A new function makima computes a piecewise cubic interpolant with continuous first-order derivatives. It produces fewer undulations than the existing spline function, with less flattening than pchip. This form of interpolant has been an option in interp1, interp2, interp3, interpn, and griddedInterpolant since R2017b and is now also a standalone function. For more details see Cleve Moler’s blog post about makima, from which the code used to generate the image above is taken.

The matchpairs function solves the linear assignment problem, which requires each row of a matrix to be assigned to a column in such a way that the total cost of the assignments (given by the “sum of assigned elements” of the matrix) is minimized (or maximized). This problem can also be described as finding a minimum-weight matching in a weighted bipartite graph. The problem arises in sparse matrix computations.

The equilibrate function take as input a matrix A (dense or sparse) and returns a permutation matrix P and nonsingular diagonal matrices R and C such that R*P*A*C has diagonal entries of magnitude 1 and off-diagonal entries of magnitude at most 1. The outputs P, R, and C are sparse. This matrix equilibration can improve conditioning and can be a useful preprocessing step both in computing incomplete LU preconditioners and in iterative solvers. See, for example, the recent paper Max-Balanced Hungarian Scalings by Hook, Pestana, Tisseur, and Hogg.

The scaling produced by equilibrate is known as Hungarian scaling, and its computation involves solving a linear assignment problem.

A powerful new feature allows a function to check that its input arguments have the desired class, size, and other aspects. Previously, this could be done by calling the function validateattributes with each argument in turn or by using the inputParser object (see Loren Shure’s comparison of the different approaches).

The new feature provides a more structured way of achieving the same effect. It also allows default values for arguments to be specified. Function argument validation contains many aspects and has a long help page. We just illustrate it with a simple example that should be self-explanatory.

function fav_example(a,b,f,t)
  arguments
    a (1,1) double               % Double scalar.
    b {mustBeNumeric,mustBeReal} % Numeric type and real.
    f double {mustBeMember(f,[-1,0,1])} = 0 % -1, 0 (default), or 1.
    t string = "-"               % String, with default value.
  end
  % Body of function.
end

The functions mustBeNumeric and mustBeMember are just two of a long list of validation functions. Here is an example call:

>> a = 0; b = 2; f = 11; fav_example(a,b,f)
Error using fav_example
Invalid input argument at position 3. Value must be a member of ...
this set:
    -1
    0
    1 

One caveat is that a specification within the arguments block

v (1,:) double    

allows not just 1-by-n vectors (for any n) but also n-by-1 vectors, because of the convention that in many situations MATLAB does not distinguish between row and column vectors.

For more extensive examples, see New with R2019b – Function Argument Validation and Spider Plots and More Argument Validation by Jiro Doke at MathWorks Blogs.

One can now dot index the result of a function call. For example, consider

>> optimset('Display','iter').Display
ans =
    'iter'

The output of the dot call can even be indexed:

>> optimset('OutputFcn',{@outfun1,@outfun2}).OutputFcn{2}
ans =
  function_handle with value:
    @outfun2

Of course, these are contrived examples, but in real codes such indexing can be useful. This syntax was not allowed in R2019a and earlier.

One can now specify hexadecimal and binary values using the prefixes 0x or 0X for hexadecimal (with upper or lower case for A to F) and 0b or 0B for binary. By default the results have integer type.

>> 0xA
ans =
  uint8
   10
>> 0b101
ans =
  uint8
   5

The Release Notes say “Starting in R2019b, MATLAB changes the rules for name resolution, impacting the precedence order of variables, nested functions, local functions, and external functions. The new rules simplify and standardize name resolution”. A number of examples are given where the meaning of code is different in R2019b, and they could cause an error to be generated for codes that ran without error in earlier versions. As far as I can see, though, these changes are not likely to affect well-written MATLAB code.

Program files larger than “approximately 128 MB” will no longer open or run. This is an interesting change because I cannot imagine writing an M-file of that size. Good programming practice dictates splitting a code into separate functions and not including all of them in one program file. Presumably, some MATLAB users have produced such long files—perhaps automatically generated.

For a modern introduction and reference to MATLAB, see MATLAB Guide (third edition, 2017)

MATLAB now has a way of avoiding unnecessarily repeating the factorization of a matrix. In the past, if we wanted to solve two linear systems Ax = b and Ay = c involving the same square, general nonsingular matrix A, writing

x = A\b;
y = A\c;

was wasteful because A would be LU factorized twice. The way to avoid this unnecessary work was to factorize A explicitly then re-use the factors:

[L,U] = lu(A);
x = U\(L\b);
y = U\(L\c);

This solution lacks elegance. It is also less than ideal from the point of view of code maintenance and re-use. If we want to adapt the same code to handle symmetric positive definite A we have to change all three lines, even though the mathematical concept is the same:

R = chol(A);
x = R\(R'\b);
y = R\(R'\c);

The new decomposition function creates an object that contains the factorization of interest and allows it to be re-used. We can now write

dA = decomposition(A);
x = dA\b;
y = dA\c;

MATLAB automatically chooses the factorization based on the properties of A (as the backslash operator has always done). So for a general square matrix it LU factorizes A, while for a symmetric positive definite matrix it takes the Cholesky factorization. The decomposition object knows to use the factors within it when it encounters a backslash. So this example is functionally equivalent to the first two.

The type of decomposition can be specified as a second input argument, for example:

dA = decomposition(A,'lu');
dA = decomposition(A,'chol');
dA = decomposition(A,'ldl');
dA = decomposition(A,'qr');

These usages make the intention explicit and save a little computation, as MATLAB does not have to determine the matrix type. Currently, 'svd' is not supported as a second input augment.

This is a very welcome addition to the matrix factorization capabilities in MATLAB. Tim Davis proposed this idea in his 2013 article Algorithm 930: FACTORIZE: An Object-Oriented Linear System Solver for MATLAB. In Tim’s approach, all relevant functions such as orth, rank, condest are overloaded to use a decomposition object of the appropriate type. MATLAB currently uses the decomposition object only in backslash, forward slash, negation, and conjugate transpose.

The notation dA used in the MathWorks documentation for the factorization object associated with A grates a little with me, since dA (or ) is standard notation for a perturbation of the matrix A and of course the derivative of a function. In my usage I will probably write something more descriptive, such as decompA or factorsA.

Functions written for versions of MATLAB prior to 2016b might assume that certain inputs are character vectors (the old way of storing strings) and might not work with the new string arrays introduced in MATLAB R2016b. The function convertStringsToChars can be used at the start of a function to convert strings to character vectors or cell arrays of character vectors:

function y = myfun(x,y,z)
[x,y,z] = convertStringToChars(x,y,z);

The pre-2016b function should then work as expected.

More functions now accept string arguments. For example, we can now write (with double quotes as opposed to the older single quote syntax for character vectors)

>> A = gallery("moler",3)
A =
     1    -1    -1
    -1     2     0
    -1     0     3

There is more support in R2017b for tall arrays (arrays that are too large to fit in memory). They can be indexed, with sorted indices in the first dimension; functions including median, plot, and polyfit now accept tall array arguments; and the random number generator used in tall array calculations can be independently controlled.

The new wordcloud function produces word clouds. The following code runs the function on a file of words from The Princeton Companion to Applied Mathematics.

words = fileread('pcam_words.txt');
words = string(words);
words = splitlines(words);
words(strlength(words)<5) = [];
words = categorical(words);
figure
wordcloud(words)

I had prepared this file in order to generate a word cloud using the Wordle website. Here is the MATLAB version followed by the Wordle version: Wordle removes certain stop words (common but unimportant words) that wordcloud leaves in. Also, whereas Wordle produces a different layout each time it is called, different layouts can be obtained from wordcloud with the syntax wordcloud(words,'LayoutNum',n) with n a nonnegative integer.

New functions islocalmin and islocalmax find local minima and maxima within discrete data. Obviously targeted at data manipulation applications, such as analysis of time series, these functions offer a number of options to define what is meant by a minimum or maximum.

The code

A = randn(10); plot(A,'LineWidth',1.5); hold on
plot(islocalmax(A).*A,'r.','MarkerSize',25); 
hold off, axis tight

finds maxima down the columns of a random matrix and produces this plot:

Whereas min and max find the smallest and largest elements of an array, the new functions mink(A,k) and maxk(A,k) find the k smallest and largest elements of A, columnwise if A is a matrix.

This code compares the speed of the arithmetics for LU factorization of a 250-by-250 matrix.

%QUAD_PREC_TEST1  LU factorization.
rng(1), n = 250;
A = randn(n); A_mp = mp(A,34); A_vpa = vpa(A,34);
t = clock; [L,U] = lu(A);     t_dp = etime(clock, t)
t = clock; [L,U] = lu(A_mp);  t_mp = etime(clock, t)
t = clock; [L,U] = lu(A_vpa); t_vpa = etime(clock, t)
fprintf('mp/double: %8.2e, vpa/double: %8.2e, vpa/mp: %8.2e\n',....
         t_mp/t_dp, t_vpa/t_dp, t_vpa/t_mp)

The output is

t_dp =
   1.0000e-03
t_mp =
   9.8000e-02
t_vpa =
   2.4982e+01
mp/double: 9.80e+01, vpa/double: 2.50e+04, vpa/mp: 2.55e+02

Here, the MP quadruple precision is 98 times slower than double precision, but VPA is 25,000 times slower than double precision.

%QUAD_PREC_TEST2  Complete eigensystem of nonsymmetric matrix.
rng(1), n = 125;
A = randn(n); A_mp = mp(A,34); A_vpa = vpa(A,34);
t = clock; [V,D] = eig(A);     t_dp = etime(clock, t)
t = clock; [V,D] = eig(A_mp);  t_mp = etime(clock, t)
t = clock; [V,D] = eig(A_vpa); t_vpa = etime(clock, t)
fprintf('mp/double: %8.2e, vpa/double: %8.2e, vpa/mp: %8.2e\n',....
         t_mp/t_dp, t_vpa/t_dp, t_vpa/t_mp)

The output is

t_dp =
   1.8000e-02
t_mp =
   1.3430e+00
t_vpa =
   1.0839e+02
mp/double: 7.46e+01, vpa/double: 6.02e+03, vpa/mp: 8.07e+01

Here, MP is 75 times slower than double, and VPA is closer to the speed of MP.

%QUAD_PREC_TEST3  Complete eigensystem of symmetric matrix.
rng(1), n = 200;
A = randn(n); A = A + A'; A_mp = mp(A,34); A_vpa = vpa(A,34);
t = clock; [V,D] = eig(A);     t_dp = etime(clock, t)
t = clock; [V,D] = eig(A_mp);  t_mp = etime(clock, t)
t = clock; [V,D] = eig(A_vpa); t_vpa = etime(clock, t)
fprintf('mp/double: %8.2e, vpa/double: %8.2e, vpa/mp: %8.2e\n',....
         t_mp/t_dp, t_vpa/t_dp, t_vpa/t_mp)

The output is

t_dp =
   1.1000e-02
t_mp =
   3.5800e-01
t_vpa =
   1.2246e+02
mp/double: 3.25e+01, vpa/double: 1.11e+04, vpa/mp: 3.42e+02

Note that there are at least three different algorithms that could be used here (the QR algorithm, divide and conquer, and multiple relatively robust representations), so the three eig invocations may be using different algorithms.

%QUAD_PREC_TEST4  Componentwise exponentiation.
rng(1), n = 1000;
A = randn(n); A_mp = mp(A,34); A_vpa = vpa(A,34);
t = clock; X = exp(A);     t_dp = etime(clock, t)
t = clock; X - exp(A_mp);  t_mp = etime(clock, t)
t = clock; X - exp(A_vpa); t_vpa = etime(clock, t)
fprintf('mp/double: %8.2e, vpa/double: %8.2e, vpa/mp: %8.2e\n',....
         t_mp/t_dp, t_vpa/t_dp, t_vpa/t_mp)

The output is

t_dp =
   7.0000e-03
t_mp =
   8.5000e-02
t_vpa =
   3.4852e+01
mp/double: 1.21e+01, vpa/double: 4.98e+03, vpa/mp: 4.10e+02

Both MP and VPA come closer to double precision on this problem of computing the scalar exponential at each matrix element.

Not too much emphasis should be put on the precise timings, which vary with the value of the dimension n. The main conclusions are that 34-digit MP arithmetic is 1 to 2 orders of magnitude slower than double precision and 34-digit VPA arithmetic is 3 to 4 orders of magnitude slower than double precision.

It is worth noting that the documentation for the Multiprecision Computing Toolbox states that the toolbox is optimized for quadruple precision.

It is also interesting to note that the authors of the GNU MPFR library (for multiple-precision floating-point computations with correct rounding) are working on optimizing the library for double precision and quadruple precision computations, as explained in a talk given by Paul Zimmermann at the 24th IEEE Symposium on Computer Arithmetic in July 2017; see the slides and the corresponding paper.

[Updated September 6, 2017 to clarify mp.Info and April 6, 2018 to change “Multiplication” to “Multiprecision” in penultimate paragraph.]

% Bad style.
x = (A')\b;
y = [1];
D = (diag(x));

The parentheses around A' are unnecessary. Without them, the statement is legal and unambiguous. But of course the parentheses would be necessary in, for example, x = (A*B)'\c;

y is a scalar. There is no need to enter it as a 1-by-1 matrix by putting it in square brackets.

There is no need for parentheses around the diag function.

% Good style.
x = A'\b;
y = 1;
D = diag(x);

% Bad style.
q=sqrt(2);
a = 1;
Z =zeros(3);
for i=1:10, q=q+1/q; end

Stick to a convention about spacing around around equals signs and plus and minus signs. I think spaces make the code more readable.

% Good style.
q = sqrt(2);
a = 1;
Z = zeros(3);
for i = 1:10, q = q + 1/q; end

% Bad style.
figure;
for i = 2:n,
    x(i) = x(i-1)^2*y(i);
end;

Semicolons suppress output, but in this example there is no need for the them since no output is returned by the figure or end functions. The comma would only be necessary if the for loop was collapsed onto one line.

% Good style.
figure
for i = 2:n
    x(i) = x(i-1)^2*y(i);
end

The following if statement is perfectly correct.

% Bad style.
if flag == 1
   fprintf('Failed to converge.\n')
elseif flag == 2
   fprintf('Overflow encountered.\n')
elseif flag == 3
   fprintf('Division by zero.\n')
elseif flag == 4
   fprintf('Tolerance too small.\n')
end

But given its regular structure, it is better written as a switch statement, which is shorter, brings out the parallelism in the logic, and is easier to check.

% Good style.
switch flag
   case 1, fprintf('Failed to converge.\n')
   case 2, fprintf('Overflow encountered.\n')
   case 3, fprintf('Division by zero.\n')
   case 4, fprintf('Tolerance too small.\n')
end

In the list of output arguments in a function call a tilde signifies that the output argument in that position is not wanted and so can be discarded (and hence need not be computed). This notation was introduced in MATLAB R2009b. Generally, there is no point in providing a tilde as the last output argument, as a well written function will check the number of requested outputs and not compute any trailing outputs that are not requested.

The singular value decomposition (SVD) of a matrix A is computed by the call [U,S,V] = svd(A). In the following example we wish to compute the left singular vectors of A but not the singular values or right singular vectors.

% Bad style.
[U,~,~] = svd(A);

% Good style
[U,~] = svd(A);

An obvious question is why the second example was not shortened to

U = svd(A);

The reason is that with one output argument the svd function has a special behavior: it returns a matrix of singular values, not the matrix U of left singular vectors.

The H1 line in a MATLAB program file is a comment line that is the second line in the file and contains the program name followed by a one-line description of the program. It is good practice to provide every program with an H1 line, as MATLAB itself does. Several MATLAB functions make use of H1 lines. For example, when the help function is invoked with a directory name and the directory in question does not contain a contents.m file, a list of contents is created from the H1 lines of the program files in the directory and displayed.

Here is an example of a function with an H1-line (this is an updated version of a function from The Matrix Computation Toolbox).

function show(x)
%SHOW   Display signs of matrix elements.
%   SHOW(X) displays X in `FORMAT +' form, that is,
%   with `+', `-' and  blank representing positive, negative
%   and zero elements respectively.

f = get(0,'Format'); % Get current numerical format.
format +
disp(x)
format(f)            % Restore original numeric format.

The parula color map has been modified slightly. The difference is subtle, but as the following example illustrates, the R2017a parula is a bit more vibrant and has a bit less cyan and yellow in the blues and greens.

I note, however, that the difference between the old and the new parula is smaller when the images are converted to the CMYK color space, as they must be for printing.

The new heatmap function plots a heatmap of tabular data. Although it is intended mainly for use with the table data type, I think heatmap will be useful for getting insight into the structure of matrices, as illustrated by the following examples.

String arrays, introduced in MATLAB R2016b, can now be formed using double quotes:

>> s = string('This is a string') % R2016b
s = 
    "This is a string"
>> t = "This is a string"         % R2017a
t = 
    "This is a string"
>> isequal(s,t)
ans =
  logical
   1
>> whos
  Name      Size            Bytes  Class     Attributes

  s         1x1               166  string              
  t         1x1               166  string

However, there is one major caveat. Many MATLAB functions that take a char as input argument have not yet been adapted to accept strings. Hence

>> A = gallery('moler',3)
A =
     1    -1    -1
    -1     2     0
    -1     0     3
>> A = gallery("moler",3)
Error using nargin
Argument must be either a character vector or a function handle.
Error in gallery (line 191)
nargs = nargin(matname);

I expect that such functions will be updated in future releases.

A new function missing creates missing values appropriate to the data type in question.

>> A = ones(2); A(2,2) = missing
A =
     1     1
     1   NaN

> d = datetime('2014-05-26')
d = 
  datetime
   26-May-2014
>> d(2) = missing
d = 
  1×2 datetime array
   26-May-2014   NaT

Until converted to the target datatype, a missing value has class missing:

>> m = missing
m = 
  missing
    
>> class(m)
ans =
    'missing'

The release notes report performance improvements under a variety of headings, including execution engine, scripts, and mathematics functions. These are very welcome, as the user automatically benefits from them. One comment that caught my eye is “The backslash command A\B is faster when operating on negative definite matrices”. I think this means that MATLAB checks whether the matrix, A, is symmetric with all negative diagonal elements and, if it is, attempts a Cholesky factorization of -A.