y_end = f(x_begin), f(x_end)
for icycle in range(1 << 30):
if x_end - x_begin <= 1:
break
niter[0] += 1
if 0: # Do Binary Search step if needed
x_mid = (x_begin + x_end) >> 1
y_mid = f(x_mid)
if y_mid > n:
x_end, y_end = x_mid, y_mid
else:
x_begin, y_begin = x_mid, y_mid
# (y_end - y_begin) / (x_end - x_begin) = (n - y_begin) / (x_n - x_begin) ->
x_n = x_begin + (n - y_begin) * (x_end - x_begin) // (y_end - y_begin)
y_n = f(x_n)
tangent_x = x_n + (n - y_n) // df(x_n) + 1
chord_x = x_n + (n - y_n) * (x_end - x_n) // (y_end - y_n)
assert chord_x <= tangent_x, (chord_x, tangent_x)
x_begin, x_end = chord_x, tangent_x
y_begin, y_end = f(x_begin), f(x_end)
assert y_begin <= n, (chord_x, y_begin, n, n - y_begin)
assert y_end > n, (icycle, tangent_x - binary_search_kth_root(n, k), y_end, n, y_end - n)
if verbose:
print('Chord Tangent Method iterations:', niter[0])
return x_begin
def test():
import random, timeit
nruns = 3
bits = 8192
n = random.randrange(1 << (bits - 1), 1 << bits)
a = binary_search_kth_root(n, 3, verbose = True)
b = binary_search_opt_kth_root(n, 3, verbose = True)
c = newton_kth_root(n, 3, verbose = True)
d = shifting_kth_root(n, 3, verbose = True)
e = chord_tangent_kth_root(n, 3, verbose = True)
assert abs(a - b) <= 0 and abs(a - c) <= 1 and abs(a - d) <= 1 and abs(a - e) <= 1, (a - b, a - c, a - d, a - e)
print()
print('Binary Search timing:', round(timeit.timeit(lambda: binary_search_kth_root(n, 3), number = nruns) / nruns, 3), 'sec')
print('Binary Search Opt timing:', round(timeit.timeit(lambda: binary_search_opt_kth_root(n, 3), number = nruns) / nruns, 3), 'sec')
print('Newton Method timing:', round(timeit.timeit(lambda: newton_kth_root(n, 3), number = nruns) / nruns, 3), 'sec')
print('Shifting Method timing:', round(timeit.timeit(lambda: shifting_kth_root(n, 3), number = nruns) / nruns, 3), 'sec')
print('Chord Tangent Method timing:', round(timeit.timeit(lambda: chord_tangent_kth_root(n, 3), number = nruns) / nruns, 3), 'sec')
if __name__ == '__main__':
test()
Output:
Binary Search iterations: 8192
Binary Search Opt iterations: 2732
Newton Method iterations: 9348
Shifting Method iterations: 2752
Chord Tangent Method iterations: 8
Binary Search timing: 0.506 sec
Binary Search Opt timing: 0.05 sec
Newton Method timing: 2.09 sec
Shifting Method timing: 0.03 sec
Chord Tangent Method timing: 0.001 sec
I came up with my own answer, which takes #Mahmoud Kassem's idea, simplifies the code, and makes it more reusable:
def cube_root(x):
return decimal.Decimal(x) ** (decimal.Decimal(1) / decimal.Decimal(3))
I tested it in Python 3.5.1 and Python 2.7.8, and it seemed to work fine.
The result will have as many digits as specified by the decimal context the function is run in, which by default is 28 decimal places. According to the documentation for the power function in the decimal module, "The result is well-defined but only “almost always correctly-rounded”.". If you need a more accurate result, it can be done as follows:
with decimal.localcontext() as context:
context.prec = 50
print(cube_root(42))
In older versions of Python, 1/3 is equal to 0. In Python 3.0, 1/3 is equal to 0.33333333333 (and 1//3 is equal to 0).
So, either change your code to use 1/3.0 or switch to Python 3.0 .
Try converting the exponent to a floating number, as the default behaviour of / in Python is integer division
n**(1/float(3))
Well, if you're not particularly worried about precision, you could convert it to a sting, chop off some digits, use the exponent function, and then multiply the result by the root of how much you chopped off.
E.g. 32123 is about equal to 32 * 1000, the cubic root is about equak to cubic root of 32 * cubic root of 1000. The latter can be calculated by dividing the number of 0s by 3.
This avoids the need for the use of extension modules.

Categories

Resources