在python中找到变化集的最小值和最大值的有效方法

Question

I need to find the min/max value in a changing large set, in C++, it could be

#include<set>
using namespace std;
int minVal(set<int> & mySet){
    return *mySet.begin();
}
int maxVal(set<int> & mySet){
    return *mySet.rbegin();
}
int main(){
    set <int> mySet;
    for(..;..;..){
       // add or delete element in mySet
       ...
       // print the min and max value in the set
       printf("%d %d\n", minVal(mySet), maxVal(mySet)); 
    }
}

In C++, each query operation is O(1), but in python, I tried to use the build-in method min and max but it's too slow. Each min/max operation takes O(n) time (n is the length of my Set). Are there any elegant and efficient way to do this? Or any datatype support these operation?

mySet=set()
for i in range(..):
  # add or delete element in mySet
  ...
  # print the min and max value in the set
  print(min(mySet),max(mySet))

Answer 1

The efficient implementation in terms of complexity is wrapping a python set (which uses a hash table) and keeping a pair of maxElement and minElement attributes in the object, and updating those accordingly when adding or removing elements. This keeps every query of existence, min and max O(1). The deletion operation though would be O(n) worst case with the simplest implementation (since you have to find the next-to-minimum element if you happen to remove the minimum element, and the same happens with the maximum).

This said, the C++ implementation uses a balanced search tree which has O(log n) existence checks, deletion and insertion operations. You can find an implementation of this type of data structure in the bintrees package.

I wouldn't use just a heapq as suggested in comments as a heap is O(n) for checking existence of elements (main point of a set data structure I guess, which I assume you need).

Answer 2

You could use two priority queues to maintain min and max values in the set, respectively. Unfortunately, the stdlib's heapq doesn't support removing entries from the queue in O(log n) time out of the box. The suggested workaround is to just mark entries as removed and discard them when you pop them from the queue (which might be ok in many scenarios, though). Below is a Python class implementing that approach:

from heapq import heappop, heappush

class MinMaxSet:
    def __init__(self):
        self.min_queue = []
        self.max_queue = []
        self.entries = {}  # mapping of values to entries in the queue

    def __len__(self):
        return len(self.entries)

    def add(self, val):
        if val not in self.entries:
            entry_min = [val, False]
            entry_max = [-val, False]

            heappush(self.min_queue, entry_min)
            heappush(self.max_queue, entry_max)

            self.entries[val] = entry_min, entry_max

    def delete(self, val):
        if val in self.entries:
            entry_min, entry_max = self.entries.pop(val)
            entry_min[-1] = entry_max[-1] = True  # deleted

    def get_min(self):
        while self.min_queue[0][-1]:
            heappop(self.min_queue)
        return self.min_queue[0][0]

    def get_max(self):
        while self.max_queue[0][-1]:
            heappop(self.max_queue)
        return -self.max_queue[0][0]

Demo:

>>> s = MinMaxSet()
>>> for x in [1, 5, 10, 14, 11, 14, 15, 2]:
...     s.add(x)
... 
>>> len(s)
7
>>> print(s.get_min(), s.get_max())
1 15
>>> s.delete(1)
>>> s.delete(15)
>>> print(s.get_min(), s.get_max())
2 14

Answer 3

Since 2020 package bintrees is depricated and should be replaced with sortedcontainers .

Example usage:

import sortedcontainers

s = sortedcontainers.SortedList()
s.add(10)
s.add(3)
s.add(25)
s.add(8)
min = s[0]      # read min value
min = s.pop(0)  # read and remove min value
max = s[-1]     # read max value
max = s.pop()   # read and remove max value

Beside SortedList you also have SortedDict and SortedSet. Here is API documentation .

Answer 4

numpy min max is twice as fast as the native method

import time as t
import numpy as np

def initialize():
    storage.reset()

def tick():

    array = data.btc_usd.period(250, 'close')

    t1 = t.time()

    a = min(array)
    b = max(array)

    t2 = t.time()

    c = np.min(array)
    d = np.max(array)

    t3 = t.time()

    storage.t1 = storage.get('t1', 0)
    storage.t2 = storage.get('t2', 0)
    storage.t1 += t2-t1
    storage.t2 += t3-t2


def stop():

    log('python: %.5f' % storage.t1)
    log('numpy: %.5f' % storage.t2)
    log('ticks: %s' % info.tick)

yeilds:

[2015-11-06 10:00:00] python: 0.45959
[2015-11-06 10:00:00] numpy: 0.26148
[2015-11-06 10:00:00] ticks: 7426

but I think you're looking for something more like this:

import time as t
import numpy as np

def initialize():
    storage.reset()

def tick():

    storage.closes = storage.get('closes', [])
    if info.tick == 0:
        storage.closes = [float(x) for x in data.btc_usd.period(250, 'close')]
    else:
        z = storage.closes.pop(0) #pop left
        price = float(data.btc_usd.close)
        storage.closes.append(price) #append right
    array = np.array(storage.closes)[-250:]

    # now we know 'z' just left the list and 'price' just entered
    # otherwise the array is the same as the previous example

    t1 = t.time()
    # PYTHON METHOD
    a = min(array)
    b = max(array)

    t2 = t.time()
    # NUMPY METHOD
    c = np.min(array)
    d = np.max(array)

    t3 = t.time()
    # STORAGE METHOD
    storage.e = storage.get('e', 0)
    storage.f = storage.get('f', 0)
    if info.tick == 0:
        storage.e = np.min(array)
        storage.f = np.max(array)
    else:
        if z == storage.e:
            storage.e = np.min(array)
        if z == storage.f:
            storage.f = np.max(array)
        if price < storage.e:
            storage.e = price
        if price > storage.f:
            storage.f = price

    t4 = t.time()

    storage.t1 = storage.get('t1', 0)
    storage.t2 = storage.get('t2', 0)
    storage.t3 = storage.get('t3', 0)    
    storage.t1 += t2-t1
    storage.t2 += t3-t2
    storage.t3 += t4-t3


def stop():

    log('python: %.5f'  % storage.t1)
    log('numpy: %.5f'   % storage.t2)
    log('storage: %.5f' % storage.t3)
    log('ticks: %s'     % info.tick)

yeilds:

[2015-11-06 10:00:00] python: 0.45694
[2015-11-06 10:00:00] numpy: 0.23580
[2015-11-06 10:00:00] storage: 0.16870
[2015-11-06 10:00:00] ticks: 7426

which brings us down to about 1/3rd of the native method with a 7500 iterations against a list of 250