====== Matlab - variables, datatypes, and indexing ====== This page covers the use of variables in Matlab, the basic (built-in) datatypes, together with some information on how they can be used, how they can be converted, and how they relate to one another. The last part covers the different indexing syntax mechanisms and some pitfalls. ===== Variables ===== A variable can be thought of as a storage container that has the following properties: * links a name (identifier) to a value (or list of values) * is available in a workspace, i.e. when a function is called, variables available in the calling workspace are hidden (different function code files can use the same variable names without conflict), unless they are **''global''** variables * can be assigned a (new) value (or list of values) using the **''=''** sign (assignment operator) * has a specific datatype, which can change during the course of a program (or command line session) * changing datatypes is considered bad coding practice and should be avoided:% define x as a number x = 1; % re-define x as a string: no error! x = 'string'; * is of arbitrary size (which means it can also be empty), which can also change (older versions would allow up to 63 dimensions, but this limit no longer exists) * that means that a variable containing a single number is of the same type as a variable containing several numbers (of the same type) * does not require declaration (like in C/C++) but can be created at any time (on the command line or at any point in a program, i.e. M-file) * this unfortunately makes it sometimes difficult to find out what an identifier stands for (function or variable? what datatype and content?) * stores this value until it is reassigned a new value or the variable is "''clear''-ed" from the workspace * allows indexing operations to access sub-parts of a list of values (both reading from and writing into parts of variables) * to read-access a sub-portion of an array, the index expression has to be provided within parentheses on the right-hand side (e.g. **''part = fullarray(portion);''**) * on the left-hand side, a part of an array can be replaced with new data (e.g. **''fullarray(portion) = newvalues;''**) * in that case, **''newvalues''** must either be a single number (all indices addressed by **''portion''** will be set to the same number) or must match in size * if the variable is smaller than indicated by the index expression, Matlab will attempt to grow the variable accordingly:% define a 2x3 array a = [10, 20, 30; 40, 50, 60]; % assign the value 100 to the second through 4th row and the 3rd through 4th column a(2:4, 3:4) = 100; % a is now a 4x4 array! * can be used in expressions (e.g. in computations, function calls, to index another variable, or to form new, compound variables) * is available for storing the data contained therein to a file on disk and can be loaded from disk as well Please note that at the end of a function (including when the keyword **''return''** is reached while executing a function) all variables that are not "returned" or marked as **''persistent''** are cleared from the workspace and memory. ==== Variable identifiers ==== There are a few rules applying to identifiers: * a valid identifier must contain only letters (lower and upper case), numbers, and underscores, but no symbols or blanks * it must begin with a letter * some valid variable names with assignments: v = 1; VAR12 = 12; A_Really_Long_Name = 'long name'; * **keywords can not be used as identifier names**, which excludes the following words from being identifier/variable names: **''break, case, catch, classdef, continue, else, elseif, end, for, function, global, if, otherwise, parfor, persistent, return, spmd, switch, try, while''** * if a variable is given the same name as an existing function, the identifier then only refers to the variable in this workspace (see also [[Matlab - precedence|precedence rules]]):% defining a new array newarray = [1, 2, 3, 4]; % computing the sum sumarray = sum(newarray); % redefining sum sum = 5; % then this leads to an error... notthesum = sum(newarray); % index exceeds dimensions!! ===== Datatypes ===== Datatypes can be, generally, divided into 5 major groups: * numeric datatypes (incl. a logical datatype for **''true''**/**''false''** values) * text (character/string) datatype * compound datatypes (to store values of different types in one variable) * function handles * user-defined datatypes/objects ==== Numeric datatypes ==== By default, a variable in Matlab that is storing a numeric value (or a list/array of numbers) has the datatype "double". So, in a simple assignment of a number (or array) to a variable (such as **''a = 1;''** or **''b = [2, 3, 4];''**), the datatype would be double, regardless of whether the number is integer or not! While this requires more memory (for large arrays of numbers), at least the user (or code writer) doesn't have to worry about datatype conversions, etc. Unless you have very specific needs (e.g. lower memory usage or increased speed for specific operations), it is recommended to use the default datatype. Numeric variables are defined by simply assigning the output of a function that returns a number to a variable or by setting the value(s) manually. Here is a list of all "numeric" datatypes: * **''double''** (default type for all numbers, supports decimal numbers/fractions): each value being stored requires 8 byte (= 64 bits) of memory, 1 bit for the sign, 11 bits for the exponent, and 52 bits for a base-2 fraction (see [[http://en.wikipedia.org/wiki/Double-precision_floating-point_format|Double precision floating point format at wikipedia]]). Special "configurations" are used to store the values **''Inf''**, **''-Inf''**, and **''NaN''**. * **''single''**: each value being stored requires 4 byte (= 32 bits) of memory; in short, the datatype has similar properties compared to double, just less "precision" (and exponent range) * **''int64''**: a 64-bit (8-byte) integer datatype (signed); lowest value is -2^63, highest value is 2^63 - 1 * **''uint64''**: a 64-bit (8-byte) integer datatype (unsigned); lowest value 0, highest value is 2^64 - 1 * **''int32''**: a 32-bit (4-byte) integer datatype (signed); lowest value is -2^31, highest value is 2^31 - 1 * **''uint32''**: a 32-bit (4-byte) integer datatype (unsigned); lowest value is 0, highest value is 2^32 - 1 * **''int16''**: a 16-bit (2-byte) integer datatype (signed); lowest value is -32768, highest value is 32767 * **''uint16''**: a 16-bit (2-byte) integer datatype (unsigned); lowest value is 0, highest value is 65535 * **''int8''**: an 8-bit (1-byte) integer datatype (signed); lowest value is -128, highest value is 127 * **''uint8''**: an 8-bit (1-byte) integer datatype (signed); lowest value is -128, highest value is 127 A special case is the **''logical''** datatype. It can only store two values: **''false''** or ''**true**''. If converted to any of the other numeric datatypes, **''false''** is converted to **''0''** and **''true''** is converted to **''1''**. Please note that instead of being keywords (such as in C/C++), datatypes are not "reserved" words, but obviously the names of datatypes should not be used as variable identifiers. The reason is that those names are also the names of the functions used to convert the numeric datatypes into one another! For instance the code snippet dvar = [7, -10, 13]; ivar = int32(dvar); converts the **''double''** variable dvar (default numeric datatype!) into a variable of type **''int32''**. If the target datatype cannot hold the value(s), for instance because the value range is too small, the value will be truncated (in precision and range) to fit the new datatype. ==== Character datatype ==== Given that Matlab variables can be arrays of arbitrary size, single characters, as well as "strings" (a series of characters, such as in a word or sentence) and also lists of strings (two-dimensional field of characters) all are stored with the same basic datatype: **''char''**. Here are some examples defining variables of type **''char''**: letter = 'x'; start = 'The letter is'; sentence = [start, ' ', letter, '.']; The resulting variable then contains the string "''The letter is x.''". Importantly, the underlying storage is yet a numeric datatype:% the difference between to characters 'd' - 'a' % is their distance in the alphabet, in this case 3! And that also means that a variable can be converted from **''char''** to **''double''** (or any other compatible numeric datatype) and back. This can be useful to store large quantities of text and also to perform arithmetic operations on characters (for instance to test if all elements of a string are letters). ==== Compound datatypes ==== In many situations it is necessary to store data of different types (e.g. a name/string together with a number, such as age) in a "dataset", which still should be accessible via a single variable. For this purpose, Matlab provides built-in compound datatypes. These compound datatypes are further sub-divided into two types: one where elements are (mainly) addressed by a numeric (or equivalent) indices, and one where elements are accessed by a field name in a tree-like structure (same rules as identifiers, but also keywords can be used as field names, given the syntax). === Cell datatype === The **''cell''** datatype allows to access individual elements (as well as groups of elements), called "cells", with numeric indexing. This is mostly useful for storing tabular data (or data where several "columns" should be accessible at once), and numeric only data is too limited (e.g. in cases where text elements cannot be reasonably matched to numeric values a priori). To define a cell array as well as to address the **content** of a cell, Matlab uses the "curly braces" symbols: **''{''** and **''}''**:% define a 1x2 cell array with a name and an age name_and_age = {'John Doe', 41}; % to access just the name we index the first cell with {1} name = name_and_age{1}; % and for the age the second cell age = name_and_age{2}; **Please be aware that a cell array can, naturally, also be indexed with the parentheses syntax. However, in that case the returned value will be of type ''cell''. In fact, every index expression on a variable of a built-in datatype using the parentheses syntax always returns a value (or values) of the same datatype!** Put differently, if you imagine a shelf with 5 jars on it. The entire shelf then represents a cell array (by the name of **''shelf''**). The expression **''shelf(3)''** will then return the third jar of the shelf. On the other hand, the expression **''shelf{3}''** returns the **content** of the third jar! Also, please note that if you use a variable (numeric, char, or compound!) when creating a new compound variable or setting one or several elements of a compound variable, the values in the compound variable will be **copies** of the content of the original variable (or rather, if the variable is altered, a copy is created). For instance:% assign a value to variables n and s n = [12, 14, 10]; s = 'size'; % store variables in compound variable c = {s, n}; % re-assign index 2 of n variable n(2) = 18; % and then c still contains [12, 14, 10] in its second cell element! === Struct datatype === The other compound datatype in Matlab is the **''struct''** type. This allows to store multiple values of different types accessible via names. Syntax-wise, the variable name (of the compound variable) is followed by a period (dot character, **''.''**) and a field name. If the field name itself is stored in a (char) variable, the syntax is **''struct_variable.(field_name)''**. The obvious advantage is that code usually becomes less cryptic, given that instead of using syntax such as **''compound_var{3} = some_function(some_value);''** you would write **''compound_var.property = some_function(some_value);''** whereas ''property'' can be a more meaningful term (such as ''name'' or ''duration'') instead of using a numeric index (as with the cell syntax). Please note that variables of type struct still can be of arbitrary size! This means that a list of structures (e.g. 5 people's names and their ages) could be stored in one variable. The syntax to access the 3rd person's name and age would then be:% read out one persons name and age this_name = people(3).name; this_age = people(3).age; In turn this means that if the index expression is omitted, the **''struct_var.fieldname''** syntax produces a list of values (possibly with different datatypes). The only useful way to capture such a list is by creating an ad-hoc cell array:% get the names of all people all_names = {people.name}; === Multi-layered compounding === Each cell and struct field can contain any of Matlab's supported datatypes, including cell and struct arrays! This means that elaborate structures (or cells) can be created:% assign a part of a subfield's value main_tree.leaf(3).subfield{4}(11:20) = 1; If you see such a piece of code this can be translated into * main_tree is of datatype struct (and must be of size 1x1, i.e. a scalar struct, to be valid) * one of the field names of main_tree is called leaf, is also of datatype struct with at least field name subfield, and it presumably has at least 3 elements (see comment below) * the subfield (of the third leaf!) is of type cell with at least 4 elements * the 4th element of the subfield is a numeric array * numeric indices 11 through 20 (of this numeric array) are set to 1 Please be aware that this line of code is also valid if main_tree is not yet defined! In that case, the **''leaf''** field will be initialized as a 1x3 struct with field name subfield, but the first two "leaf" elements will have an empty subfield. Only the third leaf's subfield will be assigned a value, which will be a 1x4 cell array, of which only the 4th element will have a value. And that will be a 1x20 double array, of which the first 10 values are 0! As you can see, while Matlab's "auto-declaration" of variables can be quite useful, it sometimes makes code difficult to understand (given that no explicit declaration is ever given that pre-determines the shape or type of content for compound variables). ==== Function handles ==== The **''function_handle''** type is an advanced datatype which is used mainly in three cases: * the type of operation that is to be applied to a variable (or expression) is not fully determinable before code is run, in which case a function handle can be used to call a variable function (instead of using a syntax construction that selects from all possible options, which still could be insufficient if a function can be user defined) * an argument in a function call itself is "an operation" to be applied to certain values * a function is only available in a certain context (private function or sub-function in an M-file), in which case a function handle can be created and returned, allowing functions outside the original scope to use this function after all Given the fact that this feature is fairly advanced, I won't be giving any in-depth examples at this point. Just be aware that variables can also be of type **''function_handle''**. ==== User-defined datatypes ==== Matlab allows users to add functionality by creating text files with the **''.m''** extension. While this allows to create new functions that can be applied to values, most modern languages also have object-oriented design patterns. Among those are method overloading, inheritance, and protected storage of properties. For this purpose, Matlab allows to define new datatypes (classes) by adding folders with a leading **''@''** (at) symbol, and placing an M-file into this folder with the same name (without the **''@''** sign). Again, this is fairly advanced coding and will not be discussed in-depth at this point. There are, however, a few important aspects I want to mention: * the internal representation of objects is supposed to be of type struct * indexing operations (incl. the struct syntax of **''variable.fieldname''**) can be overloaded on objects * NeuroElf makes use of this feature by allowing a more C++/Visual Basic style syntax:% NeuroElf xff and xfigure examples % load a VMR into an xff object vmr = xff('some.vmr'); % call the function in @xff/private/aft_Browse.m vmr.Browse; % get xfigure object of the main UI figure neuroelf_gui; global ne_gcfg; mainfig = ne_gcfg.h.MainFig; % switch to "page" 2 (this is overloaded differently in the @xfigure/subsref.m file) mainfig.ShowPage(2); ===== Indexing ===== It is both a blessing and a curse that Matlab uses the same language elements for passing arguments into a function and indexing into a non-scalar array:% creating a 3x3 variable with random numbers randvals = randn(3, 3); % accessing the value at the 3rd row and 2nd column randvals(3, 2) % computing the sum along the 2nd dimension sum(randvals, 2) In both cases, common parentheses, **''(''** and **'')''**, are used to * sub-select values (array elements) from a non-scalar variable * pass arguments (in this case the variable ''randvals'' and the scalar value ''2'') into a function One of the reasons is that even the syntax **''randvals(3, 2)''** can be seen as a function call (and, for user defined objects, actually leads to a function call!). The blessing is that, for user defined objects, this can be used to create very elegant code. The curse, on the other hand, is that it cannot be determined if an expression is an indexing operation of a function call, in cases such asvar_or_function(x, y); ==== Subscript indexing ==== Given that, in principle, all of Matlab's variables support non-scalar (single value/element) content, it is necessary to allow to access individual elements. In the example above, this is exactly what happens with **''randvals(3, 2)''**, which selects the value in the 3rd row and 2nd column of the array. But subscript indexing not only allows to select a single element, but also ranges of elements. For this purpose, each indexing expression can be a list of indices. When a variable is accessed to "read" from it, each of these lists must contain only valid entries (integer numbers, with a minimum of 1 and a maximum according to the size of the array in that dimension). Other than that, there are no restrictions (for reading from an array). When writing to an array (subscript assignment), the list must be unique, but values can be greater than the existing size, in which case the variable is expanded accordingly: % creating a 5x4, i.e. 5 rows and 4 columns, variable with random values fivebyfour = randn(5, 4); % reading the value at row 3, column 1 fivebyfour(3, 1) % reading the entire 2nd row secondrow = fivebyfour(2, :) % reading the 4th column fourthcol = fivebyfour(:, 4) % reading the 2nd to 4th row, 1st to 3rd column smaller3x3 = fivebyfour(2:4, 1:3) % reading all uneven rows and columns unevens = fivebyfour(1:2:end, 1:2:end) % reading (in this order) 4th, 1st, and 3rd rows (complete) r413 = fivebyfour([4, 1, 3], :) % repeatedly reading the 2nd column col2times3 = fivebyfour(:, [2, 2, 2]) All but the last of these expressions are also valid for assignment, in which case the value or array being assigned must either be a scalar (which is then stored in all of the written-to elements) or match in size (i.e. to write into a 3x3 sub-part, only a 1x1 or 3x3 right-hand-side value/array can be used). In addition to these expressions, write access also allows to use indices exceeding the size: % increase the size to 6-by-6 fivebyfour(4:6, 4:6) = 1; Importantly, this performs two steps: * first, the array is expanded, with all new elements being assigned the "neutral" element: * for numeric variables this is 0 * for logical variables this is **''false''** * for char variables this is also 0 (which is NOT the blank character)! so this syntax should not be used to extend a string! * for cell arrays this is an empty double array (empty cell content) * for struct arrays all fields are empty double arrays * next, the portion that is specified with the indexing expression is assigned the provided value(s) This means that in the above example, the first three values of the 5th and 6th columns will be 0! And finally, this type of indexing also allows to **shrink** an array by a special syntax:% remove 2nd row fivebyfour(2, :) = []; % remove 3rd and 4th column fivebyfour(:, 3:4) = []; Please note: for higher dimensional variables (3D/4D), all but one indexing expression **must** be **'':''** (the colon character), because the region to be eliminated must be "removable" without interfering with array storage rules. Overall, the idea behind using subscripts is the most "complete" form of indexing: using as many expressions (arguments) as dimensions. If for instance a variable stores numeric values in 3 dimensions (such as a anatomical 3D image of a subject's head) in a, say, 256-by-256-by-256 array, you can then easily access a slice in any of the three dimensions: % assuming that vol3d is a 3D array, getting three slices in the middle xyslice = vol3d(:, :, 128); xzslice = vol3d(:, 128, :); yzslice = vol3d(128, :, :); ==== Single expression indexing ==== Matlab's internal storage works like this: in a multi-dimensional array, values are stored in order of column indices, then row indices, then 3rd, then 4th dimension, and so forth. That means that in a 3x3 variable, the order of values (in memory) is as follows: % order of values in a 3x3 variable ttvar(1, 1) ttvar(2, 1) ttvar(3, 1) ttvar(1, 2) ttvar(2, 2) ttvar(3, 2) ttvar(1, 3) ttvar(2, 3) ttvar(3, 3) The total number of values is simply the product of all dimension lengths (size). There are contexts in which, for instance, an operation has to be performed to each individual element of an array, regardless of position. Matlab thus allows to access all elements with a single index expression: % accessing element ttvar(2, 2) via single index ttvar(5) Please note that this syntax can be extremely misleading, particularly for people unfamiliar with Matlab index expressions! There is, however, one extremely useful application for this, in which **all** values of an array are considered as a single column: % computing the total average over a 3D volume avgslice = mean(vol3d, 3); avgcolumn = mean(avgslice, 2); totalavg = mean(avgcolumn); % alternatively, by "columnizing" the array, compute in one step totalavg = mean(vol3d(:)); Please note that for multi-dimensional arrays, this syntax leads to an error if the index exceeds the total number of elements, i.e. an array cannot be resized with a single index expression! ==== Variables as indices ==== Matlab also allows to use (numeric) variables (and return values of functions) to be used in index expressions:% create a 4x4 array with random numbers four_by_four = randn(4, 4); % select random row rrowindex = ceil(4 * rand(1, 1)); rrowdata = four_by_four(rrowindex, :) % select random column without variable rcoldata = four_by_four(:, ceil(4 * rand(1, 1))) % and select a random value from an array without "knowing" its size randomarrayvalue = four_by_four(ceil(numel(four_by_four) * rand(1, 1))) In this context, the **''size''** and **''numel''** functions of Matlab are important! **''size(variable)''** returns a 1-by-number-of-dimensions list of array sizes of the passed in variable (or other argument), and **''numel''** returns the total number of elements. This can be used, for instance, when an operation has to be applied to a data, say, slice by slice:% inquire about the size/dimensions of a volume volsize = size(vol3d); % "loop" (iterate) over all slices (in 3rd dimension) for slice = 1:volsize(3) % compute critical value critval = critval_function(vol3d(:, :, slice)); % break (leave loop) if threshold is hit if critval >= 10 break; end end ==== Logical indexing ==== On top of using numbers (or numeric expressions, variables, etc.) in indexing expressions, Matlab also supports selecting indices using a logical expression. The most typical application is by applying a comparison operator to establish a threshold: % generate an array with 10 random numbers r = randn(10, 1); % sum all numbers that are greater (or equal) 0 pos_sum = sum(r(r >= 0)) The indexing expression **''(r >= 0)''** performs an element-by-element comparison with 0. This comparison creates (temporarily) an array of logical (true/false) values with the same size as ''r'', which is then used to "select" all values from r for which the comparison is true. If the comparison operator is used on and then returns a multi-dimensional logical array, the indexing operation automatically converts the result into a column vector (given that arbitrary elements are selected, not allowing the result to be regularly shaped). Logical indexing can also be used with the subscript notation (more than one expression), in which case each logical expression ought to have as many elements as the size of the variable in that dimension:% create a 10-by-10 array with random numbers r10x10 = randn(10, 10); % sub-select some rows and some columns randompart = r10x10(randn(1, 10) > 0, randn(1, 10) > 0)