Skip list

Skip Lists: A Probabilistic Alternative to
Balanced Trees
Skip lists are a data structure that can be used in place of balanced trees.
Skip lists use probabilistic balancing rather than strictly enforced balancing
and as a result the algorithms for insertion and deletion in skip lists are
much simpler and significantly faster than equivalent algorithms for
balanced trees.

William Pugh
Binary trees can be used for representing abstract data types Also giving every fourth node a pointer four ahead (Figure
such as dictionaries and ordered lists. They work well when 1c) requires that no more than n/4 + 2 nodes be examined.
the elements are inserted in a random order. Some sequences If every (2i)th node has a pointer 2i nodes ahead (Figure 1d),
of operations, such as inserting the elements in order, produce the number of nodes that must be examined can be reduced to
degenerate data structures that give very poor performance. If log2 n while only doubling the number of pointers. This
it were possible to randomly permute the list of items to be in- data structure could be used for fast searching, but insertion
serted, trees would work well with high probability for any in- and deletion would be impractical.
put sequence. In most cases queries must be answered on-line, A node that has k forward pointers is called a level k node.
so randomly permuting the input is impractical. Balanced tree If every (2i)th node has a pointer 2i nodes ahead, then levels
algorithms re-arrange the tree as operations are performed to of nodes are distributed in a simple pattern: 50% are level 1,
maintain certain balance conditions and assure good perfor- 25% are level 2, 12.5% are level 3 and so on. What would
mance. happen if the levels of nodes were chosen randomly, but in the
Skip lists are a probabilistic alternative to balanced trees. same proportions (e.g., as in Figure 1e)? A node’s ith forward
Skip lists are balanced by consulting a random number gen- pointer, instead of pointing 2i–1 nodes ahead, points to the
erator. Although skip lists have bad worst-case performance, next node of level i or higher. Insertions or deletions would
no input sequence consistently produces the worst-case per- require only local modifications; the level of a node, chosen
formance (much like quicksort when the pivot element is cho- randomly when the node is inserted, need never change. Some
sen randomly). It is very unlikely a skip list data structure will arrangements of levels would give poor execution times, but
be significantly unbalanced (e.g., for a dictionary of more we will see that such arrangements are rare. Because these
than 250 elements, the chance that a search will take more data structures are linked lists with extra pointers that skip
than 3 times the expected time is less than one in a million). over intermediate nodes, I named them skip lists.
Skip lists have balance properties similar to that of search
trees built by random insertions, yet do not require insertions SKIP LIST ALGORITHMS
to be random. This section gives algorithms to search for, insert and delete
Balancing a data structure probabilistically is easier than elements in a dictionary or symbol table. The Search opera-
explicitly maintaining the balance. For many applications, tion returns the contents of the value associated with the de-
skip lists are a more natural representation than trees, also sired key or failure if the key is not present. The Insert opera-
leading to simpler algorithms. The simplicity of skip list algo- tion associates a specified key with a new value (inserting the
rithms makes them easier to implement and provides signifi- key if it had not already been present). The Delete operation
cant constant factor speed improvements over balanced tree deletes the specified key. It is easy to support additional oper-
and self-adjusting tree algorithms. Skip lists are also very ations such as “find the minimum key” or “find the next key”.
space efficient. They can easily be configured to require an Each element is represented by a node, the level of which
average of 1 1 / 3 pointers per element (or even less) and do not is chosen randomly when the node is inserted without regard
require balance or priority information to be stored with each for the number of elements in the data structure. A level i
node. node has i forward pointers, indexed 1 through i. We do not
need to store the level of a node in the node. Levels are
SKIP LISTS capped at some appropriate constant MaxLevel. The level of a
We might need to examine every node of the list when search- list is the maximum level currently in the list (or 1 if the list is
ing a linked list (Figure 1a). If the list is stored in sorted order empty). The header of a list has forward pointers at levels one
and every other node of the list also has a pointer to the node through MaxLevel. The forward pointers of the header at
two ahead it in the list (Figure 1b), we have to examine no levels higher than the current maximum level of the list point
more than n/2 + 1 nodes (where n is the length of the list). to NIL.

a 3 6 7 9 12 17 19 21 25 26 NIL

b 6 9 17 21 26 NIL
3 7 12 19 25

c 9 21 NIL
6 17 26
3 7 12 19 25

d 21 NIL
9
6 17 26
3 7 12 19 25

e 6 NIL
25
9 17
3 7 12 19 21 26

FIGURE 1 - Linked lists with additional pointers

Initialization the previous maximum level of the list, we update the maxi-
An element NIL is allocated and given a key greater than any mum level of the list and initialize the appropriate portions of
legal key. All levels of all skip lists are terminated with NIL. the update vector. After each deletion, we check if we have
A new list is initialized so that the the level of the list is equal deleted the maximum element of the list and if so, decrease
to 1 and all forward pointers of the list’s header point to NIL. the maximum level of the list.
Search Algorithm Choosing a Random Level
We search for an element by traversing forward pointers that Initially, we discussed a probability distribution where half of
do not overshoot the node containing the element being the nodes that have level i pointers also have level i+1 point-
searched for (Figure 2). When no more progress can be made ers. To get away from magic constants, we say that a fraction
at the current level of forward pointers, the search moves p of the nodes with level i pointers also have level i+1 point-
down to the next level. When we can make no more progress ers. (for our original discussion, p = 1/2). Levels are generated
at level 1, we must be immediately in front of the node that randomly by an algorithm equivalent to the one in Figure 5.
contains the desired element (if it is in the list). Levels are generated without reference to the number of ele-
Insertion and Deletion Algorithms ments in the list.
To insert or delete a node, we simply search and splice, as At what level do we start a search? Defining L(n)
shown in Figure 3. Figure 4 gives algorithms for insertion and In a skip list of 16 elements generated with p = 1/2, we might
deletion. A vector update is maintained so that when the happen to have 9 elements of level 1, 3 elements of level 2, 3
search is complete (and we are ready to perform the splice), elements of level 3 and 1 element of level 14 (this would be
update[i] contains a pointer to the rightmost node of level i or very unlikely, but it could happen). How should we handle
higher that is to the left of the location of the inser- this? If we use the standard algorithm and start our search at
tion/deletion. level 14, we will do a lot of useless work.
If an insertion generates a node with a level greater than Where should we start the search? Our analysis suggests
that ideally we would start a search at the level L where we
expect 1/p nodes. This happens when L = log1/p n. Since we
will be referring frequently to this formula, we will use L(n)
Search(list, searchKey)
to denote log1/p n.
x := list→header
There are a number of solutions to the problem of deciding
-- loop invariant: x→key < searchKey
how to handle the case where there is an element with an
for i := list→level downto 1 do
unusually large level in the list.
while x→forward[i]→key < searchKey do
• Don’t worry, be happy. Simply start a search at the highest
x := x→forward[i]
level present in the list. As we will see in our analysis, the
-- x→key < searchKey ≤ x→forward[1] →key
probability that the maximum level in a list of n elements is
x := x→forward[1]
significantly larger than L(n) is very small. Starting a
if x→key = searchKey then return x→value
search at the maximum level in the list does not add more
else return failure
than a small constant to the expected search time. This is
the approach used in the algorithms described in this paper.
FIGURE 2 - Skip list search algorithm

Search path update[i]→forward[i]

6 NIL
25
9
3 7 12 19 21 26
original list, 17 to be inserted

6 NIL
25
9 17
3 7 12 19 21 26
list after insertion, updated pointers in grey
FIGURE 3 - Pictorial description of steps involved in performing an insertion

• Use less than you are given. Although an element may con-
tain room for 14 pointers, we don’t need to use all 14. We
can choose to utilize only L(n) levels. There are a number
of ways to implement this, but they all complicate the algo- Insert(list, searchKey, newValue)
rithms and do not noticeably improve performance, so this local update[1..MaxLevel]
approach is not recommended. x := list→header
• Fix the dice. If we generate a random level that is more than for i := list→level downto 1 do
one greater than the current maximum level in the list, we while x→forward[i]→key < searchKey do
simply use one plus the current maximum level in the list as x := x→forward[i]
the level of the new node. In practice and intuitively, this -- x→key < searchKey ≤ x→forward[i]→key
change seems to work well. However, it totally destroys our update[i] := x
ability to analyze the resulting algorithms, since the level of x := x→forward[1]
a node is no longer completely random. Programmers if x→key = searchKey then x→value := newValue
should probably feel free to implement this, purists should else
avoid it. lvl := randomLevel()
Determining MaxLevel if lvl > list→level then
Since we can safely cap levels at L(n), we should choose for i := list→level + 1 to lvl do
MaxLevel = L(N) (where N is an upper bound on the number update[i] := list→header
of elements in a skip list). If p = 1/2, using MaxLevel = 16 is list→level := lvl
appropriate for data structures containing up to 216 elements. x := makeNode(lvl, searchKey, value)
for i := 1 to level do
ANALYSIS OF SKIP LIST ALGORITHMS x→forward[i] := update[i]→forward[i]
The time required to execute the Search, Delete and Insert update[i]→forward[i] := x
operations is dominated by the time required to search for the
appropriate element. For the Insert and Delete operations, Delete(list, searchKey)
there is an additional cost proportional to the level of the node local update[1..MaxLevel]
being inserted or deleted. The time required to find an element x := list→header
is proportional to the length of the search path, which is de- for i := list→level downto 1 do
termined by the pattern in which elements with different while x→forward[i]→key < searchKey do
levels appear as we traverse the list. x := x→forward[i]
Probabilistic Philosophy update[i] := x
The structure of a skip list is determined only by the number x := x→forward[1]
if x→key = searchKey then
for i := 1 to list→level do
randomLevel() if update[i]→forward[i] ≠ x then break
lvl := 1 update[i]→forward[i] := x→forward[i]
-- random() that returns a random value in [0...1) free(x)
while random() < p and lvl < MaxLevel do while list→level > 1 and
lvl := lvl + 1 list→header→forward[list→level] = NIL do
return lvl list→level := list→level – 1

FIGURE 5 - Algorithm to calculate a random level FIGURE 4 - Skip List insertion and deletion algorithms

elements in the skip list and the results of consulting the ran- Our assumption that the list is infinite is a pessimistic as-
dom number generator. The sequence of operations that pro- sumption. When we bump into the header in our backwards
duced the current skip list does not matter. We assume an ad- climb, we simply climb up it, without performing any left-
versarial user does not have access to the levels of nodes; ward movements. This gives us an upper bound of (L(n)–1)/p
otherwise, he could create situations with worst-case running on the expected length of the path that climbs from level 1 to
times by deleting all nodes that were not level 1. level L(n) in a list of n elements.
The probabilities of poor running times for successive op- We use this analysis go up to level L(n) and use a different
erations on the same data structure are NOT independent; two analysis technique for the rest of the journey. The number of
successive searches for the same element will both take ex- leftward movements remaining is bounded by the number of
actly the same time. More will be said about this later. elements of level L(n) or higher in the entire list, which has an
Analysis of expected search cost expected value of 1/p.
We analyze the search path backwards, travelling up and to We also move upwards from level L(n) to the maximum
the left. Although the levels of nodes in the list are known and level in the list. The probability that the maximum level of the
fixed when the search is performed, we act as if the level of a list is a greater than k is equal to 1–(1–pk)n , which is at most
node is being determined only when it is observed while npk. We can calculate the expected maximum level is at most
backtracking the search path. L(n) + 1/(1–p). Putting our results together, we find
At any particular point in the climb, we are at a situation
Total expected cost to climb out of a list of n elements
similar to situation a in Figure 6 – we are at the ith forward
≤ L(n)/p + 1/(1–p)
pointer of a node x and we have no knowledge about the
levels of nodes to the left of x or about the level of x, other which is O(log n).
than that the level of x must be at least i. Assume the x is not Number of comparisons
the header (the is equivalent to assuming the list extends in- Our result is an analysis of the “length” of the search path.
finitely to the left). If the level of x is equal to i, then we are in The number of comparisons required is one plus the length of
situation b. If the level of x is greater than i, then we are in the search path (a comparison is performed for each position
situation c. The probability that we are in situation c is p. Each in the search path, the “length” of the search path is the num-
time we are in situation c, we climb up a level. Let C(k) = the ber of hops between positions in the search path).
expected cost (i.e, length) of a search path that climbs up k Probabilistic Analysis
levels in an infinite list: It is also possible to analyze the probability distribution of
search costs. The probabilistic analysis is somewhat more
C(0) = 0
complicated (see box). From the probabilistic analysis, we can
C(k) = (1–p) (cost in situation b) + p (cost in situation c)
calculate an upper bound on the probability that the actual cost
By substituting and simplifying, we get: of a search exceeds the expected cost by more than a specified
ratio. Some results of this analysis are shown in Figure 8.
C(k) = (1–p) (1 + C(k)) + p (1 + C(k–1))
Choosing p
C(k) = 1/p + C(k–1)
Table 1 gives the relative times and space requirements for
C(k) = k/p
different values of p. Decreasing p also increases the variabil-

Need to climb k

? levels from here

x
situation a
probability = 1-p probability = p
Need to climb
only k-1 levels
? ? from here

x
situation b x
Still need to climb k situation c
levels from here
FIGURE 6 - Possible situations in backwards traversal of the search path

ity of running times. If 1/p is a power of 2, it will be easy to Normalized search Avg. # of pointers
generate a random level from a stream of random bits (it re- p times (i.e., normalized per node
quires an average of (log2 1/p)/(1–p) random bits to generate a L(n)/p ) (i.e., 1/(1 – p))
random level). Since some of the constant overheads are re- 1/
2 1 2
lated to L(n) (rather than L(n)/p), choosing p = 1/4 (rather than 1/
e 0.94... 1.58...
1/2) slightly improves the constant factors of the speed of the 1/4 1 1.33...
algorithms as well. I suggest that a value of 1/4 be used for p 1/
8 1.33... 1.14...
unless the variability of running times is a primary concern, in 1/16 2 1.07...
which case p should be 1/2.
Sequences of operations
The expected total time for a sequence of operations is equal TABLE 1 – Relative search speed and space requirements,
to the sum of the expected times of each of the operations in depending on the value of p.
the sequence. Thus, the expected time for any sequence of m
Constant factors
searches in a data structure that contains n elements is O(m
Constant factors can make a significant difference in the
log n). However, the pattern of searches affects the probability
practical application of an algorithm. This is particularly true
distribution of the actual time to perform the entire sequence
for sub-linear algorithms. For example, assume that algo-
of operations.
rithms A and B both require O(log n) time to process a query,
If we search for the same item twice in the same data
but that B is twice as fast as A: in the time algorithm A takes to
structure, both searches will take exactly the same amount of
process a query on a data set of size n, algorithm B can pro-
time. Thus the variance of the total time will be four times the
cess a query on a data set of size n 2 .
variance of a single search. If the search times for two ele-
There are two important but qualitatively different contri-
ments are independent, the variance of the total time is equal
butions to the constant factors of an algorithm. First, the in-
to the sum of the variances of the individual searches.
herent complexity of the algorithm places a lower bound on
Searching for the same element over and over again maxi-
any implementation. Self-adjusting trees are continuously re-
mizes the variance.
arranged as searches are performed; this imposes a significant
ALTERNATIVE DATA STRUCTURES overhead on any implementation of self-adjusting trees. Skip
Balanced trees (e.g., AVL trees [Knu73] [Wir76]) and self- list algorithms seem to have very low inherent constant-factor
adjusting trees [ST85] can be used for the same problems as overheads: the inner loop of the deletion algorithm for skip
skip lists. All three techniques have performance bounds of lists compiles to just six instructions on the 68020.
the same order. A choice among these schemes involves sev- Second, if the algorithm is complex, programmers are de-
eral factors: the difficulty of implementing the algorithms, terred from implementing optimizations. For example, bal-
constant factors, type of bound (amortized, probabilistic or anced tree algorithms are normally described using recursive
worst-case) and performance on a non-uniform distribution of insert and delete procedures, since that is the most simple and
queries. intuitive method of describing the algorithms. A recursive in-
Implementation difficulty sert or delete procedure incurs a procedure call overhead. By
For most applications, implementers generally agree skip lists using non-recursive insert and delete procedures, some of this
are significantly easier to implement than either balanced tree overhead can be eliminated. However, the complexity of non-
algorithms or self-adjusting tree algorithms. recursive algorithms for insertion and deletion in a balanced
tree is intimidating and this complexity deters most program-
mers from eliminating recursion in these routines. Skip list al-

1
10-1
10-2
10-3
10-4
p = 1/4, n = 256 Prob.
p = 1/4, n = 4,096 10-5
p = 1/4, n = 65,536 10-6
p = 1/2, n = 256
p = 1/2, n = 4,096 10-7
p = 1/2, n = 65,536 10-8
10-9
1.0 2.0 3.0
Ratio of actual cost to expected cost
FIGURE 8 - This graph shows a plot of an upper bound on the probability of a search taking substantially longer than expected.
The vertical axis show the probability that the length of the search path for a search exceeds the average length by more than the
ratio on the horizontal axis. For example, for p = 1/2 and n = 4096, the probability that the search path will be more than three
times the expected length is less than one in 200 million. This graph was calculated using our probabilistic upper bound.

Implementation Search Time Insertion Time Deletion Time
Skip lists 0.051 msec (1.0) 0.065 msec (1.0) 0.059 msec (1.0)
non-recursive AVL trees 0.046 msec (0.91) 0.10 msec (1.55) 0.085 msec (1.46)
recursive 2–3 trees 0.054 msec (1.05) 0.21 msec (3.2) 0.21 msec (3.65)
Self–adjusting trees:
top-down splaying 0.15 msec (3.0) 0.16 msec (2.5) 0.18 msec (3.1)
bottom-up splaying 0.49 msec (9.6) 0.51 msec (7.8) 0.53 msec (9.0)
Table 2 - Timings of implementations of different algorithms

gorithms are already non-recursive and they are simple ments takes more than 5 times the expected time is about 1 in
enough that programmers are not deterred from performing 1018.
optimizations. Non-uniform query distribution
Table 2 compares the performance of implementations of Self-adjusting trees have the property that they adjust to non-
skip lists and four other techniques. All implementations were uniform query distributions. Since skip lists are faster than
optimized for efficiency. The AVL tree algorithms were writ- self-adjusting trees by a significant constant factor when a
ten by James Macropol of Contel and based on those in uniform query distribution is encountered, self-adjusting trees
[Wir76]. The 2–3 tree algorithms are based on those presented are faster than skip lists only for highly skewed distributions.
in [AHU83]. Several other existing balanced tree packages We could attempt to devise self-adjusting skip lists. However,
were timed and found to be much slower than the results pre- there seems little practical motivation to tamper with the sim-
sented below. The self-adjusting tree algorithms are based on plicity and fast performance of skip lists; in an application
those presented in [ST85]. The times in this table reflect the where highly skewed distributions are expected, either self-
CPU time on a Sun-3/60 to perform an operation in a data adjusting trees or a skip list augmented by a cache may be
structure containing 216 elements with integer keys. The val- preferable [Pug90].
ues in parenthesis show the results relative to the skip list time
The times for insertion and deletion do not include the time ADDITIONAL WORK ON SKIP LISTS
for memory management (e.g, in C programs, calls to malloc I have described a set of algorithms that allow multiple pro-
and free). cessors to concurrently update a skip list in shared memory
Note that skip lists perform more comparisons than other [Pug89a]. This algorithms are much simpler than concurrent
methods (the skip list algorithms presented here require an balanced tree algorithms. They allow an unlimited number of
average of L(n)/p + 1/(1–p) + 1 comparisons). For tests using readers and n busy writers in a skip list of n elements with
real numbers as keys, skip lists were slightly slower than the very little lock contention.
non-recursive AVL tree algorithms and search in a skip list Using skip lists, it is easy to do most (all?) the sorts of op-
was slightly slower than search in a 2–3 tree (insertion and erations you might wish to do with a balanced tree such as use
deletion using the skip list algorithms was still faster than us- search fingers, merge skip lists and allow ranking operations
ing the recursive 2–3 tree algorithms). If comparisons are very (e.g., determine the kth element of a skip list) [Pug89b].
expensive, it is possible to change the algorithms so that we Tom Papadakis, Ian Munro and Patricio Poblette [PMP90]
never compare the search key against the key of a node more have done an exact analysis of the expected search time in a
than once during a search. For p = 1/2, this produces an upper skip list. The upper bound described in this paper is close to
bound on the expected number of comparisons of 7/2 + their exact bound; the techniques they needed to use to derive
3/2 log2 n. This modification is discussed in [Pug89b]. an exact analysis are very complicated and sophisticated.
Type of performance bound Their exact analysis shows that for p = 1/2 and p = 1/4, the
These three classes of algorithm have different kinds of per- upper bound given in this paper on the expected cost of a
formance bounds. Balanced trees have worst-case time search is not more than 2 comparisons more than the exact
bounds, self-adjusting trees have amortized time bounds and expected cost.
skip lists have probabilistic time bounds. With self-adjusting I have adapted idea of probabilistic balancing to some
trees, an individual operation can take O(n) time, but the time other problems arising both in data structures and in incre-
bound always holds over a long sequence of operations. For mental computation [PT88]. We can generate the level of a
skip lists, any operation or sequence of operations can take node based on the result of applying a hash function to the
longer than expected, although the probability of any opera- element (as opposed to using a random number generator).
tion taking significantly longer than expected is negligible. This results in a scheme where for any set S, there is a unique
In certain real-time applications, we must be assured that data structure that represents S and with high probability the
an operation will complete within a certain time bound. For data structure is approximately balanced. If we combine this
such applications, self-adjusting trees may be undesirable, idea with an applicative (i.e., persistent) probabilistically bal-
since they can take significantly longer on an individual oper- anced data structure and a scheme such as hashed-consing
ation than expected (e.g., an individual search can take O(n) [All78] which allows constant-time structural equality tests of
time instead of O(log n) time). For real-time systems, skip applicative data structures, we get a number of interesting
lists may be usable if an adequate safety margin is provided: properties, such as constant-time equality tests for the repre-
the chance that a search in a skip lists containing 1000 ele- sentations of sequences. This scheme also has a number of
applications for incremental computation. Since skip lists are

somewhat awkward to make applicative, a probabilistically [BLLSS86] Bentley, J., F. T. Leighton, M.F. Lepley, D. Stanat and
balanced tree scheme is used. J. M. Steele, A Randomized Data Structure For
Ordered Sets, MIT/LCS Technical Memo 297, May
RELATED WORK 1986.
James Discroll pointed out that R. Sprugnoli suggested a [Knu73] Knuth, D. “Sorting and Searching,” The Art of
method of randomly balancing search trees in 1981 [Spr81]. Computer Programming, Vol. 3, Addison-Wesley
With Sprugnoli’s approach, the state of the data structure is Publishing Company, 1973.
not independent of the sequence of operations which built it. [PMP90] Papadakis, Thomas, Ian Munro and Patricio Poblette,
This makes it much harder or impossible to formally analyze Exact Analysis of Expected Search Cost in Skip Lists,
his algorithms. Sprugnoli gives empirical evidence that his al- Tech Report # ????, Dept. of Computer Science, Univ.
of Waterloo, January 1990.
gorithm has good expected performance, but no theoretical re-
[PT89] Pugh, W. and T. Teitelbaum, “Incremental Computation
sults.
via Function Caching,” Proc. of the Sixteenth
A randomized data structure for ordered sets is described conference on the Principles of Programming
in [BLLSS86]. However, a search using that data structure re- Languages, 1989.
quires O(n1/2) expected time. [Pug89a] Pugh, W., Concurrent Maintenance of Skip Lists, Tech
Cecilia Aragon and Raimund Seidel describe a probabilis- Report TR-CS-2222, Dept. of Computer Science,
tically balanced search trees scheme [AC89]. They discuss University of Maryland, College Park, 1989.
how to adapt their data structure to non-uniform query distri- [Pug89b] Pugh, W., Whatever you might want to do using
butions. Balanced Trees, you can do it faster and more simply
using Skip Lists, Tech Report CS–TR–2286, Dept. of
SOURCE CODE AVAILABILITY Computer Science, University of Maryland, College
Skip list source code libraries for both C and Pascal are avail- Park, July 1989.
able for anonymous ftp from mimsy.umd.edu. [Pug90] Pugh, W. Slow Optimally Balanced Search Strategies
vs. Cached Fast Uniformly Balanced Search Strategies,
CONCLUSIONS to appear in Information Processing Letters.
From a theoretical point of view, there is no need for skip [Spr81] Sprugnoli, R. Randomly Balanced Binary Trees,
lists. Balanced trees can do everything that can be done with Calcolo, V17 (1981), pp 99-117.
skip lists and have good worst-case time bounds (unlike skip [ST85] Sleator, D. and R. Tarjan “Self-Adjusting Binary Search
lists). However, implementing balanced trees is an exacting Trees,” Journal of the ACM, Vol 32, No. 3, July 1985,
task and as a result balanced tree algorithms are rarely imple- pp. 652-666.
mented except as part of a programming assignment in a data [Wir76] Wirth, N. Algorithms + Data Structures = Programs,
structures class. Prentice-Hall, 1976.
Skip lists are a simple data structure that can be used in
place of balanced trees for most applications. Skip lists algo-
rithms are very easy to implement, extend and modify. Skip
lists are about as fast as highly optimized balanced tree algo-
rithms and are substantially faster than casually implemented
balanced tree algorithms.

ACKNOWLEDGEMENTS
Thanks to the referees for their helpful comments. Special
thanks to all those people who supplied enthusiasm and en-
couragement during the years in which I struggled to get this
work published, especially Alan Demers, Tim Teitelbaum and
Doug McIlroy. This work was partially supported by an
AT&T Bell Labs Fellowship and by NSF grant CCR–
8908900.

REFERENCES
[AC89] Aragon, Cecilia and Raimund Seidel, Randomized
Search Trees, Proceedings of the 30th Ann. IEEE Symp
on Foundations of Computer Science, pp 540–545,
October 1989.
[AHU83] Aho, A., Hopcroft, J. and Ullman, J. Data Structures
and Algorithms, Addison-Wesley Publishing Company,
1983.
[All78] John Allen. Anatomy of LISP, McGraw Hill Book
Company, NY, 1978.

PROBABILISTIC ANALYSIS Probabilistic analysis of search cost
In addition to analyzing the expected performance of skip The number of leftward movements we need to make before
lists, we can also analyze the probabilistic performance of skip we move up a level (in an infinite list) has a negative binomial
lists. This will allow us to calculate the probability that an distribution: it is the number of failures (situations b’s) we see
operation takes longer than a specified time. This analysis is before we see the first success (situation c) in a series of
based on the same ideas as our analysis of the expected cost, independent random trials, where the probability of success is
so that analysis should be understood first. p. Using the probabilistic notation introduced above:
A random variable has a fixed but unpredictable value and
Cost to climb one level in an infinite list
a predictable probability distribution and average. If X is a
=prob 1+ NB(1, p).
random variable, Prob{ X = t } denotes the probability that X
equals t and Prob{ X > t } denotes the probability that X is We can sum the costs of climbing each level to get the
greater than t. For example, if X is the number obtained by total cost to climb up to level L(n):
throwing a unbiased die, Prob{ X > 3 } = 1/2.
It is often preferable to find simple upper bounds on values Cost to climb to level L(n) in an infinite list
=prob (L(n) – 1) + NB(L(n) – 1, p).
whose exact value is difficult to calculate. To discuss upper
bounds on random variables, we need to define a partial or- Our assumption that the list is infinite is a pessimistic
dering and equality on the probability distributions of non- assumption:
negative random variables.
Definitions (=prob and ≤prob). Let X and Y be non-negative Cost to climb to level L(n) in a list of n elements
independent random variables (typically, X and Y would ≤prob (L(n) – 1) + NB(L(n) – 1, p).
denote the time to execute algorithms AX and AY). We define Once we have climbed to level L(n), the number of
X ≤ prob Y to be true if and only if for any value t, the leftward movements is bounded by the number of elements of
probability that X exceeds t is less than the probability that Y level L(n) or greater in a list of n elements. The number of
exceeds t. More formally: elements of level L(n) or greater in a list of n elements is a
X =prob Y iff ∀ t, Prob{ X > t } = Prob{ Y > t } and random variable of the form B(n, 1/np).
Let M be a random variable corresponding to the
X ≤ prob Y iff ∀ t, Prob{ X > t } ≤ Prob{ Y > t }. s
maximum level in a list of n elements. The probability that the
For example, the graph in Figure 7shows the probability level of a node is greater than k is pk, so Prob{ M > k } = 1–
distribution of three random variables X, Y and Z. Since the (1–pk)n < npk. Since npk = pk–L(n) and Prob{ NB(1, 1–p) + 1 >
probability distribution curve for X is completely under the i} = pi, we get an probabilistic upper bound of M ≤prob L(n) +
curves for Y and Z, X ≤prob Y and X ≤prob Z. Since the NB(1, 1 – p) + 1. Note that the average of L(n) + NB(1, 1 – p)
probability curves for Y and Z intersect, neither Y ≤prob Z nor + 1 is L(n) + 1/(1–p).
Z ≤prob Y. Since the expected value of a random variable X is This gives a probabilistic upper bound on the cost once we
simply the area under the curve Prob{ X > t }, if X ≤prob Y have reached level L(n) of B(n, 1/np) + (L(n) + NB(1, 1 – p) +
then the average of X is less than or equal to the average of Y. 1) – L(n). Combining our results to get a probabilistic upper
We make use of two probability distributions: bound on the total length of the search path (i.e., cost of the
Definition (binomial distributions — B(t, p)). Let t be a entire search):
non-negative integer and p be a probability. The term B(t, p) total cost to climb out of a list of n elements
denotes a random variable equal to the number of successes ≤prob (L(n) – 1) + NB(L(n) – 1, p) + B(n, 1/np)
seen in a series of t independent random trials where the + NB(1, 1 – p) + 1
probability of a success in a trial is p. The average and
variance of B(t, p) are tp and tp(1 – p) respectively. s The expected value of our upper bound is equal to
Definition (negative binomial distributions — NB(s, p)). Let s
(L(n) – 1) + (L(n) – 1)(1 – p)/p + 1/p + p/(1–p) + 1
be a non-negative integer and p be a probability. The term
= L(n)/p + 1/(1–p),
NB(s, p) denotes a random variable equal to the number of
failures seen before the sth success in a series of random which is the same as our previously calculated upper bound
independent trials where the probability of a success in a trial on the expected cost of a search. The variance of our upper
is p. The average and variance of NB(s, p) are s(1–p)/p and bound is
s(1–p)/p2 respectively. s
(L(n) – 1)(1–p)/p2 + (1 – 1/np)/p + p/(1–p)2
< (1–p)L(n)/p2 + p/(1–p)2 + (2p–1)/p2 .
1
Prob{ X > t } Figure 8 show a plot of an upper bound on the probability
Prob{ Y > t } of an actual search taking substantially longer than average,
Prob{ Z > t } based on our probabilistic upper bound.
Prob

0
t
FIGURE 7 – Plots of three probability distributions

Skip list

More Related Content

Similar to Skip list (20)

More from Institute of Computing Technology, Chinese Academy of Sciences (8)

Recently uploaded (20)

Skip list