17.

Constrained least squares applications

 Outline

Portfolio optimization

Linear quadratic control

Linear quadratic state estimation

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.1

 Portfolio allocation weights

I we invest a total of V dollars in n different assets (stocks, bonds, . . . ) over

some period (one day, week, month, . . . )
I can include short positions, assets you borrow and sell at the beginning,
but must return to the borrower at the end of the period
I portfolio allocation weight vector w gives the fraction of our total portfolio
value held in each asset
I Vwj is the dollar value of asset j you hold
I 1T w = 1, with negative wi meaning a short position
I w = ( 0.2, 0.0, 1.2) means we take a short position of 0.2V in asset 1,
don’t hold any of asset 2, and hold 1.2V in asset 3

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.2

 Leverage, long-only portfolios, and cash

I leverage is L = |w1 | + · · · + |wn |

( (L 1)/2 is also sometimes used)

I L = 1 when all weights are nonnegative (‘long only portfolio’)

I w = 1/n is called the uniform portfolio

I we often assume asset n is ‘risk-free’ (or cash or T-bills)

I so w = en means the portfolio is all cash

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.3

 Return over a period

I r̃j is the return of asset j over the period

I r̃j is the fractional increase in price or value (decrease if negative)
I often expressed as a percentage, like +1.1% or 2.3%
I full portfolio return is
V+ V
= r̃T w
V
where V + is the portfolio value at the end of the period
I if you hold portfolio for t periods with returns r1 , . . . , rt value is

Vt+1 = V1 (1 + r1 )(1 + r2 ) · · · (1 + rt )

I portfolio value versus time traditionally plotted using V1 = $10000

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.4

 Return matrix

I hold portfolio with weights w over T periods

I define T ⇥ n (asset) return matrix, with Rtj the return of asset j in period t

I row t of R is r̃tT , where r̃t is the asset return vector over period t

I column j of R is time series of asset j returns

I portfolio returns vector (time series) is T -vector r = Rw

I if last asset is risk-free, the last column of R is µrf 1, where µrf is the
risk-free per-period interest rate

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.5

 Portfolio return and risk

I r is time series (vector) of portfolio returns

I average return or just return is avg(r)

I risk is std(r)

I these are the per-period return and risk

I for small per-period returns we have

VT+1 = V1 (1 + r1 ) · · · (1 + rT )
⇡ V1 + V1 (r1 + · · · + rT )
= V1 + T avg(r)V1

I so return approximates the average per-period increase in portfolio value

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.6

 Annualized return and risk

I mean return and risk are often expressed in annualized form (i.e., per year)

I if there are P trading periods per year

p
annualized return = P avg(r), annualized risk = P std(r)

(the squareroot in risk annualization comes from the assumption that the
fluctuations in return around the mean are independent)

I if returns are daily, with 250 trading days in a year

p
annualized return = 250 avg(r), annualized risk = 250 std(r)

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.7

 Portfolio optimization

I how should we choose the portfolio weight vector w?

I we want high (mean) portfolio return, low portfolio risk

I we know past realized asset returns but not future ones

I we will choose w that would have worked well on past returns

I . . . and hope it will work well going forward (just like data fitting)

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.8

 Portfolio optimization

minimize std(Rw) 2 = (1/T)kRw ⇢1k 2

subject to 1T w = 1
avg(Rw) = ⇢

I w is the weight vector we seek

I R is the returns matrix for past returns
I Rw is the (past) portfolio return time series
I require mean (past) return ⇢
I we minimize risk for specified value of return
I solutions w are Pareto optimal

I we are really asking what would have been the best constant allocation,
had we known future returns

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.9

 Portfolio optimization via constrained least squares

minimize kRw ⇢1k 2

" T # " #
1 1
subject to w=
µT ⇢

I µ = RT 1/T is n-vector of (past) asset returns

I ⇢ is required (past) portfolio return
I an equality constrained least squares problem, with solution

26 w 37 26 2RT R 1 µ 37 1 26 2⇢T µ 37
66 z 77 = 66 1T 77 66 1 77
66 1 77 66 0 0 77 66 77
4 z2 5 4 µ 5 4 ⇢ 5
T 0 0

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.10

 Optimal portfolios

I perform significantly better than individual assets

I risk-return curve forms a straight line

I one end of the line is the risk-free asset

I two-fund theorem: optimal portfolio w is an affine function of ⇢

26 w 37 26 2RT R 1 µ 37 1 26 0 37 26 2RT R 1 µ 37 1 26 2T µ 37
66 z 77 = 66 1T 77 66 1 77 + ⇢ 66 1T 77 66 0 77
66 1 77 66 0 0 77 66 77 66 0 0 77 66 77
4 z2 5 4 µ 5 4 5 4 µ 5 4 1 5
T 0 0 0 T 0 0

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.11

 The big assumption

I now we make the big assumption (BA):

– you are warned this is false, every time you invest

– it is often reasonably true
– in periods of ‘market shift’ it’s much less true

I if BA holds (even approximately), then a good weight vector for past

(realized) returns should be good for future (unknown) returns
I for example:
– choose w based on last 2 years of returns
– then use w for next 6 months

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.12

 Example

20 assets over 2000 days

0.4
Annualized return

0.3

0.2

0.1

0
risk-free asset

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Annualized risk

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.13

 Pareto optimal portfolios

0.4
Annualized return

0.3

0.2

0.1

1/n
0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Annualized risk

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.14

 Five portfolios

Return Risk
Portfolio Train Test Train Test Leverage
risk-free 0.01 0.01 0.00 0.00 1.00
⇢ = 10% 0.10 0.08 0.09 0.07 1.96
⇢ = 20% 0.20 0.15 0.18 0.15 3.03
⇢ = 40% 0.40 0.30 0.38 0.31 5.48
1/n (uniform weights) 0.10 0.21 0.23 0.13 1.00

I train period of 2000 days used to compute optimal portfolio

I test period is different 500-day period

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.15

 Total portfolio value

Train Test
18
150
Risk-free
Value (thousand dollars)

1/n 16
10%
100 20%
14
40%

50 12

10 10

0 400 800 1200 1600 2000 0 100 200 300 400 500
Day Day

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.16

 Outline

Portfolio optimization

Linear quadratic control

Linear quadratic state estimation

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.17

 Linear dynamical system

xt+1 = At xt + Bt ut , yt = Ct xt , t = 1, 2, . . .

I n-vector xt is state at time t

I m-vector ut is input at time t
I p-vector yt is output at time t
I n ⇥ n matrix At is dynamics matrix
I n ⇥ m matrix Bt is input matrix
I p ⇥ n matrix Ct is output matrix
I xt , ut , yt often represent deviations from a standard operating condition

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.18

 Linear quadratic control

minimize Joutput + ⇢Jinput

subject to xt+1 = At xt + Bt ut , t = 1, . . . , T 1
x1 = xinit , xT = xdes

I variables are state sequence x1 , . . . , xT and input sequence u1 , . . . , uT 1

I two objectives are quadratic functions of state and input sequences:

Joutput = ky1 k 2 + · · · + kyT k 2 = kC1 x1 k 2 + · · · + kCT xT k 2

Jinput = ku1 k 2 + · · · + kuT 1 k 2

I first constraint imposes the linear dynamics equations

I second set of constraints specifies the initial and final state
I ⇢ is positive parameter used to trade off the two objectives

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.19

 Constrained least squares formulation

minimize kC1 x1 k 2 + · · · + kCT xT k 2 + ⇢ku1 k 2 + · · · + ⇢kuT 1 k 2

subject to xt+1 = At xt + Bt ut , t = 1, . . . , T 1
x1 = xinit , xT = xdes

I can be written as
minimize k Ãz b̃k 2
subject to C̃z = d̃
I vector z contains the Tn + (T 1)m variables:

z = (x1 , . . . , xT , u1 , . . . , uT 1 )

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.20

 Constrained least squares formulation

26 C1 ··· 0 0 ··· 0 37
66 .. .. .. .. 77
66 77
.. ..
. . . . . .
66 0 ··· CT 0 ··· 0 77
à = 66 7,
0 77
p b̃ = 0
66 0 ··· 0 ⇢I ···
66 .. .. .. .. 77
. 77
.. ..
66 . . . . .
⇢I 75
p
4 0 ··· 0 0 ···

26 A1 I 0 ··· 0 0 B10 ··· 0 37 26 0 37

66 77 66 77
66 0 A2 I · · · 0 0 0 B2 · · · 0 77 66 0 77
6 .. .. .. .. .. .... . . . 77 6 .. 77
C̃ = 66 . . . . . . . . ..
77 , d̃ = 66 .
7
66 0 0 0 · · · AT I 0 0 · · · BT 77 66 0 77
66 66 xinit 77
1 1

66
I 0 0 ··· 0 0 0 0 ··· 0 77 66 7
4 0 0 0 ··· 0 I 0 0 ··· 0 75 4 xdes 75

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.21

 Example

I time-invariant system: system matrices are constant

26 0.855 1.161 0.667 37 26 0.076 37

A = 666 0.015 1.073 0.053 77 ,
77 B = 666 0.139 777 ,
64 0.084 0.059 1.022 5 64 0.342 75
f g
C= 0.218 3.597 1.683

I initial condition xinit = (0.496, 0.745, 1.394)

I target or desired final state xdes = 0

I T = 100

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.22

 Optimal trade-off curve

4.4

4.2

⇢=1
Joutput

⇢ = 0.2
3.8
⇢ = 0.05

3.6

0 1 2 3 4
Jinput

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.23

 Three points on the trade-off curve

0.5 0.4
⇢ = 0.05 0 0.2
ut

yt
0.5 0
0 20 40 60 80 100 0 20 40 60 80 100

0.5 0.4
⇢ = 0.2 0 0.2
ut

yt

0.5 0
0 20 40 60 80 100 0 20 40 60 80 100

0.5 0.4
⇢=1 0 0.2
ut

yt

0.5 0
0 20 40 60 80 100 0 20 40 60 80 100
t t
Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.24
 Linear state feedback control

I linear state feedback control uses the input

ut = Kxt , t = 1, 2, . . .

I K is state feedback gain matrix

I widely used, especially when xt should converge to zero, T is not specified

I one choice for K : solve linear quadratic control problem with xdes = 0
I solution ut is a linear function of xinit , so u1 can be written as

u1 = Kxinit

I columns of K can be found by computing u1 for xinit = e1 , . . . , en

I use this K as state feedback gain matrix

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.25

 Example

0.1 0.4

State feedback
0 Optimal
ut

0.2
yt

Optimal

State feedback
0.1
0
0 50 100 150 0 50 100 150
t t

I system matrices of previous example

I blue curve uses optimal linear quadratic control for T = 100
I red curve uses simple linear state feedback ut = Kxt

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.26

 Outline

Portfolio optimization

Linear quadratic control

Linear quadratic state estimation

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.27

 State estimation

I linear dynamical system model:

xt+1 = At xt + Bt wt , yt = Ct xt + vt , t = 1, 2, . . .

I xt is state (n-vector)
I yt is measurement (p-vector)
I wt is input or process noise (m-vector)
I vt is measurement noise or measurement residual (p-vector)
I we know At , Bt , Ct , and measurements y1 , . . . , yT

I wt , vt are unknown, but assumed small

I state estimation: estimate/guess x1 , . . . , xT

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.28

 Least squares state estimation

minimize Jmeas + Jproc

subject to xt+1 = At xt + Bt wt , t = 1, . . . , T 1

I variables: states x1 , . . . , xT and input noise w1 , . . . , wT 1

I primary objective Jmeas is sum of squares of measurement residuals:

Jmeas = kC1 x1 y1 k 2 + · · · + kCT xT yT k 2

I secondary objective Jproc is sum of squares of process noise

Jproc = kw1 k 2 + · · · + kwT 1 k 2

I > 0 is a parameter, trades off measurement and process errors

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.29

 Constrained least squares formulation

minimize kC1 x1 y1 k 2 + · · · + kCT xT yT k 2 + (kw1 k 2 + · · · + kwT 1 k 2 )

subject to xt+1 = At xt + Bt wt , t = 1, . . . , T 1

I can be written as
minimize k Ãz b̃k 2
subject to C̃z = d̃
I vector z contains the Tn + (T 1)m variables:

z = (x1 , . . . , xT , w1 , . . . , wT 1 )

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.30

 Constrained least squares formulation

26 C1 0 ··· 0 0 ··· 0 37 26 y1 37
66 0 C2 0 0 0 77 66 77
66 77 66 77
··· ··· y2
66 .. .. ..
.
.. .. .. 77 66 .. 77
6 77 6 77
. . . . . .
à = 66 0 0 ··· CT 0 ··· 0 77 , b̃ = 66 yT 77
66 p
77 66 77
66 0 0 ··· 0 I ··· 0
77 66 0
77
66 .. .. .. .. .. ..
77 66 ..
77
66 66
. . . . .
p. .
4 0 0 ··· 0 0 ··· I 75 4 0 75

26 A1 I 0 ··· 0 0 B1 0 ··· 0 37
66 77
C̃ = 666 77 ,
0 A2 I ··· 0 0 0 B2 ··· 0
66
.. .. .. .. .. .. .. .. .. 77 d̃ = 0
77
. . . . . . . . .
64 0 0 0 ··· AT 1 I 0 0 ··· BT 1 5

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.31

 Missing measurements

I suppose we have measurements yt for t 2 T , a subset of {1, . . . , T }

I measurements for t < T are missing

I to estimate states, use same formulation but with

X
Jmeas = kCt xt yt k 2
t2 T

I from estimated states x̂t , can estimate missing measurements

ŷt = Ct x̂t , t<T

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.32

 Example

26 37 26 37
6
1 0 1 0
77 6
0 0
77 " #
At = 666 Bt = 666
0 1 0 1 0 0 1 0 0 0
77 , 77 , Ct =
66 0 0 1 0 77 66 1 0 77 0 1 0 0
4 0 0 0 1 5 4 0 1 5

I simple model of mass moving in a 2-D plane

I xt = (pt , zt ) : 2-vector pt is position, 2-vector zt is the velocity

I yt = Ct xt + wt is noisy measurement of position
I T = 100

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.33

 Measurements and true positions

0
t=1

500
(xt )2

1000

1500 t = 100

0 100 200 300

(xt )1
I solid line is exact position Ct xt
I 100 noisy measurements yt shown as circles

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.34

 Position estimates

=1 = 103 = 105

blue lines show position estimates for three values of

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.35

 Cross-validation

I randomly remove 20% (say) of the measurements and use as test set

I for many values of , estimate states using other (training) measurements

I for each , evaluate RMS measurement residuals on test set

I choose to (approximately) minimize the RMS test residuals

Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.36

 Example

80

60
RMS error

Test
40

20

Train
0
10 3 10 1 101 103 105

I cross-validation method applied to previous example

I remove 20 of the 100 measurements
I suggests using ⇡ 103
Introduction to Applied Linear Algebra Boyd & Vandenberghe 17.37