Project Nayuki

MATLAB language pet peeves

MATLAB is a proprietary programming language targeted toward linear algebra and rapid development. Its design dates back from the 1980s and has evolved over the decades. But having used many other programming languages in my career, the pain points of MATLAB show up clearly.

Weakly typed Booleans

The logical type represents a Boolean value, which is separate from the numeric type. However, the logical values false and true behave exactly as the numeric values 0 and 1 (respectively) in most computations:

disp(true + true);  % Prints 2
disp(false * 5);  % Prints 0

(This behavior is the same in C, C++, JavaScript, Python; but disallowed in Java, C#.)

An important difference between logical and numeric types shows up when we take a vector subscript of an array/matrix:

array = -2 : +2;
sub0 = [true, true, true, true, true];
sub1 = [1, 1, 1, 1, 1];
disp(array(sub0));  % Equals [-2, -1, 0, 1, 2]
disp(array(sub1));  % Equals [-2, -2, -2, -2, -2]

sub2 = [false, true, false, true, true];
sub3 = [0, 1, 0, 1, 1];
disp(array(sub2));  % Equals [-1, 1, 2]
disp(array(sub3));  % Runtime error
Minimum 2 dimensions
x = 8;
y = [2, 7];
z = [1, 4; 5, 0];
w = cat(3, [0 1; 2 3], [4 5; 6 7]);
disp(size(x));  % Equals [1 1]
disp(size(y));  % Equals [1 2]
disp(size(z));  % Equals [2 2]
disp(size(w));  % Equals [2 2 2]

Scalar numbers are matrices, and vectors are matrices; in other words, every numerical object has at least 2 dimensions. The extra degenerate dimensions are impure, and can lead to silent mistakes such as unintentionally extending along a dimension. Also, for many applications that only need 1-dimensional arrays, there is a needless distinction between row vectors and column vectors. Other linear algebra systems like NumPy for Python don’t suffer from this problem, and can work with 0-dimensional, 1D, 2D, 3D, etc. arrays.

Nesting means concatenation
x = [1, 2];
x = [x, [3, 4]];
disp(x);  % Equals [1 2 3 4]

y = [5; 6];
y = [y; [7; 8]];
disp(y);  % Equals [5; 6; 7; 8]

Syntactically nesting a matrix into another one yields a new matrix representing their concatenation. This can be used as an idiom for appending or prepending one or more elements to a matrix. But this is conceptually impure, and other languages like Python and JavaScript will give you a properly nested array if you tried to use this syntax.

Matrix extension
temp = [5];
temp(3) = 9;
disp(temp);  % Prints [5 0 9]

Setting an element beyond the bounds of a vector/matrix will silently extend it. This makes it easy to mask mistakes in index calculations. JavaScript also has this same misfeature. Java and Python will crash consistently. C/C++ will exhibit undefined behavior (may work correctly, may corrupt data or crash).

End-relative indexing
temp = [15, 36, 27, 98];
disp(temp(2 : end-1));  % Equals [36, 27]

range = 2 : end-1;  % Syntax error
disp(temp(range));  % Will not work

Indexing a vector/matrix relative to the end (without using a length function) is convenient. But this use of the keyword end is confined to subscripts and cannot be used as a standalone value.

Irregular indexing syntax
temp = [1 2; 3 4];
disp(temp(:, :));  % Prints [1 2; 3 4]

function result = foo
  result = [5 6; 7 8];
end

disp(foo()(:, :));
% Error: ()-indexing must appear last in an index expression.

disp((foo)(:, :));
disp((foo())(:, :));
% Error: Unbalanced or unexpected parenthesis or bracket.

The matrix indexing/slicing syntax can’t be used everywhere; sometimes you need to introduce an intermediate variable. This might be similar to the distinction between lvalues and rvalues in C and C++. This could also suggest that the MATLAB language isn’t designed properly with a recursive grammar.

Setting function values
[a, b] = function(c, d)
  a = c;
  b = d;
  % Cannot write: return [c, d];
end

A function returns values by assigning output variables, instead of returning an expression’s value. This is reminiscent of old languages like Pascal and Visual Basic, where the value to return is assigned to the function’s name. Newer languages like C, Java, Python, Ruby, etc. do not do this.

Multiple function syntaxes

A zero-argument function can be declared without parentheses:

function foo
end

function foo()  % Equivalent
end

A zero-result function can have no result array:

function foo
end

function [] = foo  % Equivalent
end

A one-result function can have no result array brackets:

function result = foo
end

function [result] = foo  % Equivalent
end

The examples above contrast with how in most languages, there is only one way to declare a function with a certain number of arguments and return values. (One exception is C++, where a zero-argument function can optionally be declared with void as the list of arguments.)

Conflated variables and functions

A zero-argument (nullary) function can be invoked without parentheses:

function result = foo()
  result = 6;
end

disp(foo);  % Prints 6
disp(foo());  % Prints 6

A variable can be “invoked” with pointless parentheses:

a = 5;
b = {3; 0};
c = datetime('today');

disp(a);  disp(a());  % Same
disp(b);  disp(b());  % Same
disp(c);  disp(c());  % Same

Most other languages make strong distinctions between the function itself and the value returned by a function. Also, no other language I know of allows a non-function-type variable to be “called” with parentheses.

One consequence of this is that in MATLAB code, today and now are effectively magic variables whose values change even without explicitly reassigning them. Or are they actually I/O functions, who knows?

Conflated calls and indexing

Calling a function looks exactly the same as indexing into a vector/matrix:

function result = bar(a, b)
  result = a + b;
end
qux = [4 5; 6 7];

disp(bar(0, 1));  % Function call
disp(qux(2, 3));  % Matrix subscript

Whereas most popular languages use parentheses (()) for function invocations and square brackets ([]) for array subscripts.

Output-sensitive function calls

MATLAB functions can behave differently and return different values depending on the number of output elements that the caller wants to receive. An example is polyfit:

xs = [3 1 4 1 5 9];
ys = [2 7 1 8 2 8];
deg = 2;

[p0]          = polyfit(xs, ys, deg);
[p1, S1]      = polyfit(xs, ys, deg);
[p2, S2, mu2] = polyfit(xs, ys, deg);

% p0 and p1 are the same, but different from p2

In most other languages, the way that the caller receives the result of a function call cannot affect the behavior of the function being called. One would normally expect the result [p0] to be a prefix of [p1, S1], which in turn should be a prefix of [p2, S2, mu2], but this is not the case in MATLAB.

Commands vs. functions
mkdir Hello
mkdir('World');

Some features are available as commands with unquoted strings (like Command Arg1 Arg2 ...), whereas some features are available as functions taking proper expressions as values (like Function(Arg1, Arg2, ...)). This distinction of syntax is arbitrary and jarring.

Can’t pass by reference
function foo(m)
  m(1, 1) = 5;
end

mat = [0 1; 2 3];
disp(mat(1, 1));  % Prints 0
foo(mat);
disp(mat(1, 1));  % Still prints 0

When passing a matrix into a function, a copy is made. It’s not possible to change a value and make it visible to the caller, except by returning a new value and having the caller explicitly reassign a variable. Although the lack of pass-by-reference reduces confusion, it makes it harder to modularize algorithms that need to update data in place.

Strings are weird
u = 'abc';
v = ['abc'];  % Identical to u
w = [v, 'def'];  % Equal to 'abcdef'
x = ['ghi'; 'jkl'];  % OK, column of 2 strings
y = ['mn'; 'opqr'];  % Error: Mismatched row lengths
z = {'mn'; 'opqr'};  % OK, cell array

Strings essentially behave like row vectors of characters, not like atomic values. This is why a list of strings is usually chosen to be represented as a cell array, not as a matrix.

Cell arrays are weird
a = cell(2, 2);
a{1, 1} = 5;
disp(a{1, 1});  % 1*1 matrix, i.e. [5]
disp(a(1, 1));  % 1*1 cell array, i.e. {[5]}

Cell arrays are a recursive/nestable heterogeneous data structure, unlike flat matrices of numbers. Indexing into a cell array requires some care to distinguish between getting a sub-array versus getting an element value.

Dummy overloading arguments

Calculating the maximum (or minimum) of a matrix M along a certain dimension d uses the syntax max(M, [], d), with [] being a constant dummy argument. This contrasts with other summarization functions like sum, which uses the notation sum(M, d) to operate on a certain dimension. The dummy argument is necessary to disambiguate from the 2-argument form max(A, B), which calculates the maximum of each pair of elements between matrices A and B.

Semicolon suppresses printing

Many kinds of statements will print the value of the statement if it doesn’t end in a semicolon. For example, variable/array assignments and function calls are printed. Adding a semicolon suppresses the print, and it is rare to print values except when designing and debugging code. Although the semicolons make MATLAB code look more like C, they are a form of syntactic salt to work around the annoying default behavior of printing almost every value. A similar example can be found in Windows batch scripts, where every command is printed unless it’s prefixed with @.

1-based array indexing
temp = [99, 47];
disp(temp(1));  % Prints 99

A few languages oriented to beginners will count arrays from 1, such as BASIC, Lua, and Mathematica. Most serious languages count from 0, including C, C++, C#, D, Java, JavaScript, Python, Ruby, assembly, and the list goes on. There are many great reasons to start indexes at 0. One of them is that if the array is represented by a pointer to its head element, then stepping forward by 0 slots will give you the head element. Another reason is that multi-dimensional indexing is like pixels[y * width + x], but in 1-based indexing the code would be like pixels[(y - 1) * width + x].

Inclusive end index
temp = [31, 41, 59, 26, 53];
disp(temp(2 : 4));  % Equals [41, 59, 26]

For ranges, using inclusive end indexes looks more natural but makes length calculations harder. The number of elements in array(start : end) is end - start + 1. Python takes the opposite approach where end indexes are exclusive, which makes it easy to write ranges like array[start : start + len].

Optional commas between columns

Space-separated data is too easy to misuse. For example, [x y] and [x -y] are both row vectors of length 2, but [x-y] and [x - y] each only have a single element. Separating data with spaces causes subtle mistakes in other languages too:

// C code example
const char *strings[] = {
  "Alpha",
  "Bravo"  // Forgot comma
  "Charlie",
};
// So the array is effectively
// {"Alpha", "BravoCharlie"}

# Python code example
strings = [
  "Alpha",
  "Bravo"  # Forgot comma
  "Charlie",
]
# So the array is effectively
# ["Alpha", "BravoCharlie"]
Standard library in global namespace

Although MATLAB has packages and namespaces, the enormous set of all standard library functions is dumped into the global namespace. In a sufficiently large user application, it is not hard to declare a variable whose name shadows an existing function.

Other languages like C and JavaScript suffer from a bloated global namespace. Python has a large number of functions and types defined in the global namespace, but they don’t grow over time because new work goes into modules.

No compound assignment operators
x = 0;
x = x + 1;
x = x - 2;
x = x * 5;

matrixLongName = [1 2; 3 4];
fooLongName = 5;
barLongName = 3;
matrixLongName(fooLongName-barLongName, barLongName*14) = ...
matrixLongName(fooLongName-barLongName, barLongName*14) / 6;

MATLAB doesn’t have compound assignment operators, which increases the redundancy of common code idioms. Major languages like C, C++, C#, Java, JavaScript, Python, Ruby all have operators like +=, *=, &=, etc. (Also the C family has operators like ++ as a shorthand for += 1.)

Esoteric operators

Some infrequently used operations have single-character operators that are too convenient. But they hurt code readability and consume syntax elements that could be used for better things. For example, \ means mldivide(), and ' means transpose().

No index-of function
temp = [16, 39, 25, 84, 50, 47];
i = indexof(temp, 25);  % Hypothetical function
disp(i);  % Should print 3

j = find(temp == 25);  % Actual code
disp(j);  % Prints 3

Finding the index of a value in an array is a basic task, yet there is no direct way to do it. Instead, find() with == serves as a replacement.

Integer to datetime
k = 736829;
dt = datetime(datestr(k));
disp(dt);  % Prints 14-May-2017

Converting from an integer to a datetime object is needlessly painful. The standard library provides functions for many of the other conversions (e.g. datetime to int, datetime to string), but not a direct conversion from integer to datetime. Instead, we have to take an indirect route by turning it into a string and then parsing the string.

Date format strings

The format strings for the functions datetime(), datenum(), and datestr() are inconsistent. For example, a two-digit month is denoted as 'MM' in datetime() but 'mm' in datestr() and datenum().

CSV file I/O madness

  • csvwrite() is limited to writing 5 significant base-10 digits. This low precision makes the output only suitable for display, not for saving values that will be read by future computations.

  • The functions of dlmread() and dlmwrite() essentially replicate and exceed the functionality of csvread() and csvwrite(), respectively.

  • The 4 aforementioned functions all use 0-based indexing for row and column ranges, unlike the 1-based convention of MATLAB in general.

  • csvread() and dlmread() can only read CSV data from a file, not from a string in memory. Hence there is no convenient way to download a CSV file from the web and parse it. The workaround is either to save the CSV text to a temporary file and read it with a convenient function, or to keep the string in memory and manually parse each line using textscan().

  • importdata() can parse text tables that have a mix of character data and numeric data, and return both types of data separately. By contrast, dlmread() can skip text rows and columns, but only parse and the remaining numeric data and return that. Once again, a more general function appears to supersede the features of an existing function.

Tables are slow
tic;
n = 300;
x = zeros(n, 4);
for i = 1 : n
  for j = 1 : 4
    x(i, j) = i * j;
  end
end
toc;  % About 10 milliseconds
% Similar run time if matrix is replaced by cell array

tic;
n = 300;
t = zeros(n, 1);
y = table(t, t, t, t);
for i = 1 : n
  for j = 1 : 4
    y{i, j} = i * j;
  end
end
toc;  % About 1000 milliseconds

In one application I worked on, MATLAB’s tables were a tidy way to organize the data and avoid coding errors compared to matrices with raw column number indexing. But after tentatively rewriting the code to use a table, the performance was abysmal with more than 10× slowdown. So I ended up using a structure of columns instead, which provide some degree of organization and no loss of performance.

Additional notes