Purely functional data structure
This article has multiple issues. Please help improve it or discuss these issues on the talk page. (Learn how and when to remove these template messages)(Learn how and when to remove this template message)
In computer science, a purely functional data structure is a data structure that can be implemented in a purely functional language. The main difference between an arbitrary data structure and a purely functional one is that the latter is (strongly) immutable. This restriction ensures the data structure possesses the advantages of immutable objects: (full) persistency, quick copy of objects, and thread safety. Efficient purely functional data structures may require the use of lazy evaluation and memoization.
Purely functional data structures are often represented in a different way than their imperative counterparts. For example, an array with constant-time access and update is a basic component of most imperative languages. Many imperative data structures, such as the hash table and binary heap, are based on arrays. An array can be replaced by a map or random access list, which admits a purely functional implementation, but access and update operations may run in logarithmic time. Purely functional data structures can be implemented in imperative and object-oriented languages, but their time and/or space performance may be inferior to that of data structures lacking purely functional properties.
Ensuring that a data structure is purely functional
A data structure is never inherently functional. For example, a stack can be implemented as a singly-linked list. This implementation is purely functional as long as the only operations on the stack return a new stack without altering the old stack. However, if the language is not purely functional, the run-time system may be unable to guarantee immutability.
Here is a list of abstract data structures with purely functional implementations:
- Stack (first in, last out) implemented as a singly linked list,
- Queue, implemented as a real-time queue,
- Double-ended queue, implemented as a real-time double-ended queue,
- (Multi)set of ordered elements and map indexed by ordered keys, implemented as a red–black tree, or more generally by a search tree,
- Priority queue, implemented as a Brodal queue
- Random access list, implemented as a skew-binary random access list
Design and implementation
In his book Purely Functional Data Structures, computer scientist Chris Okasaki describes techniques used to design and implement purely functional data structures, a small subset of which are summarized below.
Laziness and memoization
Lazy evaluation is particularly interesting in a purely functional language because the order of the evaluation never changes the result of a function. Therefore, lazy evaluation naturally becomes an important part of the construction of purely functional data structures. It allows computations to be done only when its result is actually required. Therefore, the code of a purely functional data structure can, without loss of efficiency, consider similar data that will effectively be used and data that will be ignored. The only computation required is for the first kind of data that will actually be performed.
One of the key tools in building efficient, purely functional data structures is memoization. When a computation is done, it is saved and does not have to be performed a second time. This is particularly important in lazy implementations; additional evaluations may require the same result, but it is impossible to know which evaluation will require it first.
Amortized analysis and scheduling
Some data structures, even those that are not purely functional such as dynamic arrays, admit operation that are efficient most of the time (i.e. constant time for dynamic arrays), and rarely inefficient (i.e. linear time for dynamic arrays). Amortization can then be used to prove that the average running time of the operations are efficient. That is to say, the few inefficient operations are rare enough, and do not change the asymptotical evolution of time complexity when a sequence of operations is considered.
In general, having inefficient operations is not acceptable for persistent data structures, because this very operation can be called many times. It is not acceptable either for real-time or for imperative systems, where the user may require the time taken by the operation to be predictable. Furthermore, this unpredictability complicates the use of parallelism.
In order to avoid those problems, some data structures allow for the inefficient operation to be postponed—this is called scheduling. The only requirement is that the computation of the inefficient operation should end before its result is actually needed. A constant part of the inefficient operation is performed simultaneously with the following call to an efficient operation, so that the inefficient operation is already totally done when it is needed, and each individual operation remains efficient.[clarification needed]
Amortized queues are composed of two singly-linked lists: the front and the reversed rear. Elements are added to the rear list and are removed from the front list. Furthermore, whenever the front queue is empty, the rear queue is reversed and becomes the front, while the rear queue becomes empty. The amortized time complexity of each operation is constant. Each cell of the list is added, reversed and removed at most once. In order to avoid an inefficient operation where the rear list is reversed, real-time queues add the restriction that the rear list is only as long as the front list. To ensure that the rear list becomes longer than the front list, the front list is appended to the rear list and reversed. Since this operation is inefficient, it is not performed immediately. Instead, it is carried out for each of the operations. Thus, each cell is computed before it is needed, and the new front list is totally computed before a new inefficient operation needs to be called.
- Purely Functional Data Structures thesis by Chris Okasaki (PDF format)
- Making Data-Structures Persistent by James R. Driscoll, Neil Sarnak, Daniel D. Sleator, Robert E. Tarjan (PDF)
- Fully Persistent Lists with Catenation by James R. Driscoll, Daniel D. Sleator, Robert E. Tarjan (PDF)
- Persistent Data Structures from the MIT OpenCourseWare course Advanced Algorithms
- What's new in purely functional data structures since Okasaki? on Theoretical Computer Science StackExchange