Pohon Fenwick

1. Introduction

A Binary Indexed (Fenwick) Tree is a data structure that provides efficient methods for implementing dynamic cumulative frequency tables.

This Fenwick Tree data structure uses many bit manipulation techniques. In this visualization, we will refer to this data structure using the term Fenwick Tree as the abbreviation 'BIT' of Binary Indexed Tree is usually associated with the usual bit manipulation.

1-1. Cumulative Frequency Table

Suppose that we have a multiset of integers s = {2,4,5,6,5,6,8,6,7,9,7} (not necessarily sorted). There are n = 11 elements in s. Also suppose that the largest integer that we will ever use is m = 10 and we never use integer 0. For example, these integers represent student (integer) scores from [1..10]. Notice that n is independent of m.

We can create a frequency table f from s with a trivial O(n) time loop. We can then create cumulative frequency table cf from frequency table f in O(m) time using technique similar to DP 1D prefix sum.

Index/Score/SymbolFrequency fCumulative Frequency cf
0-- (index 0 is ignored)
524 == cf[4]+f[5]
637 == cf[5]+f[6]
10 == m011 == n

1-2. Range Sum Query: rsq(i, j)

With such cumulative frequency table cf, we can perform Range Sum Query: rsq(i, j) to return the sum of frequencies between index i and j (inclusive), in efficient O(1) time, again using the DP 1D prefix sum (i.e., the inclusion-exclusion principle). For example, rsq(5, 9) = rsq(1, 9) - rsq(1, 4) = 11-2 = 9. Since for this example, these key 5, 6, 7, 8, and 9 represent scores, rsq(5, 9) means the total number of students who scored between 5 to 9, inclusive.

Index/Score/SymbolFrequency fCumulative Frequency cf
0-- (index 0 is ignored)
412 == rsq(1, 4)
9111 == rsq(1, 9)
10 == m011 == n

1-3. Dynamic Cumulative Frequency Table

A dynamic data structure need to support (frequent) updates in between queries. For example, we may update (add) the frequency of score 7 from 2 → 5 (e.g., 3 more students score 7) and update (subtract) the frequency of score 9 from 1 → 0 (e.g., 1 student who previously scored 9 is found to have plagiarized the work and is now penalized to 0, i.e., removed from the scores), thereby updating the table into:

Index/Score/SymbolFrequency fCumulative Frequency cf
0-- (index 0 is ignored)
72 → 59 → 12
8110 → 13
91 → 011 → 13
10 == m011 → 13 == n

A pure array based data structure will need O(m) per update operation. Can we do better?

2. Mode-mode dan Mode Pertama/Default

Introducing: Fenwick Tree data structure.

There are three mode of usages of Fenwick Tree in this visualization.

The first mode is the default Fenwick Tree that can handle both Point Update (PU) and Range Query (RQ) in O(log n) where n is the largest index/key in the data structure. Remember that the actual number of keys in the data structure is denoted by another variable m. We abbreviate this default type as PU RQ that simply stands for Point Update Range Query.

This clever arrangement of integer keys idea is the one that originally appears in Peter M. Fenwick's 1994 paper.

3. Point Update Range Query (PU RQ)

You can click the 'Create' menu to create a frequency array f where f[i] denotes the frequency of appearance of key i in our original array of keys s.

IMPORTANT: This frequency array f is not the original array of keys s. For example, if you enter {0,1,0,1,2,3,2,1,1,0}, it means that you are creating 0 one, 1 two, 0 three, 1 four, ..., 0 ten (1-based indexing). The largest index/integer key is m = 10 in this example as in the earlier slides.

If you have the original array s of n elements, e.g., {2,4,5,6,5,6,8,6,7,9,7} from the earlier slides (s does not need to be necessarily sorted), you can do an O(n) pass to convert s into frequency table f of n indices/integer keys. (We will provide this alternative input method in the near future).

You can click the 'Randomize' button to generate random frequencies.

3-1. Visualisasi - Bagian 1

Although conceptually this data structure is a tree, it will be implemented as an integer array called ft that ranges from index 1 to index n (we sacrifice index 0 of our ft array). The values inside the vertices of the Fenwick Tree shown above are the values stored in the 1-based Fenwick Tree ft array.

Currently the edges of this Fenwick Tree are not shown yet. There are two versions of the tree, the interrogation/query tree and the updating Tree.

3-2. Visualisasi - Bagian 2

The values inside the vertices at the bottom are the values of the data (the frequency array f).

3-3. Visualisasi - Bagian 3

The value stored in index i in array ft, i.e., ft[i] is the cumulative frequency of keys in range [i-LSOne(i)+1 .. i]. Visually, this range is shown by the edges of the (interrogation/query version of) Fenwick Tree. For details of LSOne(i) operation, see our bitmask visualization page.

For example, ft[6] = 5 stores the cumulative frequency of keys in range of [6-LSOne(6)+1..6] (the edges between index 6 back to 4, plus 1). This is [6-2+1..6] = [5..6] and f[5]+f[6] = 2+3 = 5. Then ft[4] = 2 stores the cumulative frequency of keys in range of [4-LSOne(4)+1..4] (the edges between index 4 back to 0, plus 1). This is [4-4+1..6] = [1..4] and f[1]+f[2]+f[3]+f[4] = 0+1+0+1 = 2.

3-4. Range Query: rsq(j)

The function rsq(j) returns the cumulative frequencies from the first index 1 (ignoring index 0) to index j.

This value is the sum of sub-frequencies stored in array ft with indices related to j via this formula j' = j-LSOne(j). This relationship forms a Fenwick Tree, specifically, the 'interrogation tree' of Fenwick Tree.

We apply this formula iteratively until j is 0. (We will add that dummy vertex 0 later).

Discussion: Do you understand what does this function compute?

This function runs is O(log m), regardless of n. Discussion: Why?

3-5. Range Query: rsq(i, j)

rsq(i, j) returns the cumulative frequencies from index i to j, inclusive.

If i = 1, the previous slide is sufficient.
If i > 1, we simply need to return: rsq(j)–rsq(i–1).

Discussion: Do you understand the reason?

This function also runs in O(log m), regardless of n. Discussion: Why?

3-6. Point Update: update(i, v)

To update the frequency of a key (an index) i by v (v is either positive or negative; |v| does not necessarily be one), we use update(i, v).

Indices that are related to i via i' = i+LSOne(i) will be updated by v when i < ft.size() (Note that ft.size() is m+1 (as we ignore index 0). These relationships form a variant of Fenwick Tree structure called the 'updating tree'.

Discussion: Do you understand this operation and on why we avoided index 0?

This function also runs in O(log m), regardless of n. Discussion: Why?

4. Mode Kedua

Mode kedua dari Pohon Fenwick ini adalah mode yang dapat melakukan Range Update (RU) tetapi hanya dapat melakukan Point Query (PQ) dalam O(log N).

Kita menyingkat tipe ini sebagai RU PQ.

5. Range Update Point Query (RU PQ)

Buatlah sebuah array dan cobalah jalankan algoritma Update Range atau Query Point padanya. Pembuatan data untuk tipe ini berarti memasukkan beberapa interval. Sebagai contoh, juka anda memasukkan [2,4],[3,5], itu berarti kita meng-update range 2 hingga 4 dengan +1 serta range 3 hingga 5 dengan +1 pula. Frekuensi yang dihasilkan adalah: 0,1,2,2,1 yang berarti 0 nilai 1, 1 nilai 2, 2 nilai 3, 2 nilai 4, dan 1 nilai 5.

5-1. Visualisasi RU PQ

Simpul-simpul diatas menunjukan nilai-nilai yang disimpan dalam Pohon Fenwick (larik ft).

Simpul-simpul dibawah menunjukkan nilai-nilai dari data (tabel frekuensi f).

Catat modifikasi pintar dari Pohon Fenwick yang digunakan dalam tipe RU PQ ini: Kita menambah awal dari range sebesar +1 tetapi mengurangi satu indeks setelah akhir dari range sebesar -1 untuk mencapai hasil ini.

6. Mode Ketiga

Mode ketiga dari Pohon Fenwick ini adalah mode yang dapat melakukan Range Update (RU) dan Range Query (RQ) dalam O(log N), sehingga mode ini berjalan se-efisien Pohon Segmen dengan Lazy Update yang juga dapat melakukan RU RQ dalam O(log N).

7. Range Update Range Query (RU RQ)

Buatlah datanya dan coba jalankan algoritma Range Update atau Range Query pada data tersebut.

Pembuatan data dapat dilakukan dengan memasukkan beberapa interval seperti dalam versi RU PQ. Namun, kali ini anda juga melakukan Range Query secara efisien.

7-1. Visualisasi RU RQ

In Range Update Range Query Fenwick Tree, we need to have two Fenwick Trees. The vertices at the top shows the values of the first Fenwick Tree (BIT1[] array), the vertices at the middle shows the values of the second Fenwick Tree (BIT2[] array), while the vertices at the bottom shows the values of the data (the frequency table). The first Fenwick Tree behaves the same as in RU PQ version. The second Fenwick Tree is used to do clever offset to allow Range Query again.

8. Tambahan-Tambahan

Kita memiliki beberapa hal-hal ekstra berhubungan dengan struktur data ini.

8-1. Implementasi

Unfortunately, this data structure is not yet available in C++ STL, Java API, Python or OCaml Standard Library as of 2020. Therefore, we have to write our own implementation.

Please look at the following C++/Python/Java/OCaml implementations of this Fenwick Tree data structure in Object-Oriented Programming (OOP) fashion:
fenwicktree_ds.cpp | py | java | ml

Again, you are free to customize this custom library implementation to suit your needs.