# Introduction to Functional Programming with Anonymous Functions, Part 3 5

Posted by **Loren Shure**,

Tucker McClure is an Application Engineer with The MathWorks. He spends his time helping our customers accelerate their work with the right tools and problem-solving techniques. Today, he'll be discussing how "functional programming" can help create brief and powerful MATLAB code.

### Contents

#### Recap

For Part 1, click here. For Part 2, click here.

When we left off, we had implemented conditional statements, recursion, and multi-line statements in anonymous functions, so today we'll tackle loops.

Before we get started, let's implement the functions that we'll need again.

```
iif = @(varargin) varargin{2*find([varargin{1:2:end}], 1, 'first')}();
recur = @(f, varargin) f(f, varargin{:});
curly = @(x, varargin) x{varargin{:}};
```

#### Loops

Note that the recursive sequences we created in the last part could also have been implemented with `for` loops. For instance, here's factorial of `n`:

factorial = 1; for k = 1:n factorial = k * factorial; end

Many times, recursive functions can be written iteratively in loops. However, we can't use `for` or `while` in an anonymous function, so instead of asking how we can unwrap recursive functions into iterative loops, let's ask the reverse: how can we implement loops with recursive functions?

#### Loops via Recursion

To loop properly, one must know:

- What to do each iteration
- If the process should continue to the next iteration
- What's available when the loop begins

Allowing the "what to do" to be a function (`fcn`) of some state (`x`), the "if it should continue" to be another function (`cont`) of the state, and "what's available when the loop begins" to be the initial state (`x0`), we can write a `loop` function. This is a big step, so bear with me for some explanation!

On each step, the loop function will call the `cont` function, passing in all elements of the state, `x`, as in `cont(x{:})`. If that returns false (meaning we shouldn't continue), the current state, `x`, is returned. Otherwise, if we *should* continue, it calls `fcn` with all elements of the current state, as in `fcn(x{:})`, and passes the output from that to the next iteration. Letting this single iteration be denoted as `f`, we can build the anonymous function `loop` using our `recur` function.

loop = @(x0, cont, fcn) ... % Header recur(@(f, x) iif(~cont(x{:}), x, ... % Continue? true, @() f(f, fcn(x{:}))), ... % Iterate x0); % from x0.

For this trivial example, the state is simply the iteration count. We'll increase the count every iteration until the count `>= n` and return the final count. All this does therefore is count from 0 to the input `n`. Not very interesting, but it demonstrates the loop.

count = @(n) loop({0}, ... % Initialize state, k, to 0 @(k) k < n, ... % While k < n @(k) {k + 1}); % k = k + 1 (returned as cell array) arrayfun(count, 1:10)

ans = [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

I suppose that worked, but why are we using cell arrays to store the state, such as `{0}` and `{k+1}`? There are two reasons. First, if `x` is a cell array, then when we dump all elements of `x` into `fcn`, they become multiple arguments! That is, `fcn(x{:})` is the same as `fcn(x{1}, x{2}, ...)`. So instead of our function taking a big cell array for an input, it can take named arguments, which we'll use below. Second, we do this because it allows a function to *return* multiple elements that will be used by the next iteration, so if a function needed to return `y` and `z`, which would be arguments to the next iteration, it can simply return one cell array, `{y, z}`. It makes it easy to use. Here's a factorial example demonstrating this. The state is two different things: the iteration count, `k`, and factorial of the previous number, `x`. Note that both values of the state, `k` and `x`, are inputs to all of the functions. Note here how we're using `@(k, x)` for our functions. By allowing `x` to be a cell array, each element of the array becomes an argument such as `k` or `x`!

factorial = @(n) loop({1, 1}, ... % Start with k = 1 and x = 1 @(k, x) k <= n, ... % While k <= n @(k, x) {k + 1, ... % k = k + 1; k * x}); % x = k * x;

Call it:

factorial(5)

ans = [6] [120]

Wait, we wanted 120 (the fifth number of the factorial sequence), so what's that 6 doing there?

#### A Better Loop

Remember how we return the full state? That's not very useful for this factorial example, as we get both `k` and the number we want as outputs in that cell array. Because the whole state isn't generally useful, let's add a `cleanup` function to our loop. We'll execute this when the loop is done (right after `~cont(...)` returns `false`). Our `cleanup` function will take the full state and return only the important parts.

loop = @(x0, cont, fcn, cleanup) ... % Header recur(@(f, x) iif(~cont(x{:}), cleanup(x{:}), ... % Continue? true, @() f(f, fcn(x{:}))), ... % Iterate x0); % from x0.

Now here's factorial, with clean output.

factorial = @(n) loop({1, 1}, ... % Start with k = 1 and x = 1 @(k,x) k <= n, ... % While k <= n @(k,x) {k + 1, ... % k = k + 1; k * x}, ... % x = k * x; @(k,x) x); % End, returning x

The result:

factorial(5)

ans = 120

First seven numbers of factorial:

arrayfun(factorial, 1:7)

ans = Columns 1 through 6 1 2 6 24 120 720 Column 7 5040

That's better.

I'll be the first to admit that the loop is a bit longer and much more rigid than a normal MATLAB loop. On the other hand, it can be used in anonymous functions, and its syntax has a certain cleanliness to it in that it doesn't modify any variables that live outside the loop; it has its own scope. This is one nice feature of `loop` being a *function* that takes *code* (functions) as arguments.

#### Doing More in a Loop

Let's say we want to do something else in the loop, but don't want its output passed to the next iteration, like printing something out. Remember the `do_three_things` example from last time? We executed numerous statements by putting them in a cell array and used `curly` to access the output we cared about. We can do that here, in a loop. For example, let's write out a function to print `n` digits of the factorial sequence. We'll use an array to store two things. The first will be the number that `fprintf` returns, which we don't care about. The second will be the cell array we want to return, `k` and `x`. We'll access that cell array with `curly`, as in `curly({..., {k, x}}, 2)`, which just returns `{k, x}`.

say_it = @(k, x) fprintf('Factorial(%d): %d\n', k, x); print_factorial = @(n) loop({1, 1}, ... % Start with k=1, x=1 @(k,x) k <= n, ... % While k <= n @(k,x) curly({say_it(k,k*x),... % Print, discard {k + 1, ... % k = k + 1; k * x}}, ... % x = k * x; 2), ... % Return {k+1,k*x}. @(k,x) x); % End, returning x

print_factorial(7);

Factorial(1): 1 Factorial(2): 2 Factorial(3): 6 Factorial(4): 24 Factorial(5): 120 Factorial(6): 720 Factorial(7): 5040

Now we're executing multiple things and only returning what we want while inside a loop built built on recursion and anonymous conditionals! We've come a long way since Part 1.

As a practical note, recall that because these loops use recursion, there's a limit to the number of times they can loop (MATLAB has a recursion limit, which is a setting in Preferences). Also, a recursive implementation of a loop isn't the most efficient. For this reason, it's best to implement `loop` itself in a file that can then be used in the same way. If it's in a file, it can also be kept on the MATLAB path so that it can be used anywhere.

function x = loop(x, cont, f, cleanup) while cont(x{:}) x = f(x{:}); end if nargin == 4 x = cleanup(x{:}); end end

#### Final Example

This brings us to our final example. Below, we'll simulate a simple harmonic oscillator over time, using a structure to store dissimilar states, including a complete time history of the oscillator. This might simulate, for example, the sway of a lamp that's hanging from the ceiling after an earthquake.

% First, calculate a state transition matrix that represents a harmonic % oscillator with damping. Multiplying this by |x| produces |x| at a % slightly later time. The math here isn't important to the example. Phi = expm(0.5*[0 1; -1 -0.2]); % Now create the loop. x = loop({[1; 0], 1}, ... % Initial state, x = [1; 0] @(x,k) k <= 100, ... % While k <= 100 @(x,k) {[x, Phi * x(:, end)], ... % Update x k + 1}, ... % Update k @(x,k) x); % End, return x % Create a plot function. plot_it = @(n, x, y, t) {subplot(2, 1, n), ... % Select subplot. plot(x(n, :)), ... % Plot the data. iif(nargin==4, @() title(t), ... % If there's a true, []), ... % title, add it. ylabel(y), ... % Label y xlabel('Time (s)')}; % and x axes. % Plot the result. plot_it(1, x, 'Position (m)', 'Harmonic Oscillator'); plot_it(2, x, 'Velocity (m/s)');

#### Summary

That's it for loops via recursion!

Let's look back at what we did over these three parts. First, we started with a simple `map` utility function to demonstrate the function-of-functions idea. Then we created our ubiquitous inline if, which further enabled recursion (a conditional is necessary to make recursion stop!). We also showed using multiple statements by storing their outputs in a cell array. Finally, we created a `loop` construct on top of our recursion functions.

At this point, we've done more than just scratch the surface of functional programming. We've used MATLAB's interesting constructs, such as function handles, cell arrays, and `varargin` to implement a functional programming framework, allowing a new syntax within MATLAB, where code can be arguments to flow control functions. Here's a roundup of what we created.

map = @(val, fcns) cellfun(@(f) f(val{:}), fcns); mapc = @(val, fcns) cellfun(@(f) f(val{:}), fcns, 'UniformOutput', 0); iif = @(varargin) varargin{2*find([varargin{1:2:end}], 1, 'first')}(); recur = @(f, varargin) f(f, varargin{:}); paren = @(x, varargin) x(varargin{:}); curly = @(x, varargin) x{varargin{:}}; loop = @(x0,c,f,r)recur(@(g,x)iif(c(x{:}),@()g(g,f(x{:})),1,r(x{:})),x0);

These have also been programmed as "normal" MATLAB functions so that they can be kept on the path and used whenever they're needed. These can be found under "Functional Programming Constructs" in File Exchange, here.

Thanks for reading. I hope this has both enabled a new level of detail in anonymous functions in MATLAB and helped demonstrate the wide range of possibilities available within the MATLAB language.

Do you have other functional programming patterns you use in your code? For instance, a do-while loop is just like our loop above except that it always runs at least one iteration. Any ideas how to program this or other interesting constructs in anonymous functions? Please let us know here!

Get the MATLAB code

Published with MATLAB® R2012b

**Category:**- Function Handles,
- functional programming

## 5 CommentsOldest to Newest

**1**of 5

Tucker,

Thank you for doing this series, and I’ve been trying to follow it (although sometimes I’m stretching my brain a bit to do it.) I’ve been trying to apply this to solve a piece of programming which I’ve been working on (it’s described in Answers: http://www.mathworks.com/matlabcentral/answers/62634-how-to-make-an-anonymous-function-for-variable-amount-of-input-data

Basically, I’ve got an array of objects X(1)…X(n), each of which has a collection of vector properties (a, b, c, …) which may be of different lengths in each of the instances of X. I also have a cell array where each cell contains a list of indices for each of the X objects and I’m trying to create a new object which uses the indices from my cell array to extract a specific set of entries from each of the X values and combine them into a rectangular matrix. The best suggestion I’ve gotten on Answers so far includes creating a new method, but I feel like there must be a way to do this with an anonymous function…

Any thoughts?

Thanks,

Dan

**2**of 5

Not sure I exactly understand you, Dan, but is this what you’re trying to do?

f = @(objects, indices) ... % Function header

arrayfun(@(k) objects(k).field(indices{k}), ... % Get indices of obj(k)

1:length(objects), ... % for k = 1:n.

'UniformOutput', false); % Output in cells.

my_struct_array(1).field = (1:10)';

my_struct_array(2).field = (2:2:20)';

my_indices = {1:5, 6:10};

f(my_struct_array, my_indices)

Also, if the indices are guaranteed to be the same length, then you could output to a matrix instead of a cell array by adding a cell2mat(…) around the arrayfun(…).

I’ve copied this to Answers too. Hope that helps!

**3**of 5

Tucker,

Yes, thank you… I believe that it does do what I need (or almost- I just need to make the field the third input to the anonymous function.

Thanks,

Dan

**4**of 5

I’m trying to use function handles for subplots instead of cycles. My problem is, when I try to use ‘axes’ in a fnhandle definition with multiple commands, ie. in a cell array, an error is triggered I presume because including it in a cell array implicitly sets nargout to 1.

@(ntr) {plot(timevec,squeeze(datamat(1,:,ntr))); axis(‘off’)}

Is there a solution to this?

**5**of 5

Hi bknakker,

You’re exactly correct, and there *is* a solution. In the File Exchange entry (http://www.mathworks.com/matlabcentral/fileexchange/39735-functional-programming-constructs), I include a function called ‘void’ (there’s a hint about this at the end of Part 2). It allows one to call a function that doesn’t produce any output in a place where output is required, such as in a cell array. For example:

list = {figure(), plot(randn(50), ‘k.’), void(@() axis(‘off’))};

While it’s not quite as short as might be desired, it’s the easiest way I know of. Hope that helps!

## Recent Comments