# Comparing three data structures for dealing with probability distributions in Java

Introduction

Suppose you are given three elements $a, b, c$ with respective weights $1, 1, 3$. Now, a probability distribution data structures will return upon request $a$ with probability 20%, $b$ with probability 20%, and $c$ with probability 60%.

The API for my probability distribution data structures is defined by the following abstract class:

package net.coderodde.stat;

import java.util.Objects;
import java.util.Random;

/**
* This class implements an abstract base class for probability distributions.
* Elements are added with strictly positive weights and whenever asking this
* data structure for a random element, their respective weights are taken into
* account. For example, if this data structure contains three different
* elements (<tt>a</tt>, <tt>b</tt>, <tt>c</tt> with respective weights
* <tt>1.0</tt>, <tt>1.0</tt>, <tt>3.0</tt>), whenever asking for a random
* element, there is 20 percent chance of obtaining <tt>a</tt>, 20 percent
* chance of obtaining <tt>b</tt>, and 60 percent chance of obtaining
* <tt>c</tt>.
*
* @author Rodion "rodde" Efremov
* @version 1.6 (Jun 11, 2016)
*/
public abstract class AbstractProbabilityDistribution<E> {

/**
* The amount of elements in this probability distribution.
*/
protected int size;

/**
* The sum of all weights.
*/
protected double totalWeight;

/**
* The random number generator of this probability distribution.
*/
protected final Random random;

/**
* Constructs this probability distribution.
*/
protected AbstractProbabilityDistribution() {
this(new Random());
}

/**
* Constructs this probability distribution using the input random number
* generator.
*
* @param random the random number generator.
*/
protected AbstractProbabilityDistribution(final Random random) {
this.random =
Objects.requireNonNull(random,
"The random number generator is null.");
}

public boolean isEmpty() {
return this.size == 0;
}

public int size() {
return this.size;
}

/**
* Adds the element {@code element} to this probability distribution, and
* assigns {@code weight} as its weight.
*
* @param element the element to add.
* @param weight  the weight of the new element.
*
* @return {@code true} only if the input element did not reside in this
*         structure and was successfully added.
*/
public abstract boolean addElement(final E element, final double weight);

/**
* Returns a randomly chosen element from this probability distribution
* taking the weights into account.
*
* @return a randomly chosen element.
*/
public abstract E sampleElement();

/**
* Returns {@code true} if this probability distribution contains the
* element {@code element}.
*
* @param element the element to query.
* @return {@code true} if the input element is in this probability
*         distribution; {@code false} otherwise.
*/
public abstract boolean contains(final E element);

/**
* Removes the element {@code element} from this probability distribution.
*
* @param element the element to remove.
* @return {@code true} if the element was present in this probability
*         distribution and was successfully removed.
*/
public abstract boolean removeElement(final E element);

/**
* Removes all elements from this probability distribution.
*/
public abstract void clear();

/**
* Checks that the element weight is valid. The weight must not be a
* <tt>NaN</tt> and must be positive, but not a positive infinity.
*
* @param weight the weight to validate.
*/
protected void checkWeight(final double weight) {
if (Double.isNaN(weight)) {
throw new IllegalArgumentException("The element weight is NaN.");
}

if (weight <= 0.0) {
throw new IllegalArgumentException(
"The element weight must be positive. Received " + weight);
}

if (Double.isInfinite(weight)) {
// Once here, 'weight' is positive infinity.
throw new IllegalArgumentException(
"The element weight is infinite.");
}
}

/**
* Checks that this probability distribution contains at least one element.
*/
protected void checkNotEmpty() {
if (size == 0) {
throw new IllegalStateException(
"This probability distribution is empty.");
}
}
}


Implementations

The first probability distribution data structure relies on arrays. It has following running times:

• element addition in ammortized constant time,
• element removal in worst-case linear time,
• element sampling in worst-case linear time.

The data structure follows:

ArrayProbabilityDistribution.java:

package net.coderodde.stat.support;

import java.util.HashSet;
import java.util.Objects;
import java.util.Random;
import java.util.Set;
import net.coderodde.stat.AbstractProbabilityDistribution;

/**
* This class implements a probability distribution relying on an array of
* elements. The running times are as follows:
*
* <table>
* <tr><td>Method</td>  <td>Complexity</td></tr>
* <tr><td><tt>addElement   </tt> </td>  <td>amortized constant time,</td></tr>
* <tr><td><tt>sampleElement</tt> </td>  <td><tt>O(n)</tt>,</td></tr>
* <tr><td><tt>removeElement</tt> </td>  <td><tt>O(n)</tt>.</td></tr>
* </table>
*
* @param <E> the actual type of the elements stored in this probability
*            distribution.
*
* @author Rodion "rodde" Efremov
* @version 1.6 (Jun 11, 2016)
*/
public class ArrayProbabilityDistribution<E>
extends AbstractProbabilityDistribution<E> {

private static final int DEFAULT_STORAGE_ARRAYS_CAPACITY = 8;

private Object[] objectStorageArray;
private double[] weightStorageArray;
private final Set<E> filterSet = new HashSet<>();

public ArrayProbabilityDistribution() {
this(new Random());
}

public ArrayProbabilityDistribution(final Random random) {
super(random);
this.objectStorageArray = new Object[DEFAULT_STORAGE_ARRAYS_CAPACITY];
this.weightStorageArray = new double[DEFAULT_STORAGE_ARRAYS_CAPACITY];
}

/**
* {@inheritDoc }
*/
@Override
public boolean addElement(final E element, final double weight) {
checkWeight(weight);

if (filterSet.contains(element)) {
// 'element' is already present in this probability distribution.
return false;
}

ensureCapacity(this.size + 1);
objectStorageArray[this.size] = element;
weightStorageArray[this.size] = weight;
this.totalWeight += weight;
this.size++;
return true;
}

/**
* {@inheritDoc }
*/
@Override
public E sampleElement() {
checkNotEmpty();
double value = this.random.nextDouble() * this.totalWeight;

for (int i = 0; i < this.size; ++i) {
if (value < this.weightStorageArray[i]) {
return (E) this.objectStorageArray[i];
}

value -= this.weightStorageArray[i];
}

throw new IllegalStateException("Should not get here.");
}

/**
* {@inheritDoc }
*/
@Override
public boolean removeElement(final E element) {
if (!this.filterSet.contains(element)) {
return false;
}

final int index = indexOf(element);
this.totalWeight -= this.weightStorageArray[index];

for (int j = index + 1; j < this.size; ++j) {
objectStorageArray[j - 1] = objectStorageArray[j];
weightStorageArray[j - 1] = weightStorageArray[j];
}

objectStorageArray[--this.size] = null;
return true;
}

/**
* {@inheritDoc }
*/
@Override
public void clear() {
for (int i = 0; i < this.size; ++i) {
objectStorageArray[i] = null;
}

this.size = 0;
this.totalWeight = 0.0;
}

/**
* {@inheritDoc }
*/
@Override
public boolean contains(E element) {
return this.filterSet.contains(element);
}

private int indexOf(final E element) {
for (int i = 0; i < this.size; ++i) {
if (Objects.equals(element, this.objectStorageArray[i])) {
return i;
}
}

return -1;
}

private void ensureCapacity(final int requestedCapacity) {
if (requestedCapacity > objectStorageArray.length) {
final int newCapacity = Math.max(requestedCapacity,
2 * objectStorageArray.length);
final Object[] newObjectStorageArray = new Object[newCapacity];
final double[] newWeightStorageArray = new double[newCapacity];

System.arraycopy(this.objectStorageArray,
0,
newObjectStorageArray,
0,
this.size);

System.arraycopy(this.weightStorageArray,
0,
newWeightStorageArray,
0,
this.size);

this.objectStorageArray = newObjectStorageArray;
this.weightStorageArray = newWeightStorageArray;
}
}
}


The second probability distribution data structure relies on a linked list, and provides the following operation:

• element addition in ammortized constant time,
• element removal in constant time,
• element sampling in worst-case linear time.

The data structure follows:

package net.coderodde.stat.support;

import java.util.HashMap;
import java.util.Map;
import java.util.Random;
import net.coderodde.stat.AbstractProbabilityDistribution;

/**
* This class implements a probability distribution relying on a linked list.
* The running times of the main methods are as follows:
*
* <table>
* <tr><td>Method</td>  <td>Complexity</td></tr>
*     <td><tt>amortized constant time</tt>,</td></tr>
* <tr><td><tt>sampleElement</tt> </td>  <td><tt>O(n)</tt>,</td></tr>
* <tr><td><tt>removeElement</tt> </td>  <td><tt>O(1)</tt>.</td></tr>
* </table>
*
* This probability distribution class is best used whenever it is modified
* frequently compared to the number of queries made.
*
* @param <E> the actual type of the elements stored in this probability
*            distribution.
*
* @author Rodion "rodde" Efremov
* @version 1.6 (Jun 11, 2016)
*/
extends AbstractProbabilityDistribution<E> {

private static final class LinkedListNode<E> {

private final E element;
private final double weight;

LinkedListNode(final E element, final double weight) {
this.element = element;
this.weight  = weight;
}

E getElement() {
return this.element;
}

double getWeight() {
return this.weight;
}

return this.prev;
}

return this.next;
}

this.prev = node;
}

this.next = node;
}
}

/**
* This map maps the elements to their respective linked list nodes.
*/
private final Map<E, LinkedListNode<E>> map = new HashMap<>();

/**
* Stores the very first linked list node in this probability distribution.
*/

/**
* Stores the very last linked list node in this probability distribution.
*/

/**
* Construct a new probability distribution.
*/
super();
}

/**
* Constructs a new probability distribution using the input random number
* generator.
*
* @param random the random number generator to use.
*/
super(random);
}

/**
* {@inheritDoc }
*/
@Override
public boolean addElement(final E element, final double weight) {
checkWeight(weight);

if (this.map.containsKey(element)) {
return false;
}

} else {
}

this.map.put(element, newnode);
this.size++;
this.totalWeight += weight;
return true;
}

/**
* {@inheritDoc }
*/
@Override
public E sampleElement() {
checkNotEmpty();
double value = this.random.nextDouble() * this.totalWeight;

node != null;
if (value < node.getWeight()) {
return node.getElement();
}

value -= node.getWeight();
}

throw new IllegalStateException("Should not get here.");
}

/**
* {@inheritDoc }
*/
@Override
public boolean contains(E element) {
return this.map.containsKey(element);
}

/**
* {@inheritDoc }
*/
@Override
public boolean removeElement(E element) {

if (node == null) {
return false;
}

this.map.remove(element);
this.size--;
this.totalWeight -= node.getWeight();
return true;
}

/**
* {@inheritDoc }
*/
@Override
public void clear() {
this.size = 0;
this.totalWeight = 0.0;
this.map.clear();
}

if (left != null) {
} else {
}

if (right != null) {
} else {
}
}
}


The third data structure relies on a binary tree and runs all the three main methods in worst-case logarithmic time. It looks like this:

Above, the red nodes are the leaf nodes containing the actual elements. White nodes are called in the code relay nodes. The integers in each node denote how many leaf nodes a particular relay node contains, and the real numbers denote the sum of weights of all the leaves of a relay node.

The data structure follows:

BinaryTreeProbabilityDistribution.java:

package net.coderodde.stat.support;

import java.util.Deque;
import java.util.HashMap;
import java.util.Map;
import java.util.Random;
import net.coderodde.stat.AbstractProbabilityDistribution;

/**
* This class implements a probability distribution relying on a binary tree
* structure. It allows <tt>O(log n)</tt> worst case time for adding, removing
* and sampling an element.
*
* @author Rodion "rodde" Efremov
* @version 1.6 (Jun 11, 2016)
* @param <E> the actual type of the elements stored in this distribution.
*/
public class BinaryTreeProbabilityDistribution<E>
extends AbstractProbabilityDistribution<E> {

private static final class Node<E> {

/**
* Holds the element if this block is a leaf. Internal blocks have
* {@code null} assigned to this field.
*/
private final E element;

/**
* If this block is a leaf, specifies the weight of the {@code element}.
* Otherwise, this field caches the sum of all weights over all
* descendant leaves.
*/
private double weight;

private boolean isRelayNode;

/**
* The left child node.
*/
private Node<E> leftChild;

/**
* The right child node.
*/
private Node<E> rightChild;

/**
* The parent node.
*/
private Node<E> parent;

/**
* Caches the number of leaf nodes in the subtree starting from this
* node.
*/
private int numberOfLeafNodes;

Node(final E element, final double weight) {
this.element = element;
this.weight  = weight;
this.numberOfLeafNodes = 1;
}

Node() {
this.element = null;
this.isRelayNode = true;
}

public String toString() {
if (this.isRelayNode) {
return "[" + String.format("%.3f", this.getWeight()) +
" : " + this.numberOfLeafNodes + "]";
}

return "(" + String.format("%.3f", this.getWeight()) +
" : " + this.element + ")";
}

E getElement() {
return this.element;
}

double getWeight() {
return this.weight;
}

void setWeight(final double weight) {
this.weight = weight;
}

int getNumberOfLeaves() {
return this.numberOfLeafNodes;
}

void setNumberOfLeaves(final int numberOfLeaves) {
this.numberOfLeafNodes = numberOfLeaves;
}

Node<E> getLeftChild() {
return this.leftChild;
}

void setLeftChild(final Node<E> block) {
this.leftChild = block;
}

Node<E> getRightChild() {
return this.rightChild;
}

void setRightChild(final Node<E> block) {
this.rightChild = block;
}

Node<E> getParent() {
return this.parent;
}

void setParent(final Node<E> block) {
this.parent = block;
}

boolean isRelayNode() {
return isRelayNode;
}

boolean isLeafNode() {
return !isRelayNode;
}
}

/**
* Maps each element to the list of nodes representing the element.
*/
private final Map<E, Node<E>> map = new HashMap<>();

/**
* The root node of this distribution tree.
*/
private Node<E> root;

/**
* Constructs this probability distribution using a default random number
* generator.
*/
public BinaryTreeProbabilityDistribution() {
this(new Random());
}

/**
* Constructs this probability distribution using the input random number
* generator.
*
* @param random the random number generator to use.
*/
public BinaryTreeProbabilityDistribution(final Random random) {
super(random);
}

/**
* {@inheritDoc }
*/
@Override
public boolean addElement(E element, double weight) {
checkWeight(weight);

if (this.map.containsKey(element)) {
return false;
}

final Node<E> newnode = new Node<>(element, weight);
insert(newnode);
this.size++;
this.totalWeight += weight;
this.map.put(element, newnode);
return true;
}

/**
* {@inheritDoc }
*/
@Override
public boolean contains(E element) {
return this.map.containsKey(element);
}

/**
* {@inheritDoc }
*/
@Override
public E sampleElement() {
checkNotEmpty();

double value = this.totalWeight * this.random.nextDouble();
Node<E> node = root;

while (node.isRelayNode()) {
if (value < node.getLeftChild().getWeight()) {
node = node.getLeftChild();
} else {
value -= node.getLeftChild().getWeight();
node = node.getRightChild();
}
}

return node.getElement();
}

/**
* {@inheritDoc }
*/
@Override
public boolean removeElement(final E element) {
final Node<E> node = this.map.get(element);

if (node == null) {
return false;
}

delete(node);
this.size--;
this.totalWeight -= node.getWeight();
return true;
}

/**
* {@inheritDoc }
*/
@Override
public void clear() {
this.root = null;
this.size = 0;
this.totalWeight = 0.0;
}

/**
* Assuming that {@code leafNodeToBypass} is a leaf node, this procedure
* attaches a relay node instead of it, and assigns {@code leafNodeToBypass}
* and {@code newnode} as children of the new relay node.
*
* @param leafNodeToBypass the leaf node to bypass.
* @param newNode          the new node to add.
*/
private void bypassLeafNode(final Node<E> leafNodeToBypass,
final Node<E> newNode) {
final Node<E> relayNode = new Node<>();
final Node<E> parentOfCurrentNode = leafNodeToBypass.getParent();

relayNode.setNumberOfLeaves(1);
relayNode.setWeight(leafNodeToBypass.getWeight());
relayNode.setLeftChild(leafNodeToBypass);
relayNode.setRightChild(newNode);

leafNodeToBypass.setParent(relayNode);
newNode.setParent(relayNode);

if (parentOfCurrentNode == null) {
this.root = relayNode;
} else if (parentOfCurrentNode.getLeftChild() == leafNodeToBypass) {
relayNode.setParent(parentOfCurrentNode);
parentOfCurrentNode.setLeftChild(relayNode);
} else {
relayNode.setParent(parentOfCurrentNode);
parentOfCurrentNode.setRightChild(relayNode);
}

}

private void insert(final Node<E> node) {
if (root == null) {
root = node;
return;
}

Node<E> currentNode = root;

while (currentNode.isRelayNode()) {
if (currentNode.getLeftChild().getNumberOfLeaves() <
currentNode.getRightChild().getNumberOfLeaves()) {
currentNode = currentNode.getLeftChild();
} else {
currentNode = currentNode.getRightChild();
}
}

bypassLeafNode(currentNode, node);
}

private void delete(final Node<E> leafToDelete) {
final Node<E> relayNode = leafToDelete.getParent();

if (relayNode == null) {
this.root = null;
return;
}

final Node<E> parentOfRelayNode = relayNode.getParent();
final Node<E> siblingLeaf = relayNode.getLeftChild() == leafToDelete ?
relayNode.getRightChild() :
relayNode.getLeftChild();

if (parentOfRelayNode == null) {
this.root = siblingLeaf;
siblingLeaf.setParent(null);
return;
}

if (parentOfRelayNode.getLeftChild() == relayNode) {
parentOfRelayNode.setLeftChild(siblingLeaf);
} else {
parentOfRelayNode.setRightChild(siblingLeaf);
}

siblingLeaf.setParent(parentOfRelayNode);
}

/**
* This method is responsible for updating the metadata of this data
* structure.
*
* @param node      the node from which to start the metadata update. The
*                  predecessors of this node in the tree.
* @param weight    the weight delta to add to each predecessor node.
* @param nodeDelta the node count delta to add to each predecessor node.
*/
final double weightDelta,
final int nodeDelta) {
while (node != null) {
node.setNumberOfLeaves(node.getNumberOfLeaves() + nodeDelta);
node.setWeight(node.getWeight() + weightDelta);
node = node.getParent();
}
}

public String debugToString() {
if (root == null) {
return "empty";
}

final StringBuilder sb = new StringBuilder();
final int treeHeight = getTreeHeight(root);
final Deque<Node<E>> queue = new LinkedList<>();

for (int i = 0; i < treeHeight + 1; ++i) {
int currentQueueLength = queue.size();

for (int j = 0; j < currentQueueLength; ++j) {
final Node<E> node = queue.removeFirst();
sb.append(node == null ? "null" : node.toString()).append(" ");
}

sb.append("\n");
}

return sb.toString();
}

private void addChildren(final Node<E> node, final Deque<Node<E>> queue) {
if (node == null) {
return;
}

}

private int getTreeHeight(final Node<E> node) {
if (node == null) {
return -1;
}

return 1 + Math.max(getTreeHeight(node.getLeftChild()),
getTreeHeight(node.getRightChild()));
}
}


Finally, the demo (I don't want to get reviewed) is...

import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
import java.util.Random;
import net.coderodde.stat.AbstractProbabilityDistribution;
import net.coderodde.stat.support.ArrayProbabilityDistribution;
import net.coderodde.stat.support.BinaryTreeProbabilityDistribution;

public class Demo {

private static final int DISTRIBUTION_SIZE = 20_000;

public static void main(final String[] args) {
System.out.println("[DEMO] BinaryTreeProbabilityDistribution:");
binaryTreeProbabilityDistributionDemo();

System.out.println("[STATUS] Warming up...");
warmup();
System.out.println("[STATUS] Warming up done!");
System.out.println();

AbstractProbabilityDistribution<Integer> arraypd =
new ArrayProbabilityDistribution<>();

AbstractProbabilityDistribution<Integer> listpd =

AbstractProbabilityDistribution<Integer> treepd =
new BinaryTreeProbabilityDistribution<>();

profile(arraypd);
profile(listpd);
profile(treepd);
}

private static void binaryTreeProbabilityDistributionDemo() {
BinaryTreeProbabilityDistribution<Integer> pd =
new BinaryTreeProbabilityDistribution<>();

int[] counts = new int[4];

for (int i = 0; i < 100; ++i) {
Integer myint = pd.sampleElement();
counts[myint]++;
System.out.println(myint);
}

System.out.println(Arrays.toString(counts));
}

private static void
profile(final AbstractProbabilityDistribution<Integer> pd) {
final Random random = new Random();

System.out.println("[" + pd.getClass().getSimpleName() + "]:");

long totalDuration = 0L;

long startTime = System.currentTimeMillis();
for (int i = 0; i < DISTRIBUTION_SIZE; ++i) {
}

long endTime = System.currentTimeMillis();

System.out.println("addElement() in " + (endTime - startTime) +
" milliseconds.");
totalDuration += (endTime - startTime);

startTime = System.currentTimeMillis();

for (int i = 0; i < DISTRIBUTION_SIZE; ++i) {
pd.sampleElement();
}

endTime = System.currentTimeMillis();

System.out.println("sampleElement() in " + (endTime - startTime) +
" milliseconds.");
totalDuration += (endTime - startTime);

final List<Integer> contents = new ArrayList<>(DISTRIBUTION_SIZE);

for (int i = 0; i < DISTRIBUTION_SIZE; ++i) {
}

shuffle(contents);

startTime = System.currentTimeMillis();

for (Integer i : contents) {
pd.removeElement(i);
}

endTime = System.currentTimeMillis();

System.out.println("removeElement() in " + (endTime - startTime) +
" milliseconds.");

totalDuration += (endTime - startTime);

System.out.println("Total duration: " + totalDuration +
" milliseconds.");

System.out.println();
}

private static void shuffle(final List<Integer> list) {
final Random random = new Random();

for (int i = 0; i < list.size(); ++i) {
final int index = random.nextInt(list.size());
final Integer integer = list.get(index);
list.set(index, list.get(i));
list.set(i, integer);
}
}

private static void warmup() {
final long seed =35214717058750L; System.nanoTime();
final Random inputRandom1 = new Random(seed);
final Random inputRandom2 = new Random(seed);
final Random inputRandom3 = new Random(seed);

final AbstractProbabilityDistribution<Integer> pd1 =
new ArrayProbabilityDistribution<>(inputRandom1);

final AbstractProbabilityDistribution<Integer> pd2 =

final AbstractProbabilityDistribution<Integer> pd3 =
new BinaryTreeProbabilityDistribution<>(inputRandom3);

final Random random = new Random(seed);
final List<Integer> content = new ArrayList<>();

System.out.println("Seed = " + seed);

for (int iteration = 0; iteration < 100_000; ++iteration) {
final double coin = random.nextDouble();

if (coin < 0.3) {
final Integer element = random.nextInt();
final double weight = 30.0 * random.nextDouble();

} else if (coin < 0.5) {
// Remove an element.
if (!pd1.isEmpty()) {
final Integer element = choose(content, random);

pd1.removeElement(element);
pd2.removeElement(element);
pd3.removeElement(element);
content.remove(element);
}
} else if (!pd1.isEmpty()) {
// Sample elements:
pd1.sampleElement();
pd2.sampleElement();
pd3.sampleElement();
}
}
}

private static Integer choose(final List<Integer> list,
final Random random) {
return list.get(random.nextInt(list.size()));
}
}


The performance figures are as follows:


[STATUS] Warming up...
Seed = 35214717058750
[STATUS] Warming up done!

[ArrayProbabilityDistribution]:
sampleElement() in 321 milliseconds.
removeElement() in 500 milliseconds.
Total duration: 829 milliseconds.

sampleElement() in 1184 milliseconds.
removeElement() in 9 milliseconds.
Total duration: 1200 milliseconds.

[BinaryTreeProbabilityDistribution]:
sampleElement() in 15 milliseconds.
removeElement() in 16 milliseconds.
Total duration: 55 milliseconds.



Critique request

I would like to hear the comments regarding the following:

• API design,
• naming conventions,
• coding conventions,
• performance.

### Inconsistent use of this

I'm not sure why you used a this prefix for many of your variables, when it was optional to do so. I would have omitted it myself. But assuming you had a reason for doing so, you didn't do it consistently. It seems like around 70% of the places use it and 30% don't.

### Array implementation suggestion

You could modify your Array implementation to do the following:

1. In addition to storing an array of objects and an array of weights, you could make a third array which stores cumulative weights. In other words, cumulativeWeight[n] would be the sum of weightStorageArray[i] for i = 0..n.
2. With the cumulative weight array, sampleElement() could now be done using a binary search through the cumulative weight array. This reduces sample time to $O(\log n)$ instead of $O(n)$.
3. Adding an element remains constant time because you can compute the cumulative weight of the new element in constant time like this:

cumulativeWeight[this.size] = weight +
(this.size == 0) ? 0 : cumulativeWeight[this.size-1];

4. Removing an element still takes linear time, and you will need to adjust each element of cumulativeWeight as you shift it downwards.

This implemenation would potentially be superior to the binary tree implementation if removals were expected to be rare.

• Benchmarked the binary search version: you are right; adding 4 times faster + sampling 2 times faster. – coderodde Jun 17 '16 at 12:37
• What comes to the (inconsistent) usage of this, I am still not sure what my coding conventions should be, so I am still experimenting with it. – coderodde Jun 17 '16 at 12:39