LaTeX floats terminology

Float classes
Each float in LaTeX belongs to a class. By default, LaTeX knows about two
classes, viz., figureand table. Further classes can be added by a document class or by
packages. The class a float belongs to influences certain aspect of the float's position, such
as its default placement specification (if not overridden by the float itself).
One important aspect of the float placement algorithm is that LaTeX never violates the order
of placement within a class of floats. E.g., if you have figure1, table1, figure2 in a document,
then figure1 will always be placed before figure2. However, table1 (belonging to a different
float class) will be placed independently and hence can appear before, after, or between the
figures.
Float areas
LaTeX knows about 2 float areas within a column where it can place floats: the top area and
the bottom area of the column. In two-column layout, it also knows about a top area
spanning the two columns. There is no bottom area for page-wide floats in two-column
mode

In addition, LaTeX can make float columns and float pages, i.e, columns and pages which
contain only floats.

Finally LaTeX can place floats inline into the text (if so directed).

Float placement specifiers

To direct a float to be placed into one of these areas, a float placement specifier has to be
provided as an optional argument to the float. If no such optional argument is given then a
default placement specifier is used (which depends on the float class as mentioned above).
A float placement specifier can consist of the following characters in any order:
 ! indicates that some restrictions should be ignored (discussed later)
 h indicates that the float is allowed to be placed inline
 t indicates that the float is allowed to go into a top area
 b indicates that the float is allowed to go into a bottom area
 p indicates the the float is allowed to go on a float page or column area
The order in which these characters are stated does not influence how the algorithm tries to
place the float (e.g., [ht] or [th] will make no difference)! This is one of the common
misunderstandings, for instance when people think that bt means that the bottom area
should be tried first.
However, if a letter is not present then the corresponding area will not be tried at all.

Float algorithm parameters

There are about 20 parameters that influence the placement. Basically they define

 how many floats can go into a certain area,

 how big an area can become,
  how much text there has to be on a page (in other words, how much the top and bottom areas
can occupy) and
 how much space will be inserted
 between consecutive floats in an area and
 between the area and the text above or below.
Float reference point
The point where the float is placed in the source document affects the placement of the float
in the output because it determines when LaTeX sees the float for the first time. It's
important to understand that if a float is placed in the middle of a paragraph, this reference
point is the next line break, or page break, in the paragraph that follows the actual
placement in the source.

Basic behavioral rules Emptying the holding queue at the page boundary
After a page has been finished, LaTeX looks at the holding queue and attempts to empty it out as best
as possible. For this it will first try to build float pages.

Any floats participating in a float page (or column) must have a p as a float specifier in its float
placement specification. If not, the float cannot go on a float page and, on top of that, the absence of
the p specifier will prevent any following floats of the same class to go there too!
If the float can go there, it will be marked for inclusion on the float page, but the processor may still
abort the attempt if the float page will not get filled "enough" (depending on the parameter settings
for float pages). Only at the very end of the document, or when a \clearpage has been issued, are
these restrictions lifted, and a float will then be placed on a float page even if it has no p and would
be the only float on that page.
Creation of float pages continues until the algorithm has no further floats to place or if it fails to
produce a float page. In the latter case, all floats that have not been placed so far are then considered
for inclusion in the top and bottom areas of the next page (or column).

The process there is the same as the one described above, except that

 the h specifier has no longer any meaning (as we are, by now, far away from the original
"here") and is therefore ignored
 and the floats are this time coming not from the source document but one after the other from
the holding queue.
Any float that couldn't been placed then is put back to the holding queue, so that when LaTeX is
ready to look at further textual input from the document the holding queue may already contain
floats. A consequence of this is that a float encountered in the document may immediately get
deferred just because an earlier float of the same float class is already on hold.

Details on the parameters that restrict/influence the placement

There are four counters that control how many floats can go into areas:

 totalnumber (default 3) is the maximum number of floats on a text (!) page

 topnumber (default 2) is the maximum number of floats in the top area
  bottomnumber (default 1) is the maximum number of floats in the bottom area
 dbltopnumber (default 2) is the maximum number of full sized floats in two-column mode
going above the text columns.
The size of the areas are controlled through parameters (changed with \renewcommand) that define
the maximum (or minimum) size of the area, expressed as a fraction of the page height:
 \topfraction (default 0.7) maximum size of the top area
 \bottomfraction (default 0.3) maximum size of the bottom area
 \dbltopfraction (default 0.7) maximum size of the top area for double-column floats
 \textfraction (default 0.2) minimum size of the text area, i.e., the area that must not be
occupied by floats
The space that separates floats within an area, as well as between float areas and text areas, is defined
through the following parameters (all of which are rubber lengths, i.e., can contain some stretch or
shrink components). Their defaults depend on the document font size and change when class options
like 11pt or 12ptare used. We only show the 10pt defaults:
 \floatsep (default 12pt plus 2pt minus 2pt) the separation between floats in top or bottom
areas
 \dblfloatsep (default 12pt plus 2pt minus 2pt) the separation between double-column floats
on two column pages
 \textfloatsep (default 20pt plus 2pt minus 4pt) the separation between top or bottom area
and the text area
 \dbltextfloatsep (default 20pt plus 2pt minus 4pt) the analog of \textfloatsep for two-
column floats
For inline floats (that have been placed "here") the separation to the surrounding text is controlled by

 \intextsep (default 12pt plus 2pt minus 2pt)

In case of float pages or float columns (i.e., a page or a column of a page containing only floats)
parameters like \topfraction etc. do not apply. Instead the creation of them is controlled through
 \floatpagefraction (default 0.5) minimum part of the page (or column) that need to be
occupied by floats to be allowed to form a float page (or column)
Consequences of the algorithm
A float may appear in the the document earlier than its placement in the source
The placement of the float environment in the source determines the earliest point where it can
appear in the final document. It may move visually backward to some degree as it may be placed in
the top area on the current page. It can, however, not end up on an earlier page as the surrounding
text due to the fact that LaTeX does no backtracking an the earlier pages have already been typeset.

Thus normally a float is placed in the source near its first call-out (i.e., text like "see figure 5")
because this will ensure that the float appears either on the same page as this text or on a later page.
However, in some situations you may want to place a float on the preceding page (if that page is still
visible from the call-out). This is only possible by moving the float in the source.

Double-column floats are always deferred first

When LaTeX encounters a page-wide float environment (indicated by a * at the end of the
environment name, e.g., figure*) in two column-mode, it immediately moves it to the deferred
queue. The reason for this behavior again lies in the "greedy" behavior of its algorithm: if LaTeX is
currently assembling the second column of that page, the first column has already been assembled
and stored away; recall that because LaTeX does not back-track there is no way to fit the float on the
current page. To keep the algorithm simple, it does the same even if working on the first column
(where it could in theory do better even without back-tracking).
Thus, in order to place such a float onto the current page, one has to manually move it to an earlier
place in the source -- before the start of the current page. If this is done, obviously any further change
in the document could make this adjustment obsolete; hence, such adjustments are best done (if at
all) only at the very last stage of document production --- when all material has been written and the
focus is on fine-tuning the visual appearance.

There is no bottom float area for double-column floats

This isn't so much a consequence of the algorithm but rather a fact about it. For double-column floats
the only possible placements are the top area or a float page. Thus if somebody adds an h or a b float
placement specifier to such a float it simply gets ignored. So {figure*}[b] implies that this float
will not get typeset until either a \clearpage is encountered or the end of the document is reached.
All float parameters (normally) restrict the placement possibilities
This may be obvious but it is worth repeating: any float parameter defines a restriction on LaTeX's
ability to place the floats. There is always a way to set a parameter in such a way that it does not
affect the placement at all. Unfortunately, in doing so one invites rather poorly looking placements.

By default LaTeX has settings that are fairly liberal. For example, for a float page to be accepted the
float(s) must occupy at least half of the available page. Expressed differently, this means that such a
page is allowed to be half empty (which is certainly not the best possible placement in most cases).

What often happens is that users try to improve such settings and then get surprised when suddenly
all floats pile up at the end of the document. To stay with this example: if one changes the
parameter \floatpagefraction to require, say, 0.8 of the float page, a float that occupies
about 0.75 of the page will not be allowed to form a float page on its own. Thus, if there isn't another
float that could be added and actually fits in the remaining space, the float will get deferred and with
it all other floats of the same class. But, even worse, this specific float is too big to go into the next
top area as well because there the default maximum permissible area is 0.7. As a result all your floats
stay deferred until the next \clearpage.
For this reason it is best not to meddle with the parameters while writing a document or at least not to
do so in a way that makes it more difficult for the algorithm to place a float close to its call-out. For
proof-reading it is far more important to have a figure next to the place it is referenced then to avoid
half-empty pages. Possibilities for fine-tuning an otherwise finished document are discussed below.

Another conclusion to draw here is that there are dependencies between some of the float parameters;
it is important to take these dependencies into account when changing their values.

"Here" really just means "here if it fits"

... and often it doesn't fit. This is somewhat surprising for many people, but the way the algorithm has
been designed the hspecifier is not an unconditional command. If an unconditional command is
needed, extension packages such the float package offer H as an alternative specifier that really
means "here" (and starts a new page first if necessary).
Float specifiers do not define an order of preference
As mentioned above, the algorithm tries to place floats into available float areas in a well-defined
order that is hard-wired into the algorithm: "here", "top", "bottom" and -- on page boundaries -- first
"page" and only if that is no longer possible "top" followed by "bottom" for the next page.

Thus specifying [bt] does not mean try bottom first and only then top. It simply means allow this
float to go into top or bottom area (but not onto a float page) just like [tb] would.
Relation of floats and footnotes
This is not exactly a consequence of the algorithm but one of its implementation: Whenever LaTeX
tries to decide on a placement for a float (or a \marginpar!) it has to trigger the output routine to do
this. And as part of this process any footnotes on the page are removed from their current place in the
galley and are collected together in the \footins box. After placing the float (or deferring it) LaTeX
then returns the page material to the galley, but because of the Output Routine behavior the galley
has now changed: LaTeX has to put the footnotes somewhere but all in one place. What it does is it
reinserts the footnotes (the \footnotetext to be precise) at the end of the galley. There are some
good reasons for doing this, one of which is that LaTeX expects that all of the returned material still
fits on the current page.
However, if for some reason a page is finally taken at an earlier point then the footnotes will show up
on the wrong page or column. This is a fairly unlikely scenario but if it happens check if there is a
float near the chosen page break and either move the float or guide the algorithm by using explicit
page breaks. and example can be found in this question.
In fact that particular case is worth highlighting: Do not place a float directly after a heading, unless
it is a heading that always starts a page. The reason is that headings normally form very large objects
(as they prevent page breaks directly after it). However placing a float in the middle of this means
that the Output Routine gets triggered before LaTeX makes its decision where to break and any
footnotes get moved into the wrong place
Documentation of the algorithm
As requested here is some information on existing documentation. The algorithm and its
implementation is documented in the file ltoutput.dtx as part of the LaTeX kernel source. This can
be typeset standalone or as part of the whole kernel (i.e., by typesetting source2e.tex --- ignore the
checksum error, sorry).
This documentation is an interesting historical artifact. Parts of it show semi-formatted pseudo code
which dates back to LaTeX2.09; in other words it is from the original documentation by Leslie
Lamport. The actual code is documented using doc style and in parts is more or less properly
documented (from scratch) and dates back to 1994 or thereabout when Chris Rowley and myself
adjusted and extended the original algorithm for LaTeX2e. It also fairly openly documents the
various issues with the algorithm and/or its implementation --- in many cases we didn't dare to alter it
because of the many dependencies and, of course, because of the danger to screw up too many
existing documents that implicitly rely on the current behavior for good or worse. Near the end you'll
find a list of comments compiled back then on the algorithm, but there are also comments, questions,
and tasks (? :-) sprinkled throughout the documentation of the code.

One interesting aspect of this file (that I forgot all about) is that it contains all code necessary to trace
the behavior of the algorithm in real life. Unfortunately, I never made this officially available or so it
seems. It would probably need a good amount of cleanup and better formatting of the tracing output
it produces to be usable for the general public. But even in its current form it does give some
interesting insight in the behavior of the algorithm and how certain decisions come about.
So if somebody wants to play with it or wants to trace some strange float placements, then all that is
necessary to do (fingers crossed) is to make a short file fltrace.ins with the following content:
\input docstrip

\generateFile{fltrace.sty}{t}{%
\from{ltoutput.dtx}{fltrace,trace}
}

\endbatchfile
Run through LaTeX this will produce the style file fltrace.sty. You can then use this in your
documents via
\usepackage{fltrace}
\tracefloats % start float tracing
The command \tracefloatvals displays the current state of several float parameters and
with \notrace tracing is turned off again. As already mentioned this is not an official package, but it
might be useful in one or the other situation or purely out of interest after having studied the
documentation.