Arrays, Tuples, Ranges, and Other Fundamental Types

Overview

In Julia, arrays and tuples are the most important data type for working with numerical data

In this lecture we give more details on

  • creating and manipulating Julia arrays
  • fundamental array processing operations
  • basic matrix algebra
  • tuples and named tuples
  • ranges
  • nothing, missing, and unions
In [2]:
# using InstantiateFromURL

# activate the QuantEcon environment
# activate_github("QuantEcon/QuantEconLecturePackages", tag = "v0.9.5");

# load some packages
using LinearAlgebra, Statistics, Compat

Array Basics

(See multi-dimensional arrays documentation)

Since it is one of the most important types, we will start with arrays

Later, we will see how arrays (and all other types in Julia) are handled in a generic and extensible way

Shape and Dimension

We’ve already seen some Julia arrays in action

In [3]:
a = [10, 20, 30]
Out[3]:
3-element Array{Int64,1}:
 10
 20
 30
In [4]:
a = [1.0, 2.0, 3.0]
Out[4]:
3-element Array{Float64,1}:
 1.0
 2.0
 3.0

The output tells us that the arrays are of types Array{Int64,1} and Array{Float64,1} respectively

Here Int64 and Float64 are types for the elements inferred by the compiler

We’ll talk more about types later

The 1 in Array{Int64,1} and Array{Any,1} indicates that the array is one dimensional (i.e., a Vector)

This is the default for many Julia functions that create arrays

In [5]:
typeof(randn(100))
Out[5]:
Array{Float64,1}

In Julia, one dimensional vectors are best interpreted as column vectors, which we will see when we take transposes

We can check the dimensions of a using size() and ndims() functions

In [6]:
ndims(a)
Out[6]:
1
In [7]:
size(a)
Out[7]:
(3,)

The syntax (3,) displays a tuple containing one element – the size along the one dimension that exists

Array vs Vector vs Matrix

In Julia, Vector and Matrix are just aliases for one- and two-dimensional arrays respectively

In [8]:
Array{Int64, 1} == Vector{Int64}
Array{Int64, 2} == Matrix{Int64}
Out[8]:
true

Vector construction with , is then interpreted as a column vector

To see this, we can create a column vector and row vector more directly

In [9]:
[1, 2, 3] == [1; 2; 3]  # both column vectors
Out[9]:
true
In [10]:
[1 2 3]  # a row vector is 2-dimensional
Out[10]:
1×3 Array{Int64,2}:
 1  2  3

As we’ve seen, in Julia we have both

  • one-dimensional arrays (i.e., flat arrays)
  • arrays of size (1, n) or (n, 1) that represent row and column vectors respectively

Why do we need both?

On one hand, dimension matters for matrix algebra

  • Multiplying by a row vector is different to multiplying by a column vector

On the other, we use arrays in many settings that don’t involve matrix algebra

In such cases, we don’t care about the distinction between row and column vectors

This is why many Julia functions return flat arrays by default

 Creating Arrays

Functions that Create Arrays

We’ve already seen some functions for creating a vector filled with 0.0

In [11]:
zeros(3)
Out[11]:
3-element Array{Float64,1}:
 0.0
 0.0
 0.0
In [12]:
# This generalizes to matrices and higher dimensional arrays
zeros(2, 2)
Out[12]:
2×2 Array{Float64,2}:
 0.0  0.0
 0.0  0.0
In [13]:
# To return an array filled with a single value, use `fill`
fill(5.0, 2, 2)
Out[13]:
2×2 Array{Float64,2}:
 5.0  5.0
 5.0  5.0
In [18]:
# You can create an empty array using the `Array()` constructor
# Need to add undef
x = Array{Float64}(undef, 2, 2)
Out[18]:
2×2 Array{Float64,2}:
 3.65443e-315  1.61981e-315
 5.7182e-316   0.0         

The printed values you see here are just garbage values.

(the existing contents of the allocated memory slots being interpreted as 64 bit floats)

If you need more control over the types, fill with a non-floating point:

In [19]:
fill(0, 2, 2)  # fills with 0, not 0.0
Out[19]:
2×2 Array{Int64,2}:
 0  0
 0  0
In [20]:
# Or fill with a boolean type
fill(false, 2, 2)  # produces a boolean matrix
Out[20]:
2×2 Array{Bool,2}:
 0  0
 0  0

Creating Arrays from Existing Arrays

For the most part, we will avoid directly specifying the types of arrays, and let the compiler deduce the optimal types on its own

The reasons for this, discussed in more detail in this lecture, are to ensure both clarity and generality

One place this can be inconvenient is when we need to create an array based on an existing array

First, note that assignment in Julia binds a name to a value, but does not make a copy of that type

In [21]:
x = [1, 2, 3]
y = x
y[1] = 2
x
Out[21]:
3-element Array{Int64,1}:
 2
 2
 3

In the above, y = x simply creates a new named binding called y which refers to whatever x currently binds to

To copy the data, you need to be more explicit

In [22]:
x = [1, 2, 3]
y = copy(x)
y[1] = 2
x
Out[22]:
3-element Array{Int64,1}:
 1
 2
 3

However, rather than making a copy of x, you may want to just have a similarly sized array

In [23]:
x = [1, 2, 3]
y = similar(x)
y
Out[23]:
3-element Array{Int64,1}:
 741582032
 114708560
 114708560

We can also use similar to pre-allocate a vector with a different size, but the same shape

In [24]:
x = [1, 2, 3]
y = similar(x, 4)  # make a vector of length 4
Out[24]:
4-element Array{Int64,1}:
 741587120
 114708560
 114708560
         0
In [25]:
# higher dimensions
x = [1, 2, 3]
y = similar(x, 2, 2)  # make a 2x2 matrix
Out[25]:
2×2 Array{Int64,2}:
 1  7
 1  4

Manual Array Definitions

As we’ve seen, you can create one dimensional arrays from manually specified data like so

In [26]:
a = [10, 20, 30, 40]
Out[26]:
4-element Array{Int64,1}:
 10
 20
 30
 40

In two dimensions we can proceed as follows

In [27]:
a = [10 20 30 40]  # two dimensional, shape is 1 x n
Out[27]:
1×4 Array{Int64,2}:
 10  20  30  40
In [28]:
ndims(a)
Out[28]:
2
In [29]:
a = [10 20; 30 40]  # 2 x 2
Out[29]:
2×2 Array{Int64,2}:
 10  20
 30  40

You might then assume that a = [10; 20; 30; 40] creates a two dimensional column vector but this isn’t the case

In [30]:
a = [10; 20; 30; 40]
Out[30]:
4-element Array{Int64,1}:
 10
 20
 30
 40
In [31]:
ndims(a)
Out[31]:
1

Instead transpose the matrix (or adjoint if complex)

In [32]:
a = [10 20 30 40]'
Out[32]:
4×1 Adjoint{Int64,Array{Int64,2}}:
 10
 20
 30
 40
In [33]:
ndims(a)
Out[33]:
2

Array Indexing

We’ve already seen the basics of array indexing

In [34]:
a = [10 20 30 40]
a[end-1]
Out[34]:
30
In [35]:
a[1:3]
Out[35]:
3-element Array{Int64,1}:
 10
 20
 30

For 2D arrays the index syntax is straightforward

In [36]:
a = randn(2, 2)
a[1, 1]
Out[36]:
0.22578167013407044
In [37]:
a[1, :]  # first row
Out[37]:
2-element Array{Float64,1}:
 0.22578167013407044
 1.4078747699507168 
In [38]:
a[:, 1]  # first column
Out[38]:
2-element Array{Float64,1}:
  0.22578167013407044
 -0.22807178553184515

Booleans can be used to extract elements

In [39]:
a = randn(2, 2)
Out[39]:
2×2 Array{Float64,2}:
 -0.860958   0.756823 
  0.534699  -0.0096501
In [40]:
b = [true false; false true]
Out[40]:
2×2 Array{Bool,2}:
 1  0
 0  1
In [41]:
a[b]
Out[41]:
2-element Array{Float64,1}:
 -0.8609579427092797  
 -0.009650104166440045

This is useful for conditional extraction, as we’ll see below

An aside: some or all elements of an array can be set equal to one number using slice notation

In [42]:
a = zeros(4)
Out[42]:
4-element Array{Float64,1}:
 0.0
 0.0
 0.0
 0.0
In [43]:
a[2:end] .= 42
Out[43]:
3-element view(::Array{Float64,1}, 2:4) with eltype Float64:
 42.0
 42.0
 42.0
In [44]:
a
Out[44]:
4-element Array{Float64,1}:
  0.0
 42.0
 42.0
 42.0

Views and Slices

Using the : notation provides a slice of an array, copying the sub-array to a new array with a similar type

In [45]:
a = [1 2; 3 4]
b = a[:, 2]
@show b
a[:, 2] = [4, 5] # modify a
@show a
@show b;
b = [2, 4]
a = [1 4; 3 5]
b = [2, 4]

A view on the other hand does not copy the value

In [46]:
a = [1 2; 3 4]
@views b = a[:, 2]
@show b
a[:, 2] = [4, 5]
@show a
@show b;
b = [2, 4]
a = [1 4; 3 5]
b = [4, 5]

Note that the only difference is the @views macro, which will replace any slices with views in the expression

An alternative is to call the view function directly – though it is generally discouraged since it is a step away from the math

In [47]:
@views b = a[:, 2]
view(a, :, 2) == b
Out[47]:
true

As with most programming in Julia, it is best to avoid prematurely assuming that @views will have a significant impact on performance, and stress code clarity above all else

Another important lesson about @views is that they are not normal, dense arrays

In [48]:
a = [1 2; 3 4]
b_slice = a[:, 2]
@show typeof(b_slice)
@show typeof(a)
@views b = a[:, 2]
@show typeof(b);
typeof(b_slice) = Array{Int64,1}
typeof(a) = Array{Int64,2}
typeof(b) = SubArray{Int64,1,Array{Int64,2},Tuple{Base.Slice{Base.OneTo{Int64}},Int64},true}

The type of b is a good example of how types are not as they may seem

Similarly

In [49]:
a = [1 2; 3 4]
b = a'   # transpose
typeof(b)
Out[49]:
Adjoint{Int64,Array{Int64,2}}

To copy into a dense array

In [50]:
a = [1 2; 3 4]
b = a' # transpose
c = Matrix(b)  # convert to matrix
d = collect(b) # also `collect` works on any iterable
c == d
Out[50]:
true

Special Matrices

As we saw with transpose, sometimes types that look like matrices are not stored as a dense array

As an example, consider creating a diagonal matrix

In [51]:
d = [1.0, 2.0]
a = Diagonal(d)
Out[51]:
2×2 Diagonal{Float64,Array{Float64,1}}:
 1.0   ⋅ 
  ⋅   2.0

As you can see, the type is 2×2 Diagonal{Float64,Array{Float64,1}}, which is not a 2-dimensional array

The reasons for this are both efficiency in storage, as well as efficiency in arithmetic and matrix operations

In every important sense, matrix types such as Diagonal are just as much a “matrix” as the dense matrices we have using (see the introduction to types lecture for more)

In [52]:
@show 2a
b = rand(2,2)
@show b * a;
2a = [2.0 0.0; 0.0 4.0]
b * a = [0.4752054752832813 0.02390499529967105; 0.027894552736805966 0.10245923947320623]

Another example is in the construction of an identity matrix, where a naive implementation is

In [53]:
b = [1.0 2.0; 3.0 4.0]
b - Diagonal([1.0, 1.0])  # poor style, inefficient code
Out[53]:
2×2 Array{Float64,2}:
 0.0  2.0
 3.0  3.0
In [54]:
# should to this instead
b = [1.0 2.0; 3.0 4.0]
b - I  # good style, and note the lack of dimensions of I
Out[54]:
2×2 Array{Float64,2}:
 0.0  2.0
 3.0  3.0

While the implementation of I is a little abstract to go into at this point, a hint is:

In [55]:
typeof(I)
Out[55]:
UniformScaling{Bool}

This is a UniformScaling type rather than an identity matrix, making it much more powerful and general

Assignment and Passing Arrays

As discussed above, in Julia, the left hand side of an assignment is a “binding” or a label to a value

In [56]:
x = [1 2 3]
y = x  # name `y` binds to whatever value `x` bound to
Out[56]:
1×3 Array{Int64,2}:
 1  2  3

The consequence of this, is that you can re-bind that name

In [57]:
x = [1 2 3]
y = x        # name `y` binds to whatever `x` bound to
z = [2 3 4]
y = z        # only changes name binding, not value!
@show (x, y, z);
(x, y, z) = ([1 2 3], [2 3 4], [2 3 4])

What this means is that if a is an array and we set b = a then a and b point to exactly the same data in your RAM!

In the above, suppose you had meant to change the value of x to the values of y, you need to assign the values rather than the name

In [58]:
x = [1 2 3]
y = x       # name `y` binds to whatever `x` bound to
z = [2 3 4]
y .= z      # now dispatches the assignment of each element
@show (x, y, z);
(x, y, z) = ([2 3 4], [2 3 4], [2 3 4])

Alternatively, you could have used y[:] = z

This applies to in-place functions as well

First, define a simple function for a linear map

In [59]:
function f(x)
    return [1 2; 3 4] * x  # matrix * column vector
end

val = [1, 2]
f(val)
Out[59]:
2-element Array{Int64,1}:
  5
 11

In general, these “out-of-place” functions are preferred to “in-place” functions, which modify the arguments

In [60]:
function f(x)
    return [1 2; 3 4] * x # matrix * column vector
end

val = [1, 2]
y = similar(val)

function f!(out, x)
    out .= [1 2; 3 4] * x
end

f!(y, val)
y
Out[60]:
2-element Array{Int64,1}:
  5
 11

This demonstrates a key convention in Julia: functions which modify any of the arguments have the name ending with ! (e.g. push!)

We can also see a common mistake, where instead of modifying the arguments, the name binding is swapped

In [61]:
function f(x)
    return [1 2; 3 4] * x  # matrix * column vector
end

val = [1, 2]
y = similar(val)

function f!(out, x)
    out = [1 2; 3 4] * x   # MISTAKE! Should be .= or [:]
end
f!(y, val)
y
Out[61]:
2-element Array{Int64,1}:
 92930304
        1

In-place and Immutable Types

Note that scalars are always immutable, such that

In [62]:
y = [1 2]
y .-= 2    # y .= y .- 2, no problem

x = 5
# x .-= 2  # Fails!
x = x - 2  # subtle difference - creates a new value and rebinds the variable
Out[62]:
3

In particular, there is no way to pass any immutable into a function and have it modified

In [63]:
x = 2

function f(x)
    x = 3     # MISTAKE! does not modify x, creates a new value!
end

f(x)          # cannot modify immutables in place
@show x;
x = 2

This is also true for other immutable types such as tuples, as well as some vector types

In [64]:
using StaticArrays
xdynamic = [1, 2]
xstatic = @SVector [1, 2]  # turns it into a highly optimized static vector

f(x) = 2x
@show f(xdynamic)
@show f(xstatic)

# inplace version
function g(x)
    x .= 2x
    return "Success!"
end
@show xdynamic
@show g(xdynamic)
@show xdynamic;

# g(xstatic) # fails, static vectors are immutable
f(xdynamic) = [2, 4]
f(xstatic) = [2, 4]
xdynamic = [1, 2]
g(xdynamic) = "Success!"
xdynamic = [2, 4]

Operations on Arrays

Array Methods

Julia provides standard functions for acting on arrays, some of which we’ve already seen

In [65]:
a = [-1, 0, 1]

@show length(a)
@show sum(a)
@show mean(a)
@show std(a)      # standard deviation
@show var(a)      # variance
@show maximum(a)
@show minimum(a)
@show extrema(a)  # (mimimum(a), maximum(a))
length(a) = 3
sum(a) = 0
mean(a) = 0.0
std(a) = 1.0
var(a) = 1.0
maximum(a) = 1
minimum(a) = -1
extrema(a) = (-1, 1)
Out[65]:
(-1, 1)

To sort an array

In [66]:
b = sort(a, rev = true)  # returns new array, original not modified
Out[66]:
3-element Array{Int64,1}:
  1
  0
 -1
In [67]:
b = sort!(a, rev = true)  # returns *modified original* array
Out[67]:
3-element Array{Int64,1}:
  1
  0
 -1
In [68]:
b == a  # tests if have the same values
Out[68]:
true
In [69]:
b === a  # tests if arrays are identical (i.e share same memory)
Out[69]:
true

Matrix Algebra

For two dimensional arrays, * means matrix multiplication

In [70]:
a = ones(1, 2)
Out[70]:
1×2 Array{Float64,2}:
 1.0  1.0
In [71]:
b = ones(2, 2)
Out[71]:
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0
In [72]:
a * b
Out[72]:
1×2 Array{Float64,2}:
 2.0  2.0
In [73]:
b * a'
Out[73]:
2×1 Array{Float64,2}:
 2.0
 2.0

To solve the linear system $ A X = B $ for $ X $ use A \ B

In [74]:
A = [1 2; 2 3]
Out[74]:
2×2 Array{Int64,2}:
 1  2
 2  3
In [75]:
B = ones(2, 2)
Out[75]:
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0
In [76]:
A \ B
Out[76]:
2×2 Array{Float64,2}:
 -1.0  -1.0
  1.0   1.0
In [79]:
inv(A) * B
Out[79]:
2×2 Array{Float64,2}:
 -1.0  -1.0
  1.0   1.0

Although the last two operations give the same result, the first one is numerically more stable and should be preferred in most cases

Multiplying two one dimensional vectors gives an error – which is reasonable since the meaning is ambiguous

In [78]:
ones(2) * ones(2)
MethodError: no method matching *(::Array{Float64,1}, ::Array{Float64,1})
Closest candidates are:
  *(::Any, ::Any, !Matched::Any, !Matched::Any...) at operators.jl:529
  *(!Matched::Adjoint{#s627,#s626} where #s626<:Union{DenseArray{T,2}, Base.ReinterpretArray{T,2,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray}, Base.ReshapedArray{T,2,A,MI} where MI<:Tuple{Vararg{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64},N} where N} where A<:Union{Base.ReinterpretArray{T,N,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray}, SubArray{T,2,A,I,L} where L where I<:Tuple{Vararg{Union{Int64, AbstractRange{Int64}, Base.AbstractCartesianIndex},N} where N} where A<:Union{Base.ReinterpretArray{T,N,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, Base.ReshapedArray{T,N,A,MI} where MI<:Tuple{Vararg{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64},N} where N} where A<:Union{Base.ReinterpretArray{T,N,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, DenseArray}} where #s627, ::Union{DenseArray{S,1}, Base.ReinterpretArray{S,1,S1,A} where S1 where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray}, Base.ReshapedArray{S,1,A,MI} where MI<:Tuple{Vararg{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64},N} where N} where A<:Union{Base.ReinterpretArray{T,N,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray}, SubArray{S,1,A,I,L} where L where I<:Tuple{Vararg{Union{Int64, AbstractRange{Int64}, Base.AbstractCartesianIndex},N} where N} where A<:Union{Base.ReinterpretArray{T,N,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, Base.ReshapedArray{T,N,A,MI} where MI<:Tuple{Vararg{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64},N} where N} where A<:Union{Base.ReinterpretArray{T,N,S,A} where S where A<:Union{SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, SubArray{T,N,A,I,true} where I<:Union{Tuple{Vararg{Real,N} where N}, Tuple{AbstractUnitRange,Vararg{Any,N} where N}} where A<:DenseArray where N where T, DenseArray} where N where T, DenseArray}}) where {T<:Union{Complex{Float32}, Complex{Float64}, Float32, Float64}, S} at D:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.3\LinearAlgebra\src\matmul.jl:106
  *(!Matched::Adjoint{#s627,#s626} where #s626<:LinearAlgebra.AbstractTriangular where #s627, ::AbstractArray{T,1} where T) at D:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.3\LinearAlgebra\src\triangular.jl:1877
  ...

Stacktrace:
 [1] top-level scope at In[78]:1

If you want an inner product in this setting use dot() or the unicode \cdot<TAB>

In [80]:
dot(ones(2), ones(2))
Out[80]:
2.0

Matrix multiplication using one dimensional vectors is a bit inconsistent – pre-multiplication by the matrix is OK, but post-multiplication gives an error

In [81]:
b = ones(2, 2)
Out[81]:
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0
In [82]:
b * ones(2)
Out[82]:
2-element Array{Float64,1}:
 2.0
 2.0
In [84]:
ones(2,2) * b
Out[84]:
2×2 Array{Float64,2}:
 2.0  2.0
 2.0  2.0

Elementwise Operations

Algebraic Operations

Suppose that we wish to multiply every element of matrix A with the corresponding element of matrix B

In that case we need to replace * (matrix multiplication) with .* (elementwise multiplication)

For example, compare

In [85]:
ones(2, 2) * ones(2, 2)   # matrix multiplication
Out[85]:
2×2 Array{Float64,2}:
 2.0  2.0
 2.0  2.0
In [86]:
ones(2, 2) .* ones(2, 2)   # element by element multiplication
Out[86]:
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0
  • This is a general principle: .x means apply operator x elementwise
  • We have seen the .x before when talking about Broadcasting
  • You remember that .op(x) this just applies operation op to all elements of argument x.
In [87]:
A = -ones(2, 2)
Out[87]:
2×2 Array{Float64,2}:
 -1.0  -1.0
 -1.0  -1.0
In [88]:
A.^2  # square every element
Out[88]:
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0

However in practice some operations are mathematically valid without broadcasting, and hence the . can be omitted

In [89]:
ones(2, 2) + ones(2, 2)  # same as ones(2, 2) .+ ones(2, 2)
Out[89]:
2×2 Array{Float64,2}:
 2.0  2.0
 2.0  2.0

Scalar multiplication is similar

In [90]:
A = ones(2, 2)
Out[90]:
2×2 Array{Float64,2}:
 1.0  1.0
 1.0  1.0
In [91]:
2 * A  # same as 2 .* A
Out[91]:
2×2 Array{Float64,2}:
 2.0  2.0
 2.0  2.0

In fact you can omit the * altogether and just write 2A

Unlike MATLAB and other languages, scalar addition requires the .+ in order to correctly broadcast

In [92]:
x = [1, 2]
x .+ 1     # not x + 1
x .- 1     # not x - 1
Out[92]:
2-element Array{Int64,1}:
 0
 1

Elementwise Comparisons

Elementwise comparisons also use the .x style notation

In [101]:
a = [10, 20, 30]
Out[101]:
3-element Array{Int64,1}:
 10
 20
 30
In [102]:
b = [-100, 0, 100]
Out[102]:
3-element Array{Int64,1}:
 -100
    0
  100
In [103]:
b .> a
Out[103]:
3-element BitArray{1}:
 0
 0
 1
In [106]:
a .== b
Out[106]:
3-element BitArray{1}:
 0
 0
 0

We can also do comparisons against scalars with parallel syntax

In [107]:
b
Out[107]:
3-element Array{Int64,1}:
 -100
    0
  100
In [108]:
b .> 1
Out[108]:
3-element BitArray{1}:
 0
 0
 1

This is particularly useful for conditional extraction – extracting the elements of an array that satisfy a condition

In [111]:
a = randn(4)
Out[111]:
4-element Array{Float64,1}:
 -0.3896820762093061
 -1.4301744007044845
  1.425364737776786 
 -0.3648407272483374
In [112]:
a .< 0
Out[112]:
4-element BitArray{1}:
 1
 1
 0
 1
In [113]:
a[a .< 0]
Out[113]:
3-element Array{Float64,1}:
 -0.3896820762093061
 -1.4301744007044845
 -0.3648407272483374

Changing Dimensions

The primary function for changing the dimensions of an array is reshape()

In [114]:
a = [10, 20, 30, 40]
Out[114]:
4-element Array{Int64,1}:
 10
 20
 30
 40
In [115]:
b = reshape(a, 2, 2)
Out[115]:
2×2 Array{Int64,2}:
 10  30
 20  40
In [116]:
b
Out[116]:
2×2 Array{Int64,2}:
 10  30
 20  40

Notice that this function returns a view on the existing array

This means that changing the data in the new array will modify the data in the old one!

In [117]:
b[1, 1] = 100  # continuing the previous example
Out[117]:
100
In [118]:
b
Out[118]:
2×2 Array{Int64,2}:
 100  30
  20  40
In [119]:
a
Out[119]:
4-element Array{Int64,1}:
 100
  20
  30
  40

To collapse an array along one dimension you can use dropdims()

In [120]:
a = [1 2 3 4]  # two dimensional
Out[120]:
1×4 Array{Int64,2}:
 1  2  3  4
In [121]:
# The return value is an array with the specified dimension “flattened”
dropdims(a, dims = 1)
Out[121]:
4-element Array{Int64,1}:
 1
 2
 3
 4

Broadcasting Functions

Julia provides standard mathematical functions such as log, exp, sin, etc.

In [122]:
log(1.0)
Out[122]:
0.0

By default, these functions act elementwise on arrays

In [123]:
log.(1:4)
Out[123]:
4-element Array{Float64,1}:
 0.0               
 0.6931471805599453
 1.0986122886681098
 1.3862943611198906

Note that we can get the same result as with a comprehension or more explicit loop

In [124]:
[ log(x) for x in 1:4 ]
Out[124]:
4-element Array{Float64,1}:
 0.0               
 0.6931471805599453
 1.0986122886681098
 1.3862943611198906

Comprehensions and Generators

  • Those are very convenient for us to set up arrays
  • consider those examples
In [127]:
# A Comprehension is in square brackets:
# LinSpace is depreciated in Julia v1.3

foo(x,y,z) = sin(x)*0.5y^0.5 + z^2
d = [foo(i,j,k) for i in 1:3, j in range(0.01,stop=0.1,length=5), k in [log(l) for l in 2:7]];
In [128]:
# generator expressions work in a similar way
# just leave the square brackets away
sum(1/n^2 for n=1:1000)  # this allocates no temp array: very efficient!
Out[128]:
1.6439345666815615
In [129]:
# can have indices depend on each other
[(i,j) for i=1:3 for j=1:i]
Out[129]:
6-element Array{Tuple{Int64,Int64},1}:
 (1, 1)
 (2, 1)
 (2, 2)
 (3, 1)
 (3, 2)
 (3, 3)
In [130]:
# you can even condition on the indices
[(i,j) for i=1:3 for j=1:i if i+j == 4]
Out[130]:
2-element Array{Tuple{Int64,Int64},1}:
 (2, 2)
 (3, 1)

Linear Algebra

(See linear algebra documentation)

Julia provides some a great deal of additional functionality related to linear operations

In [131]:
A = [1 2; 3 4]
Out[131]:
2×2 Array{Int64,2}:
 1  2
 3  4
In [132]:
det(A)
Out[132]:
-2.0
In [133]:
tr(A)
Out[133]:
5
In [134]:
eigvals(A)
Out[134]:
2-element Array{Float64,1}:
 -0.3722813232690143
  5.372281323269014 
In [135]:
rank(A)
Out[135]:
2

Ranges

As with many other types, a Range can act as a vector

In [136]:
a = 10:12        # a range, equivalent to 10:1:12
@show Vector(a)  # can convert, but shouldn't

b = Diagonal([1.0, 2.0, 3.0])
b * a .- [1.0; 2.0; 3.0]
Vector(a) = [10, 11, 12]
Out[136]:
3-element Array{Float64,1}:
  9.0
 20.0
 33.0
In [137]:
# 
a = 0.0:0.1:1.0  # 0.0, 0.1, 0.2, ... 1.0
Out[137]:
0.0:0.1:1.0

But care should be taken if the terminal node is not a multiple of the set sizes

In [138]:
maxval = 1.0
minval = 0.0
stepsize = 0.15
a = minval:stepsize:maxval # 0.0, 0.15, 0.3, ...
maximum(a) == maxval
Out[138]:
false

To evenly space points where the maximum value is important, i.e., linspace in other languages

In [139]:
maxval = 1.0
minval = 0.0
numpoints = 10
a = range(minval, maxval, length=numpoints)
# or range(minval, stop=maxval, length=numpoints)

maximum(a) == maxval
Out[139]:
true
  • For the range(minval, maxval, length=numpoints) notation, until the release of Julia v1.1, you will need to have the using Compat in the header, as we do above.
  • Absent that, and until then, you have to supply the keyword stop.

Tuples and Named Tuples

(See tuples and named tuples documentation)

We were introduced to tuples earlier, which provide high-performance immutable sets of distinct types

In [140]:
t = (1.0, "test")
t[1]            # access by index
a, b = t        # unpack
# t[1] = 3.0    # would fail as tuples are immutable
println("a = $a and b = $b")
a = 1.0 and b = test

As well as named tuples, which extend tuples with names for each argument

In [141]:
t = (val1 = 1.0, val2 = "test")
t.val1      # access by index
# a, b = t  # bad style, better to unpack by name with @unpack
println("val1 = $(t.val1) and val1 = $(t.val1)") # access by name
val1 = 1.0 and val1 = 1.0

While immutable, it is possible to manipulate tuples and generate new ones

In [142]:
t2 = (val3 = 4, val4 = "test!!")
t3 = merge(t, t2)  # new tuple
Out[142]:
(val1 = 1.0, val2 = "test", val3 = 4, val4 = "test!!")

Named tuples are a convenient and high-performance way to manage and unpack sets of parameters

In [143]:
function f(parameters)
    α, β = parameters.α, parameters.β  # poor style
    # poor because we'll make errors once
    # we add more parameters!
    return α + β
end

parameters = (α = 0.1, β = 0.2)
f(parameters)
Out[143]:
0.30000000000000004

This functionality is aided by the Parameters.jl package and the @unpack macro

In [144]:
using Parameters

function f(parameters)
    @unpack α, β = parameters  # good style, less sensitive to errors
    return α + β
end

parameters = (α = 0.1, β = 0.2)
f(parameters)
Out[144]:
0.30000000000000004

In order to manage default values, use the @with_kw macro

In [145]:
using Parameters
paramgen = @with_kw (α = 0.1, β = 0.2)  # create named tuples with defaults

# creates named tuples, replacing defaults
@show paramgen()  # calling without arguments gives all defaults
@show paramgen(α = 0.2)
@show paramgen(α = 0.2, β = 0.5);
paramgen() = (α = 0.1, β = 0.2)
paramgen(α=0.2) = (α = 0.2, β = 0.2)
paramgen(α=0.2, β=0.5) = (α = 0.2, β = 0.5)

An alternative approach, defining a new type using struct tends to be more prone to accidental misuse, and leads to a great deal of boilerplate code

For that, and other reasons of generality, we will use named tuples for collections of parameters where possible

Nothing, Missing, and Unions

Sometimes a variable, return type from a function, or value in an array needs to represent the absence of a value rather than a particular value

There are two distinct use cases for this

  1. nothing (“software engineers null”): used where no value makes sense in a particular context due to a failure in the code, a function parameter not passed in, etc.
  2. missing (“data scientists null”): used when a value would make conceptual sense, but it isn’t available

Nothing and Basic Error Handling

The value nothing is a single value of type Nothing

In [146]:
typeof(nothing)
Out[146]:
Nothing

An example of a reasonable use of nothing is if you need to have a variable defined in an outer scope, which may or may not be set in an inner one

In [147]:
function f(y)
    x = nothing
    if y > 0.0
        # calculations to set `x`
        x = y
    end

    # later, can check `x`
    if x === nothing
        println("x was not set")
    else
        println("x = $x")
    end
    x
end

@show f(1.0)
@show f(-1.0);
x = 1.0
f(1.0) = 1.0
x was not set
f(-1.0) = nothing

While in general you want to keep a variable name bound to a single type in Julia, this is a notable exception

Similarly, if needed, you can return a nothing from a function to indicate that it did not calculate as expected

In [148]:
function f(x)
    if x > 0.0
        return sqrt(x)
    else
        return nothing
    end
end
x1 = 1.0
x2 = -1.0
y1 = f(x1)
y2 = f(x2)

# check results with === nothing
if y1 === nothing
    println("f($x2) successful")
else
    println("f($x2) failed");
end
f(-1.0) failed

As an aside, an equivalent way to write the above function is to use the ternary operator, which gives a compact if/then/else structure

In [149]:
function f(x)
    x > 0.0 ? sqrt(x) : nothing  # the "a ? b : c" pattern is the ternary
end

f(1.0)
Out[149]:
1.0

We will sometimes use this form when it makes the code more clear (and it will occasionally make the code higher performance)

Regardless of how f(x) is written, the return type is an example of a union, where the result could be one of an explicit set of types

In this particular case, the compiler would deduce that the type would be a Union{Nothing,Float64} – that is, it returns either a floating point or a nothing

You will see this type directly if you use an array containing both types

In [150]:
x = [1.0, nothing]
Out[150]:
2-element Array{Union{Nothing, Float64},1}:
 1.0     
  nothing

When considering error handling, whether you want a function to return nothing or simply fail depends on whether the code calling f(x) is carefully checking the results

For example, if you were calling on an array of parameters where a priori you were not sure which ones will succeed, then

In [151]:
x = [0.1, -1.0, 2.0, -2.0]
y = f.(x)

# presumably check `y`
Out[151]:
4-element Array{Union{Nothing, Float64},1}:
 0.31622776601683794
  nothing           
 1.4142135623730951 
  nothing           

On the other hand, if the parameter passed is invalid and you would prefer not to handle a graceful failure, then using an assertion is more appropriate

In [154]:
function f(x)
    @assert x > 0.0
    sqrt(x)
end

f(1.0)
Out[154]:
1.0

Finally, nothing is a good way to indicate an optional parameter in a function

In [155]:
function f(x; z = nothing)

    if(z === nothing)
        println("No z given with $x")
    else
        println("z = $z given with $x")
    end
end

f(1.0)
f(1.0, z=3.0)
No z given with 1.0
z = 3.0 given with 1.0

An alternative to nothing, which can be useful and sometimes higher performance, is to use NaN to signal that a value is invalid returning from a function

In [156]:
function f(x)
    if x > 0.0
        return x
    else
        return NaN
    end
end

f(0.1)
f(-1.0)

@show typeof(f(-1.0))
@show f(-1.0) == NaN  # note, this fails!
@show isnan(f(-1.0))  # check with this
typeof(f(-1.0)) = Float64
f(-1.0) == NaN = false
isnan(f(-1.0)) = true
Out[156]:
true

Note that in this case, the return type is Float64 regardless of the input for Float64 input

Keep in mind, though, that this only works if the return type of a function is Float64

Exceptions

(See exceptions documentation)

While returning a nothing can be a good way to deal with functions which may or may not return values, a more robust error handling method is to use exceptions

Unless you are writing a package, you will rarely want to define and throw your own exceptions, but will need to deal with them from other libraries

The key distinction for when to use an exceptions vs. return a nothing is whether an error is unexpected rather than a normal path of execution

An example of an exception is a DomainError, which signifies that a value passed to a function is invalid

In [158]:
# throws exception, turned off to prevent breaking notebook
# sqrt(-1.0)

# to see the error
try sqrt(-1.0); catch err; err end  # catches the exception and prints it
Out[158]:
DomainError(-1.0, "sqrt will only return a complex result if called with a complex argument. Try sqrt(Complex(x)).")

Another example you will see is when the compiler cannot convert between types

In [159]:
# throws exception, turned off to prevent breaking notebook
# convert(Int64, 3.12)

# to see the error
try convert(Int64, 3.12); catch err; err end  # catches the exception and prints it.
Out[159]:
InexactError(:Int64, Int64, 3.12)

If these exceptions are generated from unexpected cases in your code, it may be appropriate simply let them occur and ensure you can read the error

Occasionally you will want to catch these errors and try to recover, as we did above in the try block

In [160]:
function f(x)
    try
        sqrt(x)
    catch err                # enters if exception thrown
        sqrt(complex(x, 0))  # convert to complex number
    end
end

f(0.0)
f(-1.0)
Out[160]:
0.0 + 1.0im

Missing

(see “missing” documentation)

The value missing of type Missing is used to represent missing value in a statistical sense

For example, if you loaded data from a panel, and gaps existed

In [161]:
x = [3.0, missing, 5.0, missing, missing]
Out[161]:
5-element Array{Union{Missing, Float64},1}:
 3.0     
  missing
 5.0     
  missing
  missing

A key feature of missing is that it propagates through other function calls - unlike nothing

In [162]:
f(x) = x^2

@show missing + 1.0
@show missing * 2
@show missing * "test"
@show f(missing);      # even user-defined functions
@show mean(x);
missing + 1.0 = missing
missing * 2 = missing
missing * "test" = missing
f(missing) = missing
mean(x) = missing

The purpose of this is to ensure that failures do not silently fail and provide meaningless numerical results

This even applies for the comparison of values, which

In [163]:
x = missing

@show x == missing
@show x === missing  # an exception
@show ismissing(x);
x == missing = missing
x === missing = true
ismissing(x) = true

Where ismissing is the canonical way to test the value

In the case where you would like to calculate a value without the missing values, you can use skipmissing

In [164]:
x = [1.0, missing, 2.0, missing, missing, 5.0]

@show mean(x)
@show mean(skipmissing(x))
@show coalesce.(x, 0.0);  # replace missing with 0.0;
mean(x) = missing
mean(skipmissing(x)) = 2.6666666666666665
coalesce.(x, 0.0) = [1.0, 0.0, 2.0, 0.0, 0.0, 5.0]

As missing is similar to R’s NA type, we will see more of missing when we cover DataFrames

Exercises

Exercise 1

This exercise uses matrix operations that arise in certain problems, including when dealing with linear stochastic difference equations

If you aren’t familiar with all the terminology don’t be concerned – you can skim read the background discussion and focus purely on the matrix exercise

With that said, consider the stochastic difference equation

$$ X_{t+1} = A X_t + b + \Sigma W_{t+1} \tag{1} $$

Here

  • $ X_t, b $ and $ X_{t+1} $ are $ n \times 1 $
  • $ A $ is $ n \times n $
  • $ \Sigma $ is $ n \times k $
  • $ W_t $ is $ k \times 1 $ and $ \{W_t\} $ is iid with zero mean and variance-covariance matrix equal to the identity matrix

Let $ S_t $ denote the $ n \times n $ variance-covariance matrix of $ X_t $

Using the rules for computing variances in matrix expressions, it can be shown from (1) that $ \{S_t\} $ obeys

$$ S_{t+1} = A S_t A' + \Sigma \Sigma' \tag{2} $$

It can be shown that, provided all eigenvalues of $ A $ lie within the unit circle, the sequence $ \{S_t\} $ converges to a unique limit $ S $

This is the unconditional variance or asymptotic variance of the stochastic difference equation

As an exercise, try writing a simple function that solves for the limit $ S $ by iterating on (2) given $ A $ and $ \Sigma $

To test your solution, observe that the limit $ S $ is a solution to the matrix equation

$$ S = A S A' + Q \quad \text{where} \quad Q := \Sigma \Sigma' \tag{3} $$

This kind of equation is known as a discrete time Lyapunov equation

The QuantEcon package provides a function called solve_discrete_lyapunov that implements a fast “doubling” algorithm to solve this equation

Test your iterative method against solve_discrete_lyapunov using matrices

$$ A = \begin{bmatrix} 0.8 & -0.2 \\ -0.1 & 0.7 \end{bmatrix} \qquad \Sigma = \begin{bmatrix} 0.5 & 0.4 \\ 0.4 & 0.6 \end{bmatrix} $$

Exercise 2

Take a stochastic process for $ \{y_t\}_{t=0}^T $

$$ y_{t+1} = \gamma + \theta y_t + \sigma w_{t+1} $$

where

  • $ w_{t+1} $ is distributed Normal(0,1)
  • $ \gamma=1, \sigma=1, y_0 = 0 $
  • $ \theta \in \Theta \equiv \{0.8, 0.9, 0.98\} $

Given these parameters

  • Simulate a single $ y_t $ series for each $ \theta \in \Theta $ for $ T = 150 $. Feel free to experiment with different $ T $
  • Overlay plots of the rolling mean of the process for each $ \theta \in \Theta $, i.e. for each $ 1 \leq \tau \leq T $ plot
$$ \frac{1}{\tau}\sum_{t=1}^{\tau}y_T $$
  • Simulate $ N=200 $ paths of the stochastic process above to the $ T $, for each $ \theta \in \Theta $, where we refer to an element of a particular simulation as $ y^n_t $
  • Overlay plots a histogram of the stationary distribution of the final $ y^n_T $ for each $ \theta \in \Theta $. Hint: pass alpha to a plot to make it transparent (e.g. histogram(vals, alpha = 0.5)) or use stephist(vals) to show just the step function for the histogram
  • Numerically find the mean and variance of this as an ensemble average, i.e. $ \sum_{n=1}^N\frac{y^n_T}{N} $ and $ \sum_{n=1}^N\frac{(y_T^n)^2}{N} -\left(\sum_{n=1}^N\frac{y^n_T}{N}\right)^2 $

Later, we will interpret some of these in this lecture

Exercise 3

Let the data generating process for a variable be

$$ y = a x_1 + b x_1^2 + c x_2 + d + \sigma w $$

where $ y, x_1, x_2 $ are scalar observables, $ a,b,c,d $ are parameters to estimate, and $ w $ are iid normal with mean 0 and variance 1

First, let’s simulate data we can use to estimate the parameters

  • Draw $ N=50 $ values for $ x_1, x_2 $ from iid normal distributions

Then, simulate with different $ w $

  • Draw a $ w $ vector for the N values and then y from this simulated data if the parameters were $ a = 0.1, b = 0.2 c = 0.5, d = 1.0, \sigma = 0.1 $
  • Repeat that so you have M = 20 different simulations of the y for the N values

Finally, calculate order least squares manually (i.e., put the observables into matrices and vectors, and directly use the equations for OLS rather than a package)

  • For each of the M=20 simulations, calculate the OLS estimates for $ a, b, c, d, \sigma $
  • Plot a histogram of these estimates for each variable