Continuous Optimization

3841
Continuous Optimization
3842 Since machine learning algorithms are implemented on a computer, the

3843 mathematical formulations are expressed as numerical optimization meth-
3844 ods. This chapter describes the basic numerical methods for training ma-
3845 chine learning models. Training a machine learning model often boils
3846 down to finding a good set of parameters. The notion of “good” is de-
3847 termined by the objective function or the probabilistic model, which we
3848 will see examples of in the second part of this book. Given an objective
Since we consider 3849 function finding the best value is done using optimization algorithms.
data and models in
3850 This chapter covers two main branches of continuous optimization (Fig-
RD the
optimization 3851 ure 7.1): unconstrained and constrained optimization. We will assume in
problems we face 3852 this chapter that our objective function is differentiable (see Chapter 5),
are continuous 3853 hence we have access to a gradient at each location in the space to help
optimization 3854 us find the optimum value. By convention most objective functions in ma-
problems, as
opposed to
3855 chine learning are intended to be minimized, that is the best value is the
combinatorial 3856 minimum value. Intuitively finding the best value is like finding the val-
optimization 3857 leys of the objective function, and the gradients point us uphill. The idea is
problems for 3858 to move downhill (opposite to the gradient) and hope to find the deepest
discrete variables.
3859 point. For unconstrained optimization, this is the only concept we need,
3860 but there are several design choices which we discuss in Section 7.1. For
3861 constrained optimization, we need to introduce other concepts to man-
3862 age the constraints (Section 7.2). We will also introduce a special class
3863 of problems (convex optimization problems in Section 7.3) where we can
3864 make statements about reaching the global optimum.
global minimum Consider the function in Figure 7.2. The function has a global minimum
around the value x = 4.5 which has the objective function value of
around 47. Since the function is “smooth” the gradients can be used to
help find the minimum by indicating whether we should take a step to the
right or left. This assumes that we are in the correct bowl, as there exists
local minimum another local minimum around the value x = 0.7. Recall that we can solve
Stationary points for all the stationary points of a function by calculating its derivative and
are points that have setting it to zero. Let
zero gradient.
`(x) = x4 + 7x3 + 5x2 17x + 3. (7.1)
212
Draft chapter (September 14, 2018) from “Mathematics for Machine Learning” c 2018 by Marc
Peter Deisenroth, A Aldo Faisal, and Cheng Soon Ong. To be published by Cambridge University
Press. Report errata and feedback to http://mml-book.com. Please do not post or distribute this
file, please link to https://mml-book.com.
Continuous Optimization 213
Continuous Figure 7.1 A mind

optimization Stepsize map of the concepts
related to
optimization, as
presented in this
Unconstrained chapter. There are
optimization Gradient descent Momentum
two main ideas:
gradient descent
and convex
optimization.
Stochastic
gradient
Constrained descent
optimization
Lagrange multipliers Convex
Convex optimization
Convex conjugate
& duality
Linear
programming
Quadratic Chapter 12
programming Classification
Its gradient is given by

d`(x)
= 4x3 + 21x2 + 10x 17. (7.2)
dx
Since this is a cubic equation, it has three solutions when set to zero. Two
of them are minima and one is a maximum (around x = 1.4). Recall
that to check whether a stationary point is a minimum or maximum we
need to take the derivative a second time and check whether the second
derivative is positive or negative at the stationary point.
d2 `(x)
= 12x2 + 42x + 10 (7.3)
dx2
3865 By substituting our visually estimated values of x = 4.5, ⇣ 1.4, 0.7 we ⌘
2
3866 will observe that as expected the middle point is a maximum d dx `(x)
2 < 0
3867 and the other two stationary points are minimums.
3868 Note that we have avoided analytically solving for values of x in the pre-
3869 vious discussion, although for low order polynomials such as the above we
3870 could. In general, we are unable to find analytic solutions, and hence we In fact according to
the Abel-Ruffini
theorem, also
c 2018 Marc Peter Deisenroth, A. Aldo Faisal, Cheng Soon Ong. To be published by Cambridge University
knownPress.
as Abel’s
impossibility
theorem, there is in
general no algebraic
solution for
polynomials of
degree 5 or more.
214 Continuous Optimization
Figure 7.2 Example

objective function.
Gradients are
indicated by arrows,
and the global
minimum is
indicated by the
dashed blue line.
3871 need to start at some value, say x0 = 10 and follow the gradient. The
3872 gradient indicates that we should go right, but not how far (this is called
3873 the step size). Furthermore, if we had started at the right side (e.g. x0 = 0)
3874 the gradient would have led us to the wrong minimum. Figure 7.2 illus-
3875 trates the fact that for x > 1, the gradient points towards the minimum
3876 on the right of the figure, which has a larger objective value.
3877 We will see in Section 7.3 a class of functions called convex functions
3878 that do not exhibit this tricky dependency on the starting point of the
convex functions 3879 optimization algorithm. For convex functions all local minima are global
3880 minimum. It turns out that many machine learning objective functions
3881 are designed such that they are convex, and we will see an example in
3882 Chapter 12.
3883 The discussion in this chapter so far was about a one dimensional func-
3884 tion, where we are able to visualize the ideas of gradients, descent direc-
3885 tions and optimal values. In the rest of this chapter we develop the same
3886 ideas in high dimensions. Unfortunately we can only visualize the con-
3887 cepts in one dimension, but some concepts do not generalize directly to
3888 higher dimensions, therefore some care needs to be taken when reading.
3889 7.1 Optimization using Gradient Descent

We now consider the problem of solving for the minimum of a real-valued
function
min f (x) (7.4)

x
Draft (2018-09-14) from Mathematics for Machine Learning. Errata and feedback to https://mml-book.com.
7.1 Optimization using Gradient Descent 215
Figure 7.3 Gradient

descent on a 2
dimensional
quadratic surface
(shown as a
heatmap). See
Example 7.1 for a
description.
3890 where f : Rd ! R is an objective function that captures the machine

3891 learning problem at hand. We assume that our function f is differentiable,
3892 and we are unable to analytically find a solution in closed form.
3893 Gradient descent is a first-order optimization algorithm. To find a local
3894 minimum of a function using gradient descent, one takes steps propor-
3895 tional to the negative of the gradient of the function at the current point.
3896 Recall from Chapter 5 that the gradient points in the direction of the steep-
3897 est ascent and it is orthogonal to the contour lines of the function we wish
3898 to optimize. We use the
Let us consider multivariate functions. Imagine a surface (described by convention of row
vectors for
the function f (x)) with a ball starting at a particular location x0 . When
gradients.
the ball is released, it will move downhill in the direction of steepest de-
scent. Gradient descent exploits the fact that f (x0 ) decreases fastest if one
moves from x0 in the direction of the negative gradient ((rf )(x0 ))> of
f at x0 . We assume in this book that the functions are differentiable, and
refer the reader to more general settings in Section 7.4. Then, if
x1 = x0 ((rf )(x0 ))> (7.5)
3899 for a small step size > 0 then f (x1 ) 6 f (x0 ). Note that we use the
3900 transpose for the gradient since otherwise the dimensions will not work
3901 out.
This observation allows us to define a simple gradient-descent algo-
rithm: If we want to find a local optimum f (x⇤ ) of a function f : Rn !
R, x 7! f (x), we start with an initial guess x0 of the parameters we wish
to optimize and then iterate according to
xi+1 = xi i ((rf )(xi ))
>
. (7.6)
3902 For suitable step size i, the sequence f (x0 ) > f (x1 ) > . . . converges to
3903 a local minimum.
c 2018 Marc Peter Deisenroth, A. Aldo Faisal, Cheng Soon Ong. To be published by Cambridge University Press.
Example 7.1
Consider a quadratic function in two dimensions
✓ ◆  >   > 
x1 x 2 1 x1 5 x1
f = 1 (7.7)
x2 x2 1 20 x2 3 x2
with gradient
✓ ◆   
x1 2 1 x1 5
rf = . (7.8)
x2 1 20 x2 3
Starting at the initial location x0 = [ 3, 1]> , we iteratively apply (7.6)
to obtain a sequence of estimates that converge to the minimum value
(illustrated in Figure 7.3). We can see (both from the figure and by
plugging x0 into (7.8)) that the the gradient at x0 points north and
east, leading to x1 = [ 1.98, 1.21]> . Repeating that argument gives us
x2 = [ 1.32, 0.42]> , and so on.
3904 Remark. Gradient descent can be relatively slow close to the minimum:
3905 Its asymptotic rate of convergence is inferior to many other methods. Us-
3906 ing the ball rolling down the hill analogy, when the surface is a long thin
3907 valley the problem is poorly conditioned (Trefethen and Bau III, 1997).
3908 For poorly conditioned convex problems, gradient descent increasingly
3909 ‘zigzags’ as the gradients point nearly orthogonally to the shortest direc-
3910 tion to a minimum point, see Fig. 7.3. }
3911 7.1.1 Stepsize

3912 As mentioned earlier, choosing a good stepsize is important in gradient
The stepsize is also3913 descent. If the stepsize is too small, gradient descent can be slow. If the
called the learning3914 stepsize is chosen too large, gradient descent can overshoot, fail to con-
rate
3915 verge, or even diverge. We will discuss the use of momentum in the next
3916 section. It is a method that smoothes out erratic behavior of gradient up-
3917 dates and dampens oscillations.
3918 Adaptive gradient methods rescale the stepsize at each iteration, de-
3919 pending on local properties of the function. There are two simple heuris-
3920 tics (Toussaint, 2012):
3921 • When the function value increases after a gradient step, the step size
3922 was too large. Undo the step and decrease the stepsize.
3923 • When the function value decreases the step could have been larger. Try
3924 to increase the stepsize.
3925 Although the “undo” step seems to be a waste of resources, using this
3926 heuristic guarantees monotonic convergence.
Example 7.2 (Solving a Linear Equation System)

When we solve linear equations of the form Ax = b, in practice we solve
Ax b = 0 approximately by finding x⇤ that minimizes the the squared
error
kAx bk2 = (Ax b)> (Ax b) (7.9)
if we use the Euclidean norm. The gradient of (7.9) with respect to x is
rx = 2(Ax b)> A . (7.10)
We can use this gradient directly in a gradient descent algorithm. How-
ever for this particular special case, it turns out that there is an analytic
solution, which can be found by setting the gradient to zero. We can see
that this analytic solution is given by Ax = b. We will see more on solving
squared error problems in Chapter 9.
3927 Remark. When applied to the solution of linear systems of equations Ax =

3928 b gradient descent may converge slowly. The speed of convergence of gra-
3929 dient descent is dependent on the condition number  = (A) max
(A)min
, which condition number
3930 is the ratio of the maximum to the minimum singular value (Section 4.5)
3931 of A. The condition number essentially measures the ratio of the most
3932 curved direction versus the least curved direction, which corresponds to
3933 our imagery that poorly conditioned problems are long thin valleys: they
3934 are very curved in one direction, but very flat in the other. Instead of di-
3935 rectly solving Ax = b, one could instead solve P 1 (Ax b) = 0, where
3936 P is called the preconditioner. The goal is to design P 1 such that P 1 A preconditioner
3937 has a better condition number, but at the same time P 1 is easy to com-
3938 pute. For further information on gradient descent, pre-conditioning and
3939 convergence we refer to (Boyd and Vandenberghe, 2004, Chapter 9). }
3940 7.1.2 Gradient Descent with Momentum

3941 As illustrated in Figure 7.3, the convergence of gradient descent may be
3942 very slow if the curvature of the optimization surface is such that the there
3943 are regions which are poorly scaled. The curvature is such that the gra-
3944 dient descent steps hops between the walls of the valley, and approaches
3945 the optimum in small steps. The proposed tweak to improve convergence
3946 is to give gradient descent some memory.
Gradient descent with momentum (Rumelhart et al., 1986) is a method
that introduces an additional term to remember what happened in the
previous iteration. This memory dampens oscillations and smoothes out
the gradient updates. Continuing the ball analogy, the momentum term
Goh (2017) wrote
emulates the phenomenon of a heavy ball which is reluctant to change an intuitive blog
directions. The idea is to have a gradient update with memory to imple- post on gradient
descent with
momentum.
ment a moving average. The momentum-based method remembers the

update xi at each iteration i and determines the next update as a linear
combination of the current and previous gradients
xi+1 = xi i ((rf )(xi ))

>
+ ↵ xi (7.11)
xi = xi xi 1 = i 1 ((rf )(xi 1 ))
>
, (7.12)
3947 where ↵ 2 [0, 1]. Sometimes we will only know the gradient approxi-
3948 mately. In such cases the momentum term is useful since it averages out
3949 different noisy estimates of the gradient. One particularly useful way to
3950 obtain an approximate gradient is using a stochastic approximation, which
3951 we discuss next.
3952 7.1.3 Stochastic Gradient Descent

3953 Computing the gradient can be very time consuming. However, often it is
3954 possible to find a “cheap” approximation of the gradient. Approximating
3955 the gradient is still useful as long as it points in roughly the same direction
Stochastic gradient3956 as the true gradient.
descent 3957 Stochastic gradient descent (often shortened in SGD) is a stochastic ap-
3958 proximation of the gradient descent method for minimizing an objective
3959 function that is written as a sum of differentiable functions. The word
3960 stochastic here refers to the fact that we acknowledge that we do not
3961 know the gradient precisely, but instead only know a noisy approxima-
3962 tion to it. By constraining the probability distribution of the approximate
3963 gradients, we can still theoretically guarantee that SGD will converge.
In machine learning given n = 1, . . . , N data points, we often consider
objective functions which are the sum of the losses Ln incurred by each
example n. In mathematical notation we have the form
N
X
L(✓) = Ln (✓) (7.13)
n=1
where ✓ is the vector of parameters of interest, i.e., we want to find ✓ that

minimizes L. An example from regression (Chapter 9), is the negative log-
likelihood, which is expressed as a sum over log-likelihoods of individual
examples,
N
X
L(✓) = log p(yn |xn , ✓) (7.14)
n=1
3964 where xn 2 RD are the training inputs, yn are the training targets and ✓
3965 are the parameters of the regression model.
Standard gradient descent, as introduced previously, is a “batch” opti-
mization method, i.e., optimization is performed using the full training set
by updating the vector of parameters according to

N
X
✓ i+1 = ✓ i i (rL(✓ i ))
>
= ✓i i (rLn (✓ i ))> (7.15)
n=1
3966 for a suitable stepsize parameter i . Evaluating the sum-gradient may re-
3967 quire expensive evaluations of the gradients from all individual functions
3968 Ln . When the training set is enormous and/or no simple formulas exist,
3969 evaluating the sums of gradients becomes very expensive.
PN
3970 Consider the term n=1 (rLn (✓ i )) in (7.15) above: we can reduce the
3971 amount of computation by taking a sum over a smaller set of Ln . In con-
3972 trast to batch gradient descent, which uses all Ln for n = 1, . . . , N , we
3973 randomly choose a subset of Ln for mini-batch gradient descent. In the
3974 extreme case, we randomly select only a single Ln to estimate the gra-
3975 dient. The key insight about why taking a subset of data is sensible is
3976 to realise that for gradient descent to converge, we only require that the
3977 gradient
PN is an unbiased estimate of the true gradient. In fact the term
3978
n=1 (rLn (✓ i )) in (7.15) is an empirical estimate of the expected value
3979 (Section 6.4.1) of the gradient. Therefore any other unbiased empirical
3980 estimate of the expected value, for example using any subsample of the
3981 data, would suffice for convergence of gradient descent.
3982 Why should one consider using an approximate gradient? A major rea-
3983 son is practical implementation constraints, such as the size of CPU/GPU
3984 memory or limits on computational time. We can think of the size of the
3985 subset used to estimate the gradient in the same way that we thought of
3986 the size of a sample when estimating empirical means 6.4.1. In practice,
3987 it is good to keep the size of the mini-batch as large as possible. Large
3988 mini-batches reduce the variance in the parameter update. Furthermore This often leads to
3989 large mini-batches take advantage of highly optimized matrix operations more stable
convergence since
3990 in vectorized implementations of the cost and gradient. However when
the gradient
3991 we choose the mini-batch size, we need to make sure it fits into CPU/GPU estimator is less
3992 memory. Typical mini-batch sizes are 64, 128, 256, 512, 1024, which de- noisy.
3993 pends on the way computer memory is laid out and accessed.
3994 Remark. When the learning rate decreases at an appropriate rate, and sub-
3995 ject to relatively mild assumptions, stochastic gradient descent converges
3996 almost surely to local minimum (Bottou, 1998). }
3997 If we keep the mini-batch size small, the noise in our gradient estimate
3998 will allow us to get out of some bad local optima, which we may otherwise
3999 get stuck in.
4000 Stochastic gradient descent is very effective in large-scale machine learn-
4001 ing problems (Bottou et al., 2018), such as training deep neural networks
4002 on millions of images (Dean et al., 2012), topic models (Hoffman et al.,
4003 2013), reinforcement learning (Mnih et al., 2015) or training large-scale
4004 Gaussian process models (Hensman et al., 2013; Gal et al., 2014).
Figure 7.4
Illustration of
constrained
optimization. The
unconstrained
problem (indicated
by the contour
lines) has a
minimum on the
right side (indicated
by the circle). The
box constraints
( 1 6 x 6 1 and
1 6 y 6 1) require
that the optimal
solution are within
the box, resulting in
an optimal value
indicated by the
star.
4005 7.2 Constrained Optimization and Lagrange Multipliers

In the previous section, we considered the problem of solving for the min-
imum of a function
min f (x) (7.16)
x
4006 where f : RD ! R.
In this section we have additional constraints. That is for real valued
functions gi : RD ! R for i = 1, . . . , m we consider the constrained
optimization problem
min f (x) (7.17)
x
subject to gi (x) 6 0 for all i = 1, . . . , m

4007 It is worth pointing out that the functions f and gi could be non-convex
4008 in general, and we will consider the convex case in the next section.
One obvious, but not very practical, way of converting the constrained
problem (7.17) into an unconstrained one is to use an indicator function
m
X
J(x) = f (x) + 1(gi (x)) (7.18)
i=1
where 1(z) is an infinite step function

(
0 if z 6 0
1(z) = . (7.19)
1 otherwise
7.2 Constrained Optimization and Lagrange Multipliers 221
4009 This gives infinite penalty if the constraint is not satisfied, and hence
4010 would provide the same solution. However, this infinite step function is
4011 equally difficult to optimize. We can overcome this difficulty by introduc-
4012 ing Lagrange multipliers. The idea of Lagrange multipliers is to replace the Lagrange multipliers
4013 step function with a linear function.
We associate to problem (7.17) the Lagrangian by introducing the La- Lagrangian
grange multipliers i > 0 corresponding to each inequality constraint re-
spectively (Boyd and Vandenberghe, 2004, Chapter 4).
m
X
L(x, ) = f (x) + i gi (x)
i=1
>
= f (x) + g(x) (7.20)
4014 where in the last line we have a concatenated all constraints gi (x) into a
4015 vector g(x), and all the Lagrange multipliers into a vector 2 Rm .
4016 We now introduce the idea of Lagrangian duality. In general, duality
4017 in optimization is the idea of converting an optimization problem in one
4018 set of variables x (called the primal variables), into another optimization
4019 problem in a different set of variables (called the dual variables). We
4020 introduce two different approaches to duality: in this section we discuss
4021 Lagrangian duality, and in Section 7.3.3 we discuss Legendre-Fenchel du-
4022 ality.
Theorem 7.1. The problem in (7.17)
min f (x)
x

is known as the primal problem, corresponding to the primal variables x. primal problem
The associated Lagrangian dual problem is given by Lagrangian dual
problem
maxm D( ) (7.21)
2R
subject to > 0, (7.22)

4023 where are the dual variables and D( ) = minx2Rd L(x, ).
Proof Recall that the difference between J(x) in (7.18) and the La-
grangian in (7.20) is that we have relaxed the indicator function to a
linear function. Therefore when > 0, the Lagrangian L(x, ) is a lower
bound of J(x). Hence the maximum of L(x, ) with respect to is J(x)
J(x) = max L(x, ). (7.23)
>0
Recall that the original problem was minimising J(x),

min max L(x, ). (7.24)
x2Rd >0
By the minimax inequality (Boyd and Vandenberghe, 2004) it turns out
that, for any function swapping the order of the minimum and maximum
above results in a smaller value.
min max L(x, ) > max mind L(x, ) (7.25)
x2Rd >0 >0 x2R
weak duality 4024 This is also known as weak duality. Note that the inner part of the right
4025 hand side is the dual objective function D( ) and the theorem follows.
4026 In contrast to the original optimization problem which has constraints,
4027 minx2Rd L(x, ) is an unconstrained optimization problem for a given
4028 value of . If solving minx2Rd L(x, ) is easy, then the overall problem
4029 is easy to solve. The reason is that the outer problem (maximization over
4030 ) is a maximum over a set of affine functions, and hence is a concave
4031 function, even though f (·) and gi (·) may be non-convex. The maximum
4032 of a concave function can be efficiently computed.
4033 Assuming f (·) and gi (·) are differentiable, we find the Lagrange dual
4034 problem by differentiating the Lagrangian with respect to x and setting
4035 the differential to zero and solving for the optimal value. We will discuss
4036 two concrete examples in Section 7.3.1 and 7.3.2, where f (·) and gi (·)
4037 are convex.
Remark (Equality constraints). Consider (7.17) with additional equality
constraints
minf (x) (7.26)
x

hj (x) = 0 for all j = 1, . . . , n
4038 We can model equality constraints by replacing them with two inequality
4039 constraints. That is for each equality constraint hj (x) = 0 we equivalently
4040 replace it by two constraints hj (x) 6 0 and hj (x) > 0. It turns out that
4041 the resulting Lagrange multipliers are then unconstrained.
4042 Therefore we constrain the Lagrange multipliers corresponding to the
4043 inequality constraints in (7.26) to be non-negative, and leave the La-
4044 grange multipliers corresponding to the equality constraints unconstrained.
4045 }
4046 7.3 Convex Optimization

4047 We focus our attention of a particularly useful class of optimization prob-
4048 lems, where we can guarantee global optimality. When f (·) is a convex
4049 function, and when the constraints involving g(·) and h(·) are convex sets,
convex optimization
4050 this is called a convex optimization problem. In this setting, we have strong
problem 4051 duality: The optimal solution of the dual problem is the same as the opti-
strong duality 4052 mal solution of the primal problem. The distinction between convex func-
convex functions 4053 tions and convex sets are often not strictly presented in machine learning
convex sets
4054 literature, but one can often infer the implied meaning from context.
7.3 Convex Optimization 223
Figure 7.5 Example

of a convex
40 function.
y = 3x2 5x + 2
x2
20
0
2 0 2
x1
4055 Convex functions are functions such that a straight line between any
4056 two points of the function lie above the function. Figure 7.2 shows a non-
4057 convex function and Figure 7.3 shows a convex function. Another convex
4058 function is shown in Figure 7.5.
Definition 7.2. A function f : RD ! R is a convex function if for all x, y convex function
in the domain of f , and for any scalar ✓ with 0 6 ✓ 6 1, we have Technically, the
domain of the
f (✓x + (1 ✓)y) 6 ✓f (x) + (1 ✓)f (y) (7.27) function f must also
be a convex set.
4059 Remark. A concave function is the negative of a convex function. } concave function
4060 The constraints involving g(·) and h(·) in (7.26) truncate functions at a Figure 7.6 Example
of a convex set
4061 scalar value, resulting in sets. Another relation between convex functions
4062 and convex sets is to consider the set obtained by “filling in” a convex
4063 function. A convex function is a bowl like object, and we imagine pouring
4064 water into it to fill it up. This resulting filled in set, called the epigraph
4065 of the convex function, is a convex set. Convex sets are sets such that a
4066 straight line connecting any two elements of the set lie inside the set. Fig-
4067 ure 7.6 and Figure 7.7 illustrates convex and nonconvex sets respectively.
Definition 7.3. A set C is a convex set if for any x, y 2 C and for any Figure 7.7 Example
scalar ✓ with 0 6 ✓ 6 1, we have of a nonconvex set
✓x + (1 ✓)y 2 C (7.28)
If a function f : Rn ! R is differentiable, we can specify convexity in
terms of its gradient rx f (x) (Section 5.2). A function f (x) is convex if
and only if for any two points x, y ,
f (y) > f (x) + rx f (x)> (y x). (7.29)
convex set
4068 If we further know that a function f (x) is twice differentiable, that is the
4069 Hessian (5.144) exists for all values in the domain of x, then the function
4070 f (x) is convex if and only if r2x f (x) is positive semi-definite (Boyd and
4071 Vandenberghe, 2004).
Example 7.3
Figure 7.8 The

negative entropy
function (which is
10 tangent at x = 2
convex), and its x log2 x
tangent at x = 2.
f (x)
0 2 4
t
The negative entropy f (x) = x log2 x is convex for x > 0. A visualiza-
tion of the function is show in Figure 7.8, and we can see that the function
is convex. To illustrate the above definitions of convexity, let us check the
calculations for two points x = 2 and x = 4. Note that to prove convexity
of f (x) we would need to check for all points x 2 R.
Recall Definition 7.2. Consider a point midway between the two points
(that is ✓ = 0.5), then the left hand side is f (0.5⇥2+0.5⇥4) = 3 log2 3 ⇡
4.75. The right hand side is 0.5(2 log2 2) + 0.5(4 log2 4) = 1 + 4 = 5. And
therefore the definition is satisfied.
Since f (x) is differentiable, we can alternatively use (7.29). Calculating
the derivative of f (x), we obtain
1
rx (x log2 x) = 1 ⇥ log2 x + x ⇥ (7.30)
x
= log2 x + 1. (7.31)
Using the same two test points x = 2 and x = 4, the left hand side of
(7.29) is given by f (4) = 8. The right hand side is
f (x) + r>
x (y x) = f (2) + rf (2) ⇥ (4 2) (7.32)
= 2 + 2 ⇥ 2 = 6. (7.33)
4072 We can check that a function or set is convex from first principles by
4073 recalling the definitions. In practice we often rely on operations that pre-
4074 serve convexity to check that a particular function or set is convex. Al-
4075 though the details are vastly different, this is again the idea of closure
4076 that we introduced in Chapter 2 for vector spaces.
Example 7.4
A nonnegative weighted sum of convex functions is convex. Observe that
if f is a convex function, and ↵ > 0 is a nonnegative scalar, then the
function ↵f is convex. We can see this by multiplying ↵ to both sides of
equation in Definition 7.2, and recalling that multiplying a nonnegative
number does not change the inequality.
If f1 and f2 are convex functions, then we have by the definition
f1 (✓x + (1 ✓)y) 6 ✓f1 (x) + (1 ✓)f1 (y) (7.34)
f2 (✓x + (1 ✓)y) 6 ✓f2 (x) + (1 ✓)f2 (y). (7.35)
Summing up both sides gives us
f1 (✓x + (1 ✓)y) + f2 (✓x + (1 ✓)y)
6 ✓f1 (x) + (1 ✓)f1 (y) + ✓f2 (x) + (1 ✓)f2 (y) (7.36)
where the right hand side can be rearranged to
✓(f1 (x) + f2 (x)) + (1 ✓)(f1 (y) + f2 (y)) (7.37)
completing the proof that the sum of convex functions is convex.
Combining the two facts above, we see that ↵f1 (x) + f2 (x) is convex
for ↵, > 0. This closure property can be extended using a similar argu-
ment for nonnegative weighted sums of more than two convex functions.
4077 Remark. The inequality defining convex functions, see 7.27, is sometimes
4078 called Jensen’s inequality. In fact a whole class of inequalities for taking Jensen’s inequality
4079 nonnegative weighted sums of convex functions are all called Jensen’s
4080 inequality. }
In summary, a constrained optimization problem is called a convex opti- convex optimization
mization problem if problem
minf (x) (7.38)

x

hj (x) = 0 for all j = 1, . . . , n
4081 where all the functions f (x) and gi (x) are convex functions, and all
4082 hj (x) = 0 are convex sets. In the following two subsections, we will de-
4083 scribe two classes convex optimization problems that are widely used and
4084 well understood.
4085 7.3.1 Linear Programming

Consider the special case when all the functions above are linear, that is
min c> x (7.39)
x2Rd
Figure 7.9
Illustration of a
linear program. The
unconstrained
problem (indicated
by the contour
lines) has a
minimum on the
right side. The
optimal value given
the constraints are
shown by the star.
subject to Ax 6 b
Linear programs are

4086 where A 2 Rm⇥d and b 2 Rm . This is known as a linear program. It has d
one of the most 4087 variables and m linear constraints.
widely used
approaches in
industry.
Example 7.5
An example of a linear program is illustrated in Figure 7.9, which has
two variables. The objective function is linear, resulting in linear contour
lines. The constraint set in standard form is translated into the legend. The
optimal value must lie in the shaded (feasible) region, and is indicated by
the star.
 > 
5 x1
min2 (7.40)
x2R 3 x2
2 3 2 3
2 2 33
6 2 47  6 8 7
6 7 x1 6 7
subject to 6
6 2 17 6 7
7 x2 6 6 5 7 (7.41)
4 0 15 4 15
0 1 8
The Lagrangian is given by

>
L(x, ) = c> x + (Ax b)
where 2 Rm is the vector of non-negative Lagrange multipliers. It is

easier to see what is going on by rearranging the terms corresponding to
x.
>
L(x, ) = (c + A> )> x b
Taking the derivative of L(x, ) with respect to x and setting it to zero
gives us
c + A> = 0.
>
Therefore the dual Lagrangian is D( ) = b. Recall we would like
to maximize D( ). In addition to the constraint due to the derivative of It is convention to
L(x, ) being zero, we also have the fact that > 0, resulting in the minimize the primal
and maximize the
following dual optimization problem
dual.
maxm b> (7.42)
2R
subject to c + A> = 0
> 0.
4088 This is also a linear program, but with m variables. We have the choice
4089 of solving the primal (7.39) or the dual (7.42) program depending on
4090 whether m or d is larger. Recall that d is the number of variables and m is
4091 the number of constraints in the primal linear program.
4092 7.3.2 Quadratic Programming

Consider when the objective function is a convex quadratic function, and
the constraints are affine,
1
mind x> Qx + c> x (7.43)
x2R 2
subject to Ax 6 b
4093 where A 2 Rm⇥d , b 2 Rm and c 2 Rd . The square symmetric matrix Q 2
4094 Rd⇥d is positive definite, and therefore the objective function is convex.
4095 This is known as a quadratic program. Observe that it has d variables and
4096 m linear constraints.
Example 7.6
An example of a quadratic program is illustrated in Figure 7.4, which has
two variables. The objective function is quadratic with a positive semidefi-
nite matrix Q, resulting in elliptical contour lines. The optimal value must
lie in the shaded (feasible) region, and is indicated by the star.
 >    > 
1 x1 2 1 x1 5 x1
min + (7.44)
x2R2 2 x2 1 4 x2 3 x2
2 3 2 3
1 0  1
6 1 7
0 7 x1 6 17
subject to 6
4 0 5 66
4
7 (7.45)
1 x2 15
0 1 1
The Lagrangian is given by

1 >
L(x, ) = x Qx + c> x + > (Ax b)
2
1 >
= x> Qx + (c + A> )> x b
2
where again we have rearranged the terms. Taking the derivative of L(x, )
with respect to x and setting it to zero gives
Qx + (c + A> ) = 0
Assuming that Q is invertible, we get
x= Q 1 (c + A> ) (7.46)
Substituting (7.46) into the primal Lagrangian L(x, ) we get the dual
Lagrangian
1 >
D( ) = (c + A> )Q 1 (c + A> ) b
2
Therefore the dual optimization problem is given by
1 >
maxm (c + A> )Q 1 (c + A> ) b (7.47)
2R 2
subject to > 0. (7.48)
4097 We will see an application of Quadratic Programming in machine learning
4098 in Chapter 12.
4099 7.3.3 Legendre-Fenchel Transform and Convex Conjugate

4100 Let us revisit the idea of duality, which we saw in Section 7.2, without
4101 considering constraints. One useful fact about convex sets is that a convex
4102 set can be equivalently described by its supporting hyperplanes. A hyper-
supporting 4103 plane is called a supporting hyperplane of a convex set if it intersects the
hyperplane 4104 convex set and the convex set is contained on just one side of it. Recall
4105 that for a convex function, we can fill it up to obtain the epigraph which
4106 is a convex set. Therefore we can also describe convex functions in terms
4107 of their supporting hyperplanes. Furthermore observe that the supporting
4108 hyperplane just touches the convex function, and is in fact the tangent to
4109 the function at that point. And recall that the tangent of a function f (x) at
4110 a given point x0 is the evaluation of the gradient of that function at that
4111 point dfdx
(x)
. In summary, because convex sets can be equivalently
x=x0
4112 described by its supporting hyperplanes, convex functions can be equiv-
4113 alently described by a function of their gradient. The Legendre transform Legendre transform
4114 formalizes this concept . Physics students are
4115 We begin with the most general definition which unfortunately has a often introduced to
the Legendre
4116 counterintuitive form, and look at special cases to try to relate the defini-
transform as
4117 tion to the intuition above. The Legendre-Fenchel transform is a transfor- relating the
4118 mation (in the sense of a Fourier transform) from a convex differentiable Lagrangian and the
4119 function f (x) to a function that depends on the tangents s(x) = rx f (x). Hamiltonian in
classical mechanics.
4120 It is worth stressing that this is a transformation of the function f (·) and
Legendre-Fenchel
4121 not the variable x or the function evaluated at a value. The Legendre- transform
4122 Fenchel transform is also known as the convex conjugate (for reasons
4123 we will see soon) and is closely related to duality (Hiriart-Urruty and
4124 Lemaréchal, 2001, Chapter 5).
Definition 7.4. The convex conjugate of a function f : RD ! R is a convex conjugate
function f ⇤ defined by
f ⇤ (s) = sup hs, xi f (x) (7.49)
x2RD
4125 Note that the convex conjugate definition above does not need the func-
4126 tion f to be convex nor differentiable. In the definition above, we have
4127 used a general inner product (Section 3.2) but in the rest of this section
4128 we will consider the standard dot product between finite dimensional vec-
4129 tors (hs, xi = s> x) to avoid too many technical details.
To understand the above definition in a geometric fashion, consider an This derivation is
nice simple one dimensional convex and differentiable function, for ex- easiest to
understand by
ample f (x) = x2 . Note that since we are looking at a one dimensional
drawing the
problem, hyperplanes reduce to a line. Consider a line y = sx + c. Recall reasoning as it
that we are able to describe convex functions by their supporting hyper- progresses.
planes, so let us try to describe this function f (x) by its supporting lines.
Fix the gradient of the line s 2 R and for each point (x0 , f (x0 )) on the
graph of f , find the minimum value of c such that the line still inter-
sects (x0 , f (x0 )). Note that the minimum value of c is the place where a
line with slope s “just touches” the function f (x) = x2 . The line passing
through (x0 , f (x0 )) with gradient s is given by
y f (x0 ) = s(x x0 ) (7.50)
The y -intercept of this line is sx0 + f (x0 ). The minimum of c for which
y = sx + c intersects with the graph of f is therefore
inf sx0 + f (x0 ). (7.51)
x0
4130 The convex conjugate above is by convention defined to be the negative

4131 of this. The reasoning in this paragraph did not rely on the fact that we
4132 chose a one dimensional convex and differentiable function, and holds for
4133 f : RD ! R which are nonconvex and non differentiable.
The classical
Legendre transform Remark. Convex differentiable functions such as the example f (x) = x2
is defined on convex is a nice special case, where there is no need for the supremum, and there
differentiable
functions in RD
is a one to one correspondence between a function and its Legendre trans-
form. Let us derive this from first principles. For a convex differentiable
function, we know that at x0 the tangent touches f (x0 ), therefore
f (x0 ) = sx0 + c. (7.52)
Recall that we want to describe the convex function f (x) in terms of its
gradient rx f (x), and that s = rx f (x0 ). We rearrange to get an expres-
sion for c to obtain
c = sx0 f (x0 ). (7.53)
Note that c changes with x0 and therefore with s, which is why we can
think of it as a function of s, which we call f ⇤ (s).
f ⇤ (s) = sx0 f (x0 ). (7.54)
4134 Compare (7.54) with Definition 7.4, and observe that (7.54) is a special
4135 case (without the supremum). }
4136 The conjugate function has nice properties, for example for convex
4137 functions, applying the Legendre transform again gets us back to the origi-
4138 nal function. In the same way that the slope of f (x) is s, the slope of f ⇤ (s)
4139 is x. The following two examples show common uses of convex conjugates
4140 in machine learning.
Example 7.7
To illustrate the application of convex conjugates, consider the quadratic
function based on a positive definite matrix K 2 Rn⇥n . We denote the
primal variable to be y 2 Rn and the dual variable to be ↵ 2 Rn .
y>K 1y
f (y) = (7.55)
2
Applying Definition 7.4, we obtain the function
1
f ⇤ (↵) = sup hy, ↵i y>K y. (7.56)
y2Rn 2
Observe that the function is differentiable, and hence we can find the
maximum by taking the derivative and with respect to y setting it to zero.
⇥ ⇤
@ hy, ↵i 2 y > K 1 y
= (↵ K 1 y)> (7.57)
@y
and hence when the gradient is zero we have y = 1
K↵. Substituting
into (7.56) yields

✓ ◆> ✓ ◆
1 1 1 1
⇤
f (↵) = ↵ K↵>
K↵ K K↵ (7.58)
2
1 >
= ↵ K↵ . (7.59)
2
Example 7.8
In machine learning we often use sums of functions, for example the ob-
jective function of the training set includes a sum of the losses for each ex-
ample in the training set. In the following, we derive the convex conjugate
of a sum of losses `(t), where ` : R ! R. This also illustratesPnthe applica-
tion of the convex conjugate to the vector case. Let L(t) = i=1 ì (ti ),
n
X
L⇤ (z) = sup hz, ti ì (ti ) (7.60)
t2Rn i=1
n
X
= sup zi t i ì (ti ) definition of dot product (7.61)
t2Rn i=1
Xn
= sup zi ti ì (ti ) (7.62)
n
i=1 t2R
Xn
= `⇤i (zi ) definition of conjugate (7.63)
i=1
4141 Recall that in Section 7.2 we derived a dual optimization problem using
4142 Lagrange multipliers. Furthermore for convex optimization problems we
4143 have strong duality, that is the solutions of the primal and dual problem
4144 match. The Fenchel-Legendre transform described here also can be used
4145 to derive a dual optimization problem. Furthermore then the function is
4146 convex and differentiable, the supremum is unique. To further investigate
4147 the relation between these two approaches, let us consider a linear equal-
4148 ity constrained convex optimization problem.
Example 7.9
Let f (y) and g(x) be convex functions, and A a real matrix of appropriate
dimensions such that Ax = y . Then
min f (Ax) + g(x) = min f (y) + g(x). (7.64)
x Ax=y
By introducing the Lagrange multiplier u for the constraints Ax = y ,

min f (y) + g(x) = min max f (y) + g(x) + (Ax y)> u (7.65)
Ax=y x,y u
= max min f (y) + g(x) + (Ax y)> u (7.66)

u x,y
where the last step of swapping max and min is due to the fact that f (y)
and g(x) are convex functions. By splitting up the dot product term and
collecting x and y ,
max min f (y) + g(x) + (Ax y)> u (7.67)
u x,y
 h i
= max min y > u + f (y) + min(Ax)> u + g(x) (7.68)
u y x
 h i
= max min y > u + f (y) + min x> A> u + g(x) (7.69)
u y x
For general inner Recall the convex conjugate (Definition 7.4) and the fact that dot prod-
products, A> is
ucts are symmetric,
replaced by the
 h i
adjoint A⇤ .
max min y > u + f (y) + min x> A> u + g(x) (7.70)
u y x
= max f ⇤ (u) g ⇤ ( A> u). (7.71)

u
Therefore we have shown that

min f (Ax) + g(x) = max f ⇤ (u) g ⇤ ( A> u). (7.72)
x u
4149 The Legendre-Fenchel conjugate turns out to be quite useful for ma-
4150 chine learning problems that can be expressed as convex optimization
4151 problems. In particular for convex loss functions that apply independently
4152 to each example, the conjugate loss is a convenient way to derive a dual
4153 problem.
4154 7.4 Further Reading

4155 Continuous optimization is an active area of research, and we do not try
4156 to provide a comprehensive account of recent advances.
4157 From a gradient descent perspective, there are two major weaknesses
4158 which each have their own set of literature. The first challenge is the fact
4159 that gradient descent is a first order algorithm, and does not use informa-
4160 tion about the curvature of the surface. When there are long valleys, the
4161 gradient points perpendicularly to the direction of interest. Conjugate gra-
4162 dient methods avoid the issues faced by gradient descent by taking pre-
4163 vious directions into account (Shewchuk, 1994). Second order methods
Exercises 233
4164 such as Newton methods use the Hessian to provide information about the
4165 curvature. Many of the choices for choosing stepsizes and ideas like mo-
4166 mentum arise by considering the curvature of the objective function (Goh,
4167 2017; Bottou et al., 2018). Quasi-Newton methods such as L-BFGS try to
4168 use cheaper computational methods to approximate the Hessian (Nocedal
4169 and Wright, 2006). Recently there has been interest in other metrics for
4170 computing descent directions, resulting in approaches such as mirror de-
4171 scent (Beck and Teboulle, 2003) and natural gradient (Toussaint, 2012).
4172 The second challenge are non-differentiable functions. Gradient meth-
4173 ods are not well defined when there are kinks in the function. In these
4174 cases, subgradient methods can be used (Shor, 1985). For further infor-
4175 mation and algorithms for optimizing non-differentiable functions, we re-
4176 fer to the book by Bertsekas (1999). There is a vast amount of literature
4177 on different approaches for numerically solving continuous optimization
4178 problems, including algorithms for constrained optimization problems. A
4179 good starting point to appreciate this literature are Luenberger (1969)
4180 and Bonnans et al. (2006). A recent survey of continuous optimization is
4181 Bubeck (2015).
4182 Modern applications of machine learning often mean that the size of
4183 datasets prohibit the use of batch gradient descent, and hence stochastic
4184 gradient descent is the current workhorse of large scale machine learning
4185 methods. Recent surveys of the literature include (Hazan, 2015; Bottou
4186 et al., 2018).
4187 For duality and convex optimization, the book by Boyd and Vanden-
4188 berghe (Boyd and Vandenberghe, 2004) includes lectures and slides on-
4189 line. A more mathematical treatment is provided by Bertsekas (2009).
4190 Convex optimization is based upon convex analysis, and the reader inter-
4191 ested in more foundational results about convex functions is referred to
4192 Hiriart-Urruty and Lemaréchal (2001); Rockafellar (1970); Borwein and
4193 Lewis (2006). Legendre-Fenchel transforms are also covered in the above
4194 books on convex analysis, but more beginner friendly presentations are
4195 available at Zia et al. (2009); Gonçalves (2014). The role of Legendre-
4196 Fenchel transforms in the analysis of convex optimization algorithms is
4197 surveyed in Polyak (2016).
4198 Exercises
7.1 Consider the univariate function
f (x) = x3 + 2x2 + 5x 3.
4199 Find its stationary points and indicate whether they are maximum, mini-
4200 mum or saddle points.
4201 7.2 Consider the update equation for stochastic gradient descent (Equation (7.15)).
4202 Write down the update when we use a mini-batch size of one.
7.3 Express the following optimization problem as a standard linear program in
matrix notation
max p> x + ⇠
x2R2 ,⇠2R
4203 subject to the constraints that ⇠ > 0, x0 6 0 and x1 6 3.

7.4 The hinge loss (which is the loss used by the Support Vector Machine) is
given by
L(↵) = max{0, 1 ↵}
If we are interested in applying gradient methods such as L-BFGS, and do

not want to resort to subgradient methods, we need to smooth the kink in
the hinge loss. Compute the convex conjugate of the hinge loss L⇤ ( ) where
is the dual variable. Add a `2 proximal term, and compute the conjugate
of the resulting function
L⇤ ( ) + 2
2
4204 where is a given hyperparameter.

Continuous Optimization

Uploaded by

Copyright:

Available Formats

Continuous Optimization

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Continuous Optimization

Uploaded by

Copyright:

Available Formats

3841

3842 Since machine learning algorithms are implemented on a computer, the

`(x) = x4 + 7x3 + 5x2 17x + 3. (7.1)

Continuous Figure 7.1 A mind

Lagrange multipliers Convex

Its gradient is given by

Figure 7.2 Example

3889 7.1 Optimization using Gradient Descent

min f (x) (7.4)

Figure 7.3 Gradient

3890 where f : Rd ! R is an objective function that captures the machine

3911 7.1.1 Stepsize

Example 7.2 (Solving a Linear Equation System)

3927 Remark. When applied to the solution of linear systems of equations Ax =

3940 7.1.2 Gradient Descent with Momentum

ment a moving average. The momentum-based method remembers the

xi+1 = xi i ((rf )(xi ))

3952 7.1.3 Stochastic Gradient Descent

where ✓ is the vector of parameters of interest, i.e., we want to find ✓ that

by updating the vector of parameters according to

4005 7.2 Constrained Optimization and Lagrange Multipliers

subject to gi (x) 6 0 for all i = 1, . . . , m

where 1(z) is an infinite step function

subject to gi (x) 6 0 for all i = 1, . . . , m

subject to > 0, (7.22)

Recall that the original problem was minimising J(x),

By the minimax inequality (Boyd and Vandenberghe, 2004) it turns out

subject to gi (x) 6 0 for all i = 1, . . . , m

4046 7.3 Convex Optimization

Figure 7.5 Example

Figure 7.8 The

minf (x) (7.38)

subject to gi (x) 6 0 for all i = 1, . . . , m

4085 7.3.1 Linear Programming

Linear programs are

The Lagrangian is given by

where 2 Rm is the vector of non-negative Lagrange multipliers. It is

4092 7.3.2 Quadratic Programming

The Lagrangian is given by

4099 7.3.3 Legendre-Fenchel Transform and Convex Conjugate

4130 The convex conjugate above is by convention defined to be the negative

into (7.56) yields

By introducing the Lagrange multiplier u for the constraints Ax = y ,

= max min f (y) + g(x) + (Ax y)> u (7.66)

= max f ⇤ (u) g ⇤ ( A> u). (7.71)

Therefore we have shown that

4154 7.4 Further Reading

4203 subject to the constraints that ⇠ > 0, x0 6 0 and x1 6 3.

If we are interested in applying gradient methods such as L-BFGS, and do

You might also like