by Sven Woltmann – May 28, 2020

The big O notation¹ is used to describe the complexity of algorithms.

On Google and YouTube, you can find numerous articles and videos explaining the big O notation. But to understand most of them (like this Wikipedia article), you should have studied mathematics as a preparation. ;-)

That' s why, in this article, I will explain the big O notation (and the time and space complexity described with it) only using examples and diagrams – and entirely without mathematical formulas, proofs and symbols like θ, Ω, ω, ∈, ∀, ∃ and ε.

You can find all source codes from this article in my GitHub repository.

¹ also known as "Bachmann-Landau notation" or "asymptotic notation"

Computational time complexity describes the change in the runtime of an algorithm, depending on the change in the input data's size.

In other words: "How much does an algorithm degrade when the amount of input data increases?"

Examples:

- How much longer does it take to find an element within an
*unsorted*array when the size of the array doubles? (Answer: twice as long) - How much longer does it take to find an element within a
*sorted*array when the size of the array doubles? (Answer: one more step)

Space complexity describes how much additional memory an algorithm needs depending on the size of the input data.

This does not mean the memory required for the input data itself (i.e., that twice as much space is naturally needed for an input array twice as large), but the additional memory needed by the algorithm for loop and helper variables, temporary arrays, etc.

We divide algorithms into so-called complexity classes. A complexity class is identified by the Landau symbol *O* ("big O").

In the following section, I will explain the most common complexity classes, starting with the easy to understand classes and moving on to the more complex ones. Accordingly, the classes are not sorted by complexity.

Pronounced: "Order 1", "O of 1", "big O of 1"

The runtime is constant, i.e., independent of the number of input elements *n*.

In the following graph, the horizontal axis represents the number of input elements *n* (or more generally: the size of the input problem), and the vertical axis represents the time required. Since complexity classes can only be used to classify algorithms, but not to calculate their exact running time, the axes are not labeled.

The following two problems are examples of constant time:

- Accessing a specific element of an array of size
*n*: No matter how large the array is, accessing it via`array[index]`

always takes the same time². - Inserting an element at the beginning of a linked list: This always requires setting one or two (for a doubly linked list) pointers (or references), regardless of the list's size. (In an array, on the other hand, this would require moving all values one field to the right, which takes longer with a larger array than with a smaller one).

² This statement is not one hundred percent correct. Effects from CPU caches also come into play here: If the data block containing the element to be read is already (or still) in the CPU cache (which is more likely the smaller the array is), then access is faster than if it first has to be read from RAM.

The following source code (class ConstantTimeSimpleDemo in the GitHub repository) shows a simple example to measure the time required to insert an element at the beginning of a linked list:

```
public static void main(String[] args) {
for (int n = 32; n <= 8_388_608; n *= 2) {
LinkedList<Integer> list = new LinkedList<>();
for (int i = 0; i < n; i++) {
list.add(i);
}
long time = System.nanoTime();
list.add(0, 1);
time = System.nanoTime() - time;
System.out.printf("n = %d -> time = %d%n", n, time);
}
}
```

Code language: Java (java)

On my system, the times are between 1,200 and 19,000 ns, unevenly distributed over the various measurements. This is sufficient for a quick test. But we don't get particularly good measurement results here, as both the HotSpot compiler and the garbage collector can kick in at any time.

The test program TimeComplexityDemo with the ConstantTime class provides better measurement results. The test program first runs several warmup rounds to allow the HotSpot compiler to optimize the code. Only after that are measurements performed five times, and the median of the measured values is displayed.

Here is an extract of the results:

Code language: plaintext (plaintext)`--- ConstantTime (results 5 of 5) --- ConstantTime, n = 32 -> fastest: 31,700 ns, median: 44,900 ns ConstantTime, n = 16,384 -> fastest: 14,400 ns, median: 40,200 ns ConstantTime, n = 8,388,608 -> fastest: 34,000 ns, median: 51,100 ns`

The effort remains about the same, regardless of the size of the list. The complete test results can be found in the file test-results.txt.

Pronounced: "Order n", "O of n", "big O of n"

The time grows linearly with the number of input elements *n*: If *n* doubles, then the time approximately doubles, too.

"Approximately" because the effort may also include components with lower complexity classes. These become insignificant if *n* is sufficiently large so they are omitted in the notation.

In the following diagram, I have demonstrated this by starting the graph slightly above zero (meaning that the effort also contains a constant component):

The following problems are examples for linear time:

- Finding a specific element in an array: All elements of the array have to be examined – if there are twice as many elements, it takes twice as long.
- Summing up all elements of an array: Again, all elements must be looked at once – if the array is twice as large, it takes twice as long.

It is essential to understand that the complexity class makes no statement about the absolute time required, but only about the change in the time required depending on the change in the input size. The two examples above would take much longer with a linked list than with an array – but that is irrelevant for the complexity class.

The following source code (class LinearTimeSimpleDemo) measures the time for summing up all elements of an array:

```
public static void main(String[] args) {
for (int n = 32; n <= 536_870_912; n *= 2) {
int[] array = new int[n];
for (int i = 0; i < n; i++) {
array[i] = i;
}
long sum = 0;
long time = System.nanoTime();
for (int i = 0; i < n; i++) {
sum += array[i];
}
time = System.nanoTime() - time;
System.out.printf("n = %d -> time = %d ns%n", n, time);
}
}
```

Code language: Java (java)

On my system, the time degrades approximately linearly from 1,100 ns to 155,911,900 ns. Better measurement results are again provided by the test program TimeComplexityDemo and the LinearTime class. Here is an extract of the results:

Code language: plaintext (plaintext)`--- LinearTime (results 5 of 5) --- LinearTime, n = 512 -> fastest: 300 ns, median: 300 ns LinearTime, n = 524,288 -> fastest: 159,300 ns, median: 189,400 ns LinearTime, n = 536,870,912 -> fastest: 164,322,600 ns, median: 168,681,700 ns`

You can find the complete test results again in test-results.txt.

A function is linear if it can be represented by a straight line, e.g. *f(x) = 5x + 3*.

*Proportional* is a particular case of *linear*, where the line passes through the point (0,0) of the coordinate system, for example, *f(x) = 3x*.

As there may be a constant component in *O(n)*, it's time is *linear*.

Pronounced: "Order n squared", "O of n squared", "big O of n squared"

The time grows linearly to the square of the number of input elements: If the number of input elements *n* doubles, then the time roughly quadruples. (And if the number of elements increases tenfold, the effort increases by a factor of one hundred!)

Examples of quadratic time are simple sorting algorithms like Insertion Sort, Selection Sort, and Bubble Sort.

The following example (QuadraticTimeSimpleDemo) shows how the time for sorting an array using *Insertion Sort* changes depending on the size of the array:

```
public static void main(String[] args) {
for (int n = 32; n <= 262_144; n *= 2) {
ThreadLocalRandom random = ThreadLocalRandom.current();
int[] array = new int[n];
for (int i = 0; i < n; i++) {
array[i] = random.nextInt();
}
long time = System.nanoTime();
insertionSort(array);
time = System.nanoTime() - time;
System.out.printf("n = %d -> time = %d ns%n", n, time);
}
}
private static void insertionSort(int[] elements) {
for (int i = 1; i < elements.length; i++) {
int elementToSort = elements[i];
int j = i;
while (j > 0 && elementToSort < elements[j - 1]) {
elements[j] = elements[j - 1];
j--;
}
elements[j] = elementToSort;
}
}
```

Code language: Java (java)

On my system, the time required increases from 7,700 ns to 5.5 s. You can see reasonably well how time quadruples each time the array size doubles. However, I also see a reduction of the time needed about halfway through the test – obviously, the HotSpot compiler has optimized the code there.

We can obtain better measurement results with the test program TimeComplexityDemo and the QuadraticTime class. Here is an excerpt of the results, where you can see the approximate quadrupling of the effort each time the problem size doubles:

Code language: plaintext (plaintext)`QuadraticTime, n = 8,192 -> fastest: 4,648,400 ns, median: 4,720,200 ns QuadraticTime, n = 16,384 -> fastest: 19,189,100 ns, median: 19,440,400 ns QuadraticTime, n = 32,768 -> fastest: 78,416,700 ns, median: 79,896,000 ns QuadraticTime, n = 65,536 -> fastest: 319,905,300 ns, median: 330,530,600 ns QuadraticTime, n = 131,072 -> fastest: 1,310,702,600 ns, median: 1,323,919,500 ns`

You can find the complete test results in test-results.txt.

At this point, I would like to point out again that the effort can contain components of lower complexity classes and constant factors. Both are irrelevant for the big O notation since they are no longer of importance if *n* is sufficiently large.

It is therefore also possible that, for example, *O(n²)* is faster than *O(n)* – at least up to a certain size of *n*.

The following example diagram compares three fictitious algorithms: one with complexity class *O(n²)* and two with *O(n)*, one of which is faster than the other. It is good to see how up to *n = 4*, the orange *O(n²)* algorithm takes less time than the yellow *O(n)* algorithm. And even up to *n = 8*, less time than the cyan *O(n)* algorithm.

Above sufficiently large *n* – i.e., from *n = 9* – *O(n²)* is and remains the slowest algorithm.

Let's move on to two, not quite so intuitively understandable complexity classes.

Pronounced: "Order log n", "O of log n", "big O of log n"

The effort increases approximately by a constant amount when the number of input elements doubles.

For example, if the time increases by one second when the number of input elements increases from 1,000 to 2,000, it only increases by another second when the effort increases to 4,000. And again by one more second when the effort grows to 8,000.

An example of logarithmic effort is the binary search for a specific element in a sorted array of size *n*.

Since we halve the area to be searched with each search step, we can, in turn, search an array twice as large with only one more search step.

(The older ones among us may remember this from searching the telephone book or an encyclopedia.)

The following example (LogarithmicTimeSimpleDemo) measures how the time for binary search in a sorted array changes in relation to the size of the array.

```
public static void main(String[] args) {
for (int n = 32; n <= 536_870_912; n *= 2) {
int[] array = new int[n];
for (int i = 0; i < n; i++) {
array[i] = i;
}
long time = System.nanoTime();
Arrays.binarySearch(array, 0);
time = System.nanoTime() - time;
System.out.printf("n = %d -> time = %d ns%n", n, time);
}
}
```

Code language: Java (java)

I can recognize the expected constant growth of time with doubled problem size to some extent. As before, we get better measurement results with the test program TimeComplexityDemo and the class LogarithmicTime. Here are the results:

Code language: plaintext (plaintext)`LogarithmicTime, n = 32 -> fastest: 77,800 ns, median: 107,200 ns LogarithmicTime, n = 2,048 -> fastest: 173,500 ns, median: 257,400 ns LogarithmicTime, n = 131,072 -> fastest: 363,400 ns, median: 413,100 ns LogarithmicTime, n = 8,388,608 -> fastest: 661,100 ns, median: 670,800 ns LogarithmicTime, n = 536,870,912 -> fastest: 770,500 ns, median: 875,700 ns`

In each step, the problem size *n* increases by factor 64. The time does not always increase by exactly the same value, but it does so sufficiently precisely to demonstrate that logarithmic time is significantly cheaper than linear time (for which the time required would also increase by factor 64 each step).

As before, you can find the complete test results in the file test-results.txt.

Pronounced: "Order n log n", "O of n log n", "big O of n log n"

The effort grows slightly faster than linear because the linear component is multiplied by a logarithmic one. For clarification, you can also insert a multiplication sign: *O(n × log n)*.

This is best illustrated by the following graph. We see a curve whose gradient is visibly growing at the beginning, but soon approaches a straight line as *n* increases:

Efficient sorting algorithms like Quicksort, Merge Sort, and Heapsort are examples for quasilinear time.

The following sample code (class QuasiLinearTimeSimpleDemo) shows how the effort for sorting an array with Quicksort³ changes in relation to the array size:

```
public static void main(String[] args) {
for (int n = 64; n <= 67_108_864; n *= 2) {
ThreadLocalRandom random = ThreadLocalRandom.current();
int[] array = new int[n];
for (int i = 0; i < n; i++) {
array[i] = random.nextInt();
}
long time = System.nanoTime();
Arrays.sort(array);
time = System.nanoTime() - time;
System.out.printf("n = %d -> time = %d ns%n", n, time);
}
}
```

Code language: Java (java)

On my system, I can see very well how the effort increases roughly in relation to the array size (where at *n = 16,384*, there is a backward jump, obviously due to HotSpot optimizations). The test program TimeComplexityDemo with the class QuasiLinearTime delivers more precise results. Here is an extract:

Code language: plaintext (plaintext)`QuasiLinearTime, n = 256 -> fastest: 12,200 ns, med.: 12,500 ns QuasiLinearTime, n = 4,096 -> fastest: 228,600 ns, med.: 234,200 ns QuasiLinearTime, n = 65,536 -> fastest: 4,606,500 ns, med.: 4,679,800 ns QuasiLinearTime, n = 1,048,576 -> fastest: 93,933,500 ns, med.: 95,216,300 ns QuasiLinearTime, n = 16,777,216 -> fastest: 1,714,541,900 ns, med.: 1,755,715,000 ns`

The problem size increases each time by factor 16, and the time required by factor 18.5 to 20.3. You can find the complete test result, as always, in test-results.txt.

³ More precisely: Dual-Pivot Quicksort, which switches to Insertion Sort for arrays with less than 44 elements. For this reason, this test starts at 64 elements, not at 32 like the others.

Here are, once again, the described complexity classes, sorted in ascending order of complexity (for sufficiently large values of *n*):

*O(1)*– constant time*O(log n)*– logarithmic time*O(n)*– linear time*O(n log n)*– quasilinear time*O(n²)*– quadratic time

And here the comparison graphically:

I intentionally shifted the curves along the time axis so that the worst complexity class *O(n²)* is fastest for low values of *n*, and the best complexity class *O(1)* is slowest. To then show how, for sufficiently high values of *n*, the efforts shift as expected.

Further complexity classes are, for example:

*O(n*– polynomial time,^{m})*O(2*– exponential time,^{n})*O(n!)*– factorial time.

However, these are so bad that we should avoid algorithms with these complexities, if possible.

I have included these classes in the following diagram (*O(n ^{m})* with

I had to compress the y-axis by factor 10 compared to the previous diagram to display the three new curves.

Time complexity describes how the runtime of an algorithm changes depending on the amount of input data. The most common complexity classes are (in ascending order of complexity): *O(1)*, *O(log n)*, *O(n)*, *O(n log n)*, *O(n²)*.

Algorithms with constant, logarithmic, linear, and quasilinear time usually lead to an end in a reasonable time for input sizes up to several billion elements. Algorithms with quadratic time can quickly reach theoretical execution times of several years for the same problem sizes⁴. You should, therefore, avoid them as far as possible.

⁴ Quicksort, for example, sorts a billion items in 90 seconds on my laptop; Insertion Sort, on the other hand, needs 85 seconds for a million items; that would be 85 million seconds for a billion items - or in other words: little over two years and eight months!

If you liked the article, please leave me a comment, share the article via one of the share buttons, or subscribe to my mailing list to be informed about new articles.

About the author

I'm a freelance software developer with more than two decades of experience in scalable Java enterprise applications. My focus is on optimizing complex algorithms and on advanced topics such as concurrency, the Java memory model, and garbage collection.
Here on HappyCoders.eu, I want to help you become a better Java programmer. Read more about me here.

Thanks, great explanation. I'm happy coder now ;)

I am happy to hear that. I know it's a difficult subject, I had a great deal of trouble understanding it in college.