icl/doc/implementation.qbk

[/
    Copyright (c) 2008-2009 Joachim Faulhaber

    Distributed under the Boost Software License, Version 1.0.
    (See accompanying file LICENSE_1_0.txt or copy at
    http://www.boost.org/LICENSE_1_0.txt)
]

[section Implementation]

The [link boost_icl.interface previous section] gave an overview over the interface
of the *icl* outlining
[link boost_icl.interface.class_templates class templates],
[link boost_icl.interface.associated_types associated types]
and polymorphic
[link boost_icl.interface.function_synopsis functions and operators].
In preparation to the
[link boost_icl.function_reference next section],
that describes the *icl's* polymorphic functions in
more detail together with ['*complexity characteristics*],
this section summarizes some general information on
implementation and performance.

[h5 STL based implementation]

The *implementation* of the *icl's* containers is based on
[*std::set] and [*std::map]. So the underlying data structure of
interval containers is a red black tree of intervals or
interval value pairs.
The element containers __icl_set__ and __icl_map__ are wrapper
classes of `std::set` and `std::map`.
Interval containers are then using __icl_sets__ of intervals
or __icl_maps__ of interval value pairs as implementing
containers.
So all the ['*complexity characteristics*] of icl containers
are based on and limited by the ['*red-black tree implementation*]
of the underlying std::AssociativeContainers.


[section Iterative size]

Throughout the documentation on complexity,
big /O/ expressions like __On__ or __Omlgn__ refer to container sizes
/n/ and /m/. In this documentation these sizes ['*do not*] denote
to the familiar `size` function that returns
the ['*number of elements*] of a container. Because for an interval container
``
interval_set<int> mono;
mono += interval<int>::closed(1,5); // {[1 ... 5]}
mono.size()           == 5;         // true, 5 elements
mono.interval_count() == 1;         // true, only one interval
``

it's size and the number of contained intervals is usually different.
To refer uniformly to a /size/ that matters for iteration, which is
the decisive kind of size concerning algorithmic behavior there is a function
``
bool T::iterative_size()const; // Number of entities that can be iterated over.
``
for all element and interval containers of the icl. So for
complexity statements throughout the icl's documentation
the sizes will be `iterative_sizes` and big /O/ expressions like
__Omlgn__ will refer to sizes
``
n = y.iterative_size();
m = x.iterative_size();
``
for containers `y` and `x`.
Note that ``iterative_size`` refers to the primary entities,
that we can iterate over. For interval containers these
are intervals or segments. ``Itervative_size`` never refers
to element iteration for interval containers.

[endsect][/ Iterative size]


[section Complexity]

[h4 Complexity of element containers]

Since ['element containers] __icl_set__ and __icl_map__ are only extensions of
stl::set and stl::map, their complexity characteristics are
accordingly. So their major operations insertion (addition),
deletion and search are all using logarithmic time.

[h4 Complexity of interval containers]

The operations on ['interval containers] behave differently
due to the fact that intervals unlike elements can overlap
any number of other intervals in a container. As long as
intervals are relatively small or just singleton, interval
containers behave like containers of elements.
For large intervals however time consumption of operations
on interval containers may be worse, because most or all
intervals of a container may have to be visited.
As an example, time complexity of __biLAddition__ on
interval containers is briefly discussed.

More information on ['*complexity characteristics*]
of *icl's* functions is contained in section
[link boost_icl.function_reference Function Reference]


[h5 Time Complexity of Addition]

The next table
gives the time complexities for the overloaded
`operator +=` on interval containers.
The instance types of `T` are given as column headers.
Instances of type parameter `P` are denoted in
the second column.
The third column contains the specific kind of complexity statement.
If column three is empty ['*worst case*] complexity is given
in the related row.


[table Time Complexity of Addition:
[[]                                             [`P`]               [][interval\nset][separate\ninterval\nset][split\ninterval\nset][interval\nmap][split\ninterval\nmap]]
[/ 1operation                                   2granul.            3case        4itvset       5se_itvset    6sp_itvset    7itv_map      8sp_itvmap]
[[`T& operator +=(T& object, const P& addend)`] [`T::element_type`] []           [__Olgn__]    [__Olgn__]    [__Olgn__]    [__Olgn__]    [__Olgn__]    ]
[[]                                             [`T::segment_type`] [best case]  [__Olgn__]    [__Olgn__]    [__Olgn__]    [__Olgn__]    [__Olgn__]    ]
[[]                                             []                  [worst case] [__On__]      [__On__]      [__On__]      [__On__]      [__On__]      ]
[[]                                             []                  [amortized]  [__Olgn__]    [__Olgn__]    []            []            []            ]
[[]                                             [`interval_sets`]   []           [__Omlgnpm__] [__Omlgnpm__] [__Omlgnpm__] [           ] [           ] ]
[[]                                             [`interval_maps`]   []           [           ] [           ] [           ] [__Omlgnpm__] [__Omlgnpm__] ]
]

Adding an /element/ or
/element value pair/ is always done in /logarithmic time/,
where /n/ is the number of intervals in the interval container.
The same row of complexities applies to the insertion
of a /segment/ (an interval or an interval value pair)
in the ['*best case*], where the inserted segment does overlap
with only a ['*small*] number of intervals in the container.

In the ['*worst case*], where the inserted segment overlaps with
all intervals in the container, the algorithms
iterate over all the overlapped segments.
Using inplace manipulations of segments and
hinted inserts, it is possible to perform
all necessary operations on each iteration step
in /constant time/.
This results in ['*linear worst case time*] complexity for
segment addition for all interval containers.

After performing
a worst case addition
for an __itv_set__ or a __sep_itv_sets__
adding an interval that overlaps /n/
intervals, we
need /n/ non overlapping additions of
/logarithmic time/ before we can launch
another __On__ worst case addition.
So we have only a ['*logarithmic amortized
time*] for the addition of an interval or interval value pair.

For the addition of ['*interval containers*] complexity is __Omlgnpm__.
So for the ['*worst case*], where the container sizes /n/ and /m/
are equal and both containers cover the same ranges,
the complexity of container addition is ['*loglinear*].
For other cases, that occur frequently in real world applications
performance can be much better.
If the added container `operand` is much smaller than
`object` and the intervals in `operand` are relatively
small, performance can be /logarithmic/.
If /m/ is small compared with /n/ and intervals in `operand`
are large, performance tends to be /linear/.

[endsect][/ Complexity]

[section Inplace and infix operators]

For the major operations /addition, subtraction, intersection/
of *icl* containers and for /symmetric difference/
inplace `operator`s `+= |=, -=, &=` and `^=`
are provided.

For every ['*inplace*] operator
``
T& operator o= (T& object, const P& operand)
``
the *icl* provides corresponding ['*infix*] operators.
``
T operator o (T object, const P& operand){ return object o= operand; }
T operator o (const P& operand, T object){ return object o= operand; }
``

From this implementation of the infix `operator o`
the compiler will hopefully use return value optimization
[@http://en.wikipedia.org/wiki/Return_value_optimization (RVO)]
creating no temporary object and performing one copy of
the first argument `object`.

[caution
Compared to the /inplace/ `operator o=` every use of an
/infix/ `operator o` requires ['*one extra copy*]
of the first argument `object` that passes a container.]

Use infix operators only, if

* efficiency is not crucial, e.g. the containers copied are small.
* a concise and short notation is more important than efficiency in your context.
* you need the result of operator `o=` as a copy anyway.

[h5 Time Complexity of infix `operators o`]

The time complexity of all infix operators of the *icl*
is biased by the extra copy of the `object` argument.
So all infix `operators o` are at least
/linear/ in `n = object.iterative_size()`.
Taking this into account, the complexities of all
infix operators can be determined
from the corresponding inplace `operators o=` they depend
on.

[endsect][/ Inplace and infix operators]


[endsect][/ Implementation]