1

Dynamics and Control of Supply and Demand

Mordechai Shefer
© Copyright M. Shefer 2012 All Rights Reserved

Abstract
Dynamic model of a single-commodity, single-supplier, supply-and-demand
interaction is defined, which includes a psychological factor. The effect of such
factor in terms of stability is rigorously analyzed. The possibility of a maximum-
profit control is further explored. Finally, an optimal-quadratic control is devised
and applied to the process and its outcomes are presented and interpreted.

Introduction
We regard a simple dynamic model of an economic process (e.g., [3]), herein a
single-commodity, single-supplier, supply-and-demand interaction, as a 2 nd-order
linear equation where the demand d and the supply s both in terms of product-units
per process-stage are the state-variables and the market-price p of a product-unit
and the supply-command u s are the control variables, both are to be determined by

the supplier. In the continuous-time domain we herein propose:

d   ap  c(d  s ) (1)
Where:
p = Market-price of a product-unit
a = Market sensitivity to price, a real positive economical parameter
c = Market “hysteria-factor”, a real positive psychological parameter
As for the supply dynamics we assume that it just follows the supplier-command u s

with a 1st-order lag whose time-constant  is of a real positive value. This is:
s u
s    s (2)
 
Where:
s = Supply
 = Supply time-constant
u s = Supply-command
 2

Augmenting (1) and (2) in a matrix-vector notation we obtain the continuous-time

2nd-order linear process:
x  Fc x  Gc u (3)

Where:
d   p
x   ; u ;
s u s 
c  c   a 0 
Fc    ; Gc   ;
0  1 /    0 1/ 
Passing to the discrete-time domain with a unity (say, one day) time-step or stage,
we have:
xk 1  Fxk  Guk (4)

Where:
k = Stage (day) index
F  exp( Fc )  I  Fc ; G  Gc ;

Conduct of the Uncontrolled System

The characteristic behavior of the uncontrolled, commonly called: “Open-Loop”
system, can be assessed from the eigenvalues of the process matrix. We see from
either Fc or F that for any c  0 , the open-loop system shall have one unstable

eigenvalue. To demonstrate the open-loop behavior with a positive c, we run the

process state equations with the parameters:
c 1
 3
And the initial conditions:
d ( 0)  1
s ( 0)  1
And with zero supply-command and price. The discrete open-loop eigenvalues are:
Ev1 = 2.000, an unstable Ev.
Ev2 = 0.6667, a stable Ev.
Figures 1 and 2 show the open-loop characteristic histories of the supply and the
demand. We see that while the supply is decaying to zero, the demand grows
 3

exponentially without limit. We define this characteristic behavior as a state of

crisis.

Open-Loop Supply 7 Open-Loop Demand

x 10
1 14

12
0.8
10
0.6
8

0.4 6

4
0.2
2

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time [day] Time [day]

Figure 1; Open-Loop Supply Figure 2; Open-loop Demand

Maximum-Profit Control
Assuming a unity production-cost per product-unit, the supplier theoretical
maximum (TM) daily profit g k is given by:

g k  pk d k  sk (5)

Generally denoted as:

g k  xkT Amp uk  Bmp xk

Where:
1 0
Amp   ; Bmp  0  1 ;
0 0 
The supplier objective is to maximize its accumulated TM profit over time, that is,
to maximize the functional:
N
J   (xTk Ampu k  Bmp xk ) (6)
k 1

Where N is an arbitrarily large natural number. We denote the TM profit-to-go from

the k day to the N day by:
N
J k   (xTk Ampuk  Bmp xk ) (7)
k

Denoting by J k* the maximum TM profit-to-go from the k stage, we have, via

Bellman’s Principle of Optimality [1,2], the discrete functional equation (DFE):

 4

J k*  max{( xTk Amp uk  Bmp xk )  J k* 1} (8)

u

We assume a solution to the DFE (8) by:

1 T
J k*  xk  k xk   k xk (9)
2
Where:
 k = Matrix parameter

 k = Row vector parameter

We substitute the assumed solution (9) into the DFE (8) and obtain:
1 T 1
xk  k xk   k xk  max {xkT Amp uk  Bmp xk  xkT1 k 1xk 1   k 1xk 1}
2 u 2
(10)
Which is:
1 T 1
xk  k xk   k xk  max {xkT Ampu k  Bmp xk  ( Fx  Gu )Tk  k 1 ( Fx  Gu ) k
2 u 2
  k 1 ( Fxk  Guk )}
(11)
We maximize the right-hand side of (11) over uk by equating its derivative with

respect to uk to zero. This yields the maximum-profit control:

uk  (GT  k 1G ) 1 ( Amp xk  GT  k 1Fxk  GT  kT1 ) (12)

Shortly denoted as:

uk  C x, k xk  Co, k (13)

Where:

C x, k  (GT  k 1G ) 1 ( Amp  GT  k 1F ) (14)

Co, k   k 1G (GT  k 1G ) T (15)

The maximum-profit control (13) is a full-state feedback law whose validity

depends on the computability of the gains: C x, k and Co, k offline ahead of time. In

order to compute these gains, we substitute (13) into (11). This yields (terms
without a stage index belong to the k stage):
 5

1 T
xk  k xk   k xk   xkT Amp (C x x  Co )k  Bmp xk
2
1
 [( F  C xG ) x  GCo )]Tk  k 1[( F  C xG ) x  GCo )]k
2
  k 1[( F  C xG ) x  GCo )]k
(16)
Equating the coefficients of the quadratic terms in xk on both sides of (16) yields:

 k  2 Amp C x  ( F  GC x )T  k 1 ( F  GC x ) (17)

Which together with (14) is a discrete Riccati equation that can be formally solved
backwards offline, ahead of time from properly prescribed end condition, e.g.,:
N  0
C x, N  0

Equating the coefficients of xk on both sides of (16) yields:

T
 k  Co Amp  Bmp  CoG k 1 ( F  C x G )   k 1 ( F  C x G ) (18)

Which together with (15) is a linear regression in  that can be formally solved
backwards simultaneously with (16), offline, ahead of time from properly
prescribed end condition, e.g.,:
N  0
Co , N  0

In practice we solve for C x (and  ), Co (and  ), until they sufficiently converge

to their steady-state values and use these values as constant gains online in real-
time. The existence of solutions to regressions (17) and (18) would validate the
maximum-profit solution (13). Unfortunately however, equations (17) and (18) do
not appear to have solutions, therefore the maximum-profit control in its present
definition does not seem to exist.

Optimal Linear-Quadratic Control

The optimal linear-quadratic (LQ) control methodology seeks the minimum over
the command vector u of a quadratic functional:
N 1 1
J   ( xTk Axk  ukT Buk ) (19)
k 1 2 2
 6

Where A and B are prescribed symmetrical matrices with positive diagonals.

The DFE in the discrete cost-to-go J k* in the quadratic case has been defined in [2]

as:

J k*  min{( xTk Axk  uTk Buk )  J k*1} (20)

u

Whose solution is:

1 T
J k*  xk S k xk (21)
2

Where S is a symmetrical matrix with positive diagonal and the optimal control uk

is given per any stage as a full-state feedback law:

uk  ( B  GT S k 1G )1GT S k 1Fxk  Ck xk (22)

Where: S k , Ck are precomputable offline, ahead of time, via the discrete backwards

Riccati equation:

S k  A  C T BC  ( F  GC )T S k 1 ( F  GC ) (23)

With properly prescribed end conditions: S N , C N . It was shown in [1] that

quadratic optimal control always stabilizes controllable processes. Controlling the

process while satisfying a given set of two supplier requests is possible by

subtracting a vector xo  d o so T from the fed-back state vector in the closed

loop, where d o and so are positive control constant parameters to be prescribed by

the supplier. The closed-loop state-model is then given by:

xk 1  Fxk  GC ( xk  xo )  ( F  GC ) xk  GCxo (24)

In the steady-state we accordingly have:

d ss 
xk 1  xk  xss     ( I  F  GCss ) 1GCss xo  Kxo (25)
 sss 

Where: K  ( I  F  GCss ) 1GCss is the matrix sensitivities of xss to xo . We also

have:
u ss  C ss ( xss  xo )  Css ( K  I ) xo (26)

Which may be conveniently used in the control design (see below) in order to
foresee and avoid a too low steady-state price.
 7

Numerical Example with Optimal Linear-Quadratic Control

Example of the present optimal LQ control for our supply and demand process has
been done with the parameters:
a 1
c 1
 3
And the boundary conditions are:
d ( 0)  1
s ( 0)  0
0 0 
S(N )   
0 0 
The weighting matrices are:
 1  1 1 0
A  ; B ;
 1 1  0 1 
 - 1.5095 1.1697 
The steady-state gain matrix comes out as: C ss   .
- 0.4149 0.3366
The steady-state closed-loop eigenvalues in this example are those of the matrix:
F  GCss , which are: Ev1 = 0.3660, a stable Ev., Ev2 = 0.6790, a stable Ev.,

reassuring that the optimal LQ control strictly stabilizes the system. The matrix
sensitivities of the steady-state demand and supply to the control parameters d o and

 3.1891 - 2.4670
so in this example comes out as: K   . To design the optimal
0.6796 - 0.5141
control, the supplier first assigns his desired values for the steady-state supply and
demand. Let us assume that the supplier production-line is capable of producing
100 product-units per day. Then, the supplier might feel reasonably safe to request:
sss  90
d ss  100

With this choice and by: xo  K 1xss , the supplier now assigns: d o  4594.0 and

so  5898.2 to the control-law. To trigger the closed-loop system we select:

d (0)  1 which may be achieved by advertizing of the new product. The simulation
histories of the closed-loop system states are given in Figs. 3 to 6 below. Figs. 3 and
4 show good agreement to the supplier desired steady-state supply and demand
 8

values, where in order to secure a stable market demand, the supply continuously
lags the demand by the presently desired safety-margin of 10%. In Fig. 6 we further
see that the optimal price in the first 3 days of the process is negative. These
supplier expenses can be interpreted as the cost of penetrating the market.
Nevertheless, the present example steady-state price of about 10 expresses a quite
solid supplier profit of about 9:1.

Closed-Loop Supply Closed-Loop Demand

100 100

80 80

60 60

40 40

20 20

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time [day] Time [day]

Figure 3; Closed-Loop Supply Figure 4; Closed-loop Demand

Closed-Loop Us Closed-Loop Price

90 10

88
0

86
-10
84
-20
82

-30
80

78 -40
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time [day] Time [day]

Figure 5; Closed-Loop Supply-Command Figure 6; Closed-loop Price

Conclusions
Dynamic model of a single-commodity, single-supplier, supply-and-demand
interaction has been defined and analyzed. It has been shown that a positive
hysteria-factor (c) destabilizes the open-loop process and brings it to a state of
crisis. It has been further shown that there are no price and supply policies that can
maximize the supplier profit. Finally, it has been demonstrated that since optimal
 9

linear-quadratic control always strictly stabilizes the process, it may as well put it in
a steady supplier-profitable state.

References
[1] E.G. Al’brekht, “On the Optimal Stabilization of Nonlinear Systems,”
PMM, Vol. 25, No. 5, 1961, pp. 836-844, J. Appl. Math. Mech., 1962, pp.
1254-1266.
[2] R.E. Bellman, “Introduction to the Mathematical Theory of Control
Processes,” Volume II, Academic Press, 1971, pp. 16-21.
[3] D.G. Luenberger, “Introduction to Dynamic Systems,” John Wiley & Sons,
1979, pp. 6-8.