Math Meth Oper Res (2011) 73:2954

DOI 10.1007/s00186-010-0332-3
ORIGINAL ARTICLE
A continuous framework for open pit mine planning
Felipe Alvarez Jorge Amaya
Andreas Griewank Nikolai Strogies
Received: 2 February 2010 / Accepted: 29 September 2010 / Published online: 17 October 2010
Springer-Verlag 2010
Abstract This paper proposes a newmathematical framework for the open pit mine
planning problem, based on continuous functional analysis. The main challenge for
engineers is to determine a sequence of nested proles maximizing the net present
value of the mining operation. The traditional models for this problem have been con-
structed by using binary decision variables, giving rise to large-scale combinatorial
and Mixed Integer Programming problems. Instead, we use a continuous approach
which allows for a rened imposition of slope constraints associated with geotechni-
cal stability. The framework introduced here is posed in a suitable functional space,
essentially the real-valued functions that are Lipschitz continuous on a given two
dimensional bounded region. We derive existence results and investigate qualitative
properties of the solutions.
Keywords Mine planning Continuous optimization Calculus of variations
Functional analysis
F. Alvarez J. Amaya
Departamento de Ingeniera Matemtica and Centro de Modelamiento Matemtico (CNRS UMI 2807),
Universidad de Chile, Santiago, Chili
e-mail: [email protected]
J. Amaya
e-mail: [email protected]
A. Griewank
DFG Research Center Matheon Mathematics for Key Technologies, Berlin, Germany
e-mail: [email protected]
N. Strogies (B)
Humboldt Universitt zu Berlin, Berlin, Germany
e-mail: [email protected]
1 3
30 F. Alvarez et al.
Mathematics Subject Classication (2000) 49J300 80M50 49J50 90C26
1 Introduction and motivation
Generally speaking, three different problems are usually considered by mining engi-
neers for the economic valuation, design and planning of open pit mines as pointed out
by Hustrulid and Kuchta (2006). The rst one is the Final Open Pit (FOP) problem,
also called Ultimate Pit Limit problem, UPL for short, which aims to nd the region
of maximal economic value for exploitation under certain geotechnical stability con-
straints and assuming innite extraction capacity. Another more realistic problem is
what we call here the Capacitated Final Open Pit (CFOP) which considers an addi-
tional constraint on the total capacity for a one-period exploitation. The third problem
is a multi-period version of the latter, which we call the Capacitated Dynamic Open Pit
(CDOP) problem, with the goal of nding an optimal sequence of extracted volumes
in a certain nite time horizon for bounded capacities at each period, the optimality
criterion being the total discounted prot.
A common practice for the formulation of these problems consists in describing an
ore reserve (copper, for example) via the construction of a three-dimensional block
model of the mineralization. Each block corresponds to a unitary volume of extraction
characterized by several geologic and economic properties which are estimated from
sample data. Block models can be represented as directed graphs where nodes are
associated with blocks, while arcs correspond to the precedence of these blocks in the
ore reserve. The precedence order is induced by physical and operational constraints
as those derived fromthe geomechanics of slope stability. This discrete approach gives
rise naturally to huge combinatorial problems whose mathematical formulations are
special large-scale instances of Integer Programming (IP) optimization problems (see,
for instance, Cacetta 2007). This explains why the optimal planning of open pit mines
based on block models is usually addressed by approximation methods, heuristics
and mixed IP techniques as Linear Programming relaxations of integer variables and
branch-and-bound algorithms.
A great number of publications dealing with discrete block modeling for open pit
mines have been published since the sixties. Aseminal paper by Lerchs and Grossman
(1965) proposes a practical procedure to obtain the ultimate pit limit, which have been
extensively applied in real mines for many years. The capacitated dynamic problem
is more difcult to solve and many methods using discrete optimization techniques
have been proposed by Boland et al. (2006), Cacetta and Hill (2003) and Hochbaum
and Chen (2000). Some dynamic programming formulations as in Johnson and Sharp
(1971) and in Wilke and Wright (1984) give interesting results, but the applicability of
these techniques is still not well established for large-scale problems. Metaheuristic
and evolutionary algorithms have also been extensively tested by Denby and Schoeld
and Ferland et al. (2007).
In this paper we propose an alternative approach to the above mentioned problems
based on a continuous framework for the ore reserve as well as the mining activity. The
basic idea is to describe the pit contours through a Lipschitz continuous real-valued
function, a prole which maps each pair of horizontal coordinates to the corresponding
1 3
A continuous framework for open pit mine planning 31
vertical depth. The stability of steep slopes is ensured by a spatially distributed con-
straint on the local Lipschitz constant of the prole function. The maximal feasible
local slope may vary depending on the geotechnical properties of the possibly hetero-
geneous mineral deposit. The extraction capacity and operational costs are described
by a possibly discontinuous effort density, a scalar function dened on the three-dimen-
sional region occupied by the ore reserve. In order to quantify the economic value of
an extracted volume described by a given prole function, we consider a gain density
dened on the deposit, which again may be a discontinuous function.
Our goal here is to develop a complete existence theory and investigate qualitative
properties of the optimal solutions to the proposed continuous versions of the FOP,
CFOP and CDOP problems. The numerical resolution of these problems based on
strategies from continuous optimization in functional spaces will be the subject of
future research.
It is worth mentioning that the rst documented approach applying continuous
functions for a parametrized variant of the CFOP problem seems to be the work by
Matheron (1975). He explicitly exploited the underlying lattice structure of the set
of feasible proles for his framework in order to obtain existence results as well as
interesting characterizations for optimal solutions but neither a proof of existence nor
optimality conditions were given there. More recently, a simple related continuous
scheme was introduced by Morales (2002), for underground mines, but no study on
existence nor optimality conditions are given there. On the other hand, Guzmn (2008)
has proposed a continuous framework for the FOP problem based on shape and topo-
logical optimization using level-set techniques, reporting computational results for
a very simplied instance of the problem, but again no rigorous existence theory is
provided.
This paper is organized as follows. In Sect. 2 we describe the stationary problem
in terms of continuous prole functions by introducing nonnegativity, boundary and
stability conditions, and we prove that such a set of admissible proles is compact for
the uniform-convergence topology. Furthermore we establish structural properties of
this set related to the lattice structure induced by pointwise min and max operations.
In addition, we introduce an effort and a gain function which are related to the capac-
ity constraints and the prot objective function, respectively. In Sect. 3 we state the
optimization problems for the stationary case, derive nonconstructive existence results
for them and describe qualitative properties of the corresponding optimal solutions by
exploiting the lattice structure of the set of feasible proles. In Sect. 4 we introduce a
dynamic planning problem with discount rates, investigate properties of the dynamic
feasible set and give an existence result. Finally, in Sect. 5 we briey summarize the
main contributions of this paper and indicate some lines for future research.
2 The stationary problem
2.1 Continuous prole functions
Throughout this paper represents either a bounded connected domain R
2
with
Lipschitz boundary , or a bounded open interval =(a, b) R with ={a, b}.
1 3
32 F. Alvarez et al.
Fig. 1 Example of a continuous prole function on a one dimensional domain
In any case, the state of excavation at any particular time is dened by a function
p : R called prole so that z = p(x) for x , where the vertical coordinate
z indicates the depth of the pit at point x (see Fig. 1). In this paper p, as not stated
otherwise, is always an element of the Banach space of continuous real valued func-
tions C() endowed with the supremum norm given by p

sup
x
| p(x)|. The
initial state (prole) is dened by a function p
0
C() so that all admissible proles
p must satisfy the nonnegativity condition
p(x) p
0
(x) 0 for x . (1)
Moreover, we assume that no excavation happens on the boundary of and thus
impose the Dirichlet condition
p(x) p
0
(x) = 0 for x . (2)
In other words p p
0
must belong to the nonnegative orthant of the linear space
C
0
() C() of continuous functions on that satisfy homogeneous boundary
conditions.
The admissible proles are not only bounded frombelowby z = min{ p
0
(x)|x }
but also from above by a value z > z due to physical and operational conditions. Thus
for any admissible prole we assume that
p(x) Z [z, z] for x . (3)
Of course, with no loss of generality we may assume z 0. The general situation is
sketched in Fig. 1 for the one dimensional case.
1 3
A continuous framework for open pit mine planning 33
2.2 Geotechnical stability condition and compactness
In order to measure the local slope associated with a given prole p C(), we
dene
L
p
(x) limsup
xx x
| p( x) p( x)|
x x
for x . (4)
Here and throughout denotes the Euclidean norm. One can easily show that for
each p C() the corresponding function L
p
: [0, ] is upper semi-contin-
uous. Where L
p
(x) is nite it induces sharp local Lipschitz constant for p around x
in the sense that for all > 0, there exists for all > 0 a > 0 such that for any x, x
with x x < > x x we have
| p( x) p( x)| (L
p
(x) +) x x,
and if L is any local Lipschitz constant for p in a neighborhood of x then L
p
(x) L.
The key assumption on the admissible proles p is the pointwise stability condition
L
p
(x) (x, p(x)) for x , (5)
where : Z [0, ) is an upper bound on the limiting local Lipschitz constant
of p. It prescribes the maximal stable local slope and may vary on Z depending
on the local geotechnical properties of the material. Rather than assuming continuity
of the slope function we allow for horizontal and vertical jumps, which might be
caused by layers of different material. In particular we may have soft and hard mate-
rial layers next to each other, so that may jump upwards as one crosses into the
harder material. Wherever is discontinuous we may dene it as its upper envelope
over a neighborhood and thus we can assume without loss of generality the upper
semi-continuity property:
limsup
j
(x
j
, z
j
) (x, z) (6)
for all convergent sequences (x
j
, z
j
) (x, z) Z. This assumption immediately
implies that attains a maximum
max
(x,z)Z
(x, z). (7)
To restrict the feasible proles to the volume Z we assume without significant
loss of generality for notational simplicity that is less than 0.5|Z|/di am(). Apart
from the supremum norm in C() we will utilize the extended-real valued Lipschitz
semi-norm
p
Li p
= sup
x
L
p
(x) [0, ] for p C().
1 3
34 F. Alvarez et al.
For all p C() satisfying (5) we have p
Li p
as dened in (7). The linear space
of all p C() for which p
Li p
is indeed nite is denoted by Li p(), which can
be endowed with the norm p
1,
p

+p
Li p
to obtain the standard Banach
space of all Lipschitz functions on . On the subspace Li p
0
() Li p() C
0
()
the quantity p
Li p
denes a proper norm which is equivalent to p
1,
. In fact,
it is not difcult to see that for any p Li p
0
() we have that x , | p(x)|
p
Li p
di st

(x, ), where di st

(x, ) stands for the distance, relative to , from

x to the boundary of . As is bounded, we conclude that for a constant C <
depending only on we have that p

Cp
Li p
. In particular, the embed-
ding Li p
0
() C
0
() is completely continuous. It is well-known that the resulting
Banach space Li p
0
() is equivalent to the Sobolev space W
1,
0
() (see Evans 1998).
Furthermore, by virtue of Rademachers theorem, every p Li p() is a.e. differ-
entiable in . It follows that for every p satisfying (5) we have that
p(x) (x, p(x)) for a.e. x .
Throughout, we assume that p
0
satises (5). The set
P := { p C() | p satises (1), (2) and (5)}
is called set of admissible proles. In fact the bounded set Pis contained in the afne
subspace p
0
+ Li p
0
(). Notice that the u.s.c. assumption (6) on is necessary for
the closedness of P in C(). Indeed let us consider the following simple example:
Take = (1, 1) and [0, 1]. Set

(x, z) = 0 if x < 0, if x = 0 and 1 if

x > 0. This function is discontinuous at x = 0 and is not u.s.c. if < 1. For each
> 0, the prole p

(x) = max(0, x ) satises (5) with =

for any [0, 1];

nevertheless, its uniform limit p(x) = max(x, 0) is admissible if and only if = 1.
Furthermore, under (6) we have that P is compact in C() as we will show below.
Proposition 1 If satises (6) then P is compact and has empty interior in C().
Proof As our feasible functions are xed on we know that P p
0
C
0
().
The uniform Lipschitz continuity property ensures the equicontinuity of P, while all
functions in P have values in the compact interval Z = [z, z]. Hence the asserted
compactness of P in C() follows by the well known theorem of Arzela-Ascoli
provided we can show that it is closed w.r.t.

.
Let p p
0
+ C
0
() be a function in the closure of P w.r.t.

. By virtue of
(6), we have that for given x and > 0 there exists > 0 such that for any x
and q P with x x < > q p

then
( x, q( x)) < (x, p(x)) +/4. (8)
From this it follows that for all x = x in the ball B

(x) we have
|q( x) q( x)|
x x
< (x, p(x)) +

2
. (9)
1 3
A continuous framework for open pit mine planning 35
which will be shown by contradiction. If equality (9) is violated by a couple of points
{ x, x = x +x} B

(x) with x = 0, then it would be also violated by at least one

of the pairs { x, x +x/2} or { x x/2, x} because of
|q( x) q( x)|
x x

1/2(|q( x) q( x +x/2)| +|q( x) q( x x/2)|)
1/2 x x

1/2(2 max{|q( x) q( x +x/2)|, |q( x) q( x x/2)|})

1/2 x x
=
max{|q( x) q( x +x/2)|, |q( x) q( x x/2)|}
1/2 x x
and 1/2 x x = x x + x/2 = x x + x/2. Recursively, we can con-
tinue the bisection process to generate a family of nested segments {[ x
j
, x
j
]}
j 0
of
length x/2
j
, all of them contained in B

(x), such that the corresponding pairs

x
j
, x
j
violates (9) for each j , and moreover one would have that there exists a limit
x

[ x
j
, x
j
] B

(x) such that x

j
x

and x
j
x

as j . By construction,
using (8) and since q P, we get:
(x

, q(x

)) +

4
< (x, p(x)) +

2

|q( x
j
) q( x
j
)|
x
j
x
j

(x

, q(x

)) + R
j
(x)
where the remainder R
j
(x) is of the formo(x/2
j
)/(x/2
j
). Letting j ,
we obtain a contradiction in (8) as R
j
(x) vanishes. Thus (9) is indeed implied by
(8). Let us return to the function p. Given (0, 1], let us consider the corresponding
> 0 as in (8). For any x = x with x x < and x x < there exists a
prole q P such that p q

x x/4 and therefore

| p( x) p( x)|
x x

|q( x) q( x)| +2p q

x x
(x, p(x)) +2

2
(x, p(x)) +.
Since may be chosen arbitrary small p must satisfy the stability condition (5) and is
therefore contained in P, which is thus closed as asserted. This completes the proof
of the compactness of P.
To see that P has empty interior w.r.t. the

norm we only have to consider a

triangle wave function p
t
around an admissible prole p P. For any > 0, scaling
the amplitude of p
t
we can ensure p
t
p

. But letting the wavelength of p

t
go down, the limiting local slope L
p
t
increases and thus the stability condition (5) will
be violated. Hence in every neighborhood of a feasible prole w.r.t.

there are
infeasible proles.
As an immediate consequence of Proposition 1 we get that any functional F :
C() R which is continuous w.r.t.

attains on P a minimum and a max-

imum. This applies in particular to the distances F( p) p p

for any xed

p L

() C() P. Hence we have (non unique) least distance projections

from L

to P.
1 3
36 F. Alvarez et al.
2.3 Additional conditions on the slope constraint
One might want to impose two additional conditions on (x, z) concerning its depen-
dence on the vertical coordinate z. The rst one is that the restrictions
x
(z) (x, z)
be right continuous, i.e.
lim
zz

x
( z) =
x
(z) for x (10)
The physical motivation for this property is the exclusion of certain pathological situ-
ations where one may have two regions with soft material lying below the hard one. If
is only u.s.c. with respect to all variables, a prole might be feasible but physically
unstable because the maximal slope permitted in the hard material cannot be supported
by the soft material below which only allows a milder slope. The assumption (10) of
right-continuity w.r.t. z means that the slope constraint can not simply jump up from
a soft layer below a hard one, but must build up gradually.
The second additional condition on is definitely optional, namely we may require
concavity of (x, z) w.r.t. z Z. This rather strong condition, while clearly not very
realistic in the general case, does allow for the possibility of hard material in the mid-
dle sandwiched in between soft material on top and below. Of course, one may also
consider the case of a concave which is monotonically increasing or decreasing
w.r.t. z according to the geomechanics of the material.
Lemma 1 If (x, z) is concave w.r.t. z then P is convex.
Proof Take two proles p, q P and 0 < < 1. For any x , we have by
definition of L
p
in (4) and the triangle inequality
L
(1) p+q
(x) (1 )L
p
(x) +L
q
(x)
(1 )(x, p(x)) +(x, q(x)) (x, (1 ) p(x) +q(x))
where the last inequality is a consequence of the concavity property on . Thus
(1 ) p +q P.
We end this section with a sufcient condition for an admissible prole to be in the
interior of P in Li p().
Proposition 2 If is continuous on ( Z) then any prole p P for which
inf
x
{(x, p(x)) L
p
(x)} > 0 (11)
lies in the interior of P in Li p().
Proof We have to showthat any p Psatisfying (11) has a neighborhood which only
contains feasible proles. Due to the continuity of on there exists for all > 0
a > 0 such that at all x , we have |(x, z) (x, z)| < /2 if |z z| < .
Pick any q with q p
Li p
< /8 and q p

< . The set of such q is an open

1 3
A continuous framework for open pit mine planning 37
neighborhood of p in P w.r.t. the Lipschitz norm
Li p
. Any x is contained
in some ball B

such that on that ball p q has a Lipschitz constant of size /4 and

p has a Lipschitz constant of size L
p
(x) + /4. Now we see that for any two points
x, x B

(x)
|q( x) q( x)|/ x x = |q( x) p( x) (q( x) p( x)) + p( x) p( x)|/ x x
|q( x) p( x) (q( x) p( x))|/ x x + L
p
(x) +/4
L
p
(x) +/2 (x, p(x)) /2 (x, q(x))
where the last estimation follows from the condition on the difference of p and q
w.r.t. the supremum norm. Hence q satises the slope constraint in that certainly
L
q
(x) (x, q(x)) and thus q P.
2.4 Structural properties of the admissible proles set
The next result establishes closedness properties of P under pointwise minima and
maxima operations. These properties ensure the connectedness of P as a subset of
C() even when (x, z) is not assumed to be concave w.r.t. z so that P can be non-
convex.
Moreover we will consider the so called intermediate level proles
q

= max{ p, min{q, }} for R (12)

concerning two arbitrary, essentially bounded, ordered functions p q L

().
Proposition 3 Under (6) we have that:
(i) Pis closed with respect to pointwise minima and maxima in that for any subset

P P the functions p(x) and p(x) dened by

p(x) inf{ p(x)| p

P} and p(x) sup{ p(x)| p

P} (13)
also belong to P. Thus Pcontains a unique maximal element p
u
max
pP
{ p}.
(ii) If p, q P are such that p q then q

P for any level . Moreover

the path q

is continuous w.r.t.

.
(iii) Any two proles p, q P are connected via min{ p, q} and max{ p, q}.
Proof (i) We will apply induction and a classical diagonalization argument to
generalize the claim from the binary case to subsets containing innitely many
proles. First consider

mat hscr P = { p, q} P. At points x where
r(x) max{ p(x), q(x)} = p(x) > q(x) the same is true for all x x by
continuity and we have thus L
r
(x) = L
p
(x) (x, p(x)) = (x, r(x)). At
points x where there is a tie r(x) = p(x) = q(x) we nd that for any two
1 3
38 F. Alvarez et al.
sequences y
j
x and z
j
x
|r(y
j
) r(z
j
)| = | max{ p(y
j
), q(y
j
)} max{ p(z
j
), q(z
j
)}|
max{| p(y
j
) p(z
j
)|, |q(y
j
) q(z
j
)|}
The last inequality follows from the inverse triangle inequality for the supre-
mum norm on R
2
. After division by y
j
z
j
and taking the limit y
j
x
z
j
we nd that
L
r
(x) max{L
p
(x), L
q
(x)} max{(x, p(x)), (x, q(x))}
= (x, r(x)). (14)
Thus r = max{ p, q} P.
The same argument applies to r = min{ p, q} = max{p, q} for the slope
stability condition. Obviously it follows by induction that maxima and minima
of nitely many elements in P also belong to P.
Now suppose

P contains innitely many elements. First of all we note that
from p(x) sup{ p(x)| p

P} it follows that p L

(). Now we pick

a dense subset {x
j
}
j N
in . By induction on i we now choose sequences
p
(i )
k


P such that
lim
k
p
(i )
k
(x
j
) = p(x
j
) sup
p

P
{ p(x
j
)} for j < i
Consider a subsequence ( p
k
)

P with p
k
(x
i
) p(x
i
) and set p
(i +1)
k

max{ p
(i )
k
, p
k
} so that p
(i +1)
k
(x
j
) p(x
j
) for j i . Now we take the diago-
nal sequence p

k
= p
(k)
k
and get
lim
k
p

k
(x
j
) = p(x
j
) for j i (15)
We know for j k
p(x
j
) p

k
(x
j
) p
(i +1)
k
(x
j
)
. , .
p(x
j
)
and get a subsequence so that for all j N, p(x
j
) = lim
k
p

k
(x
j
) by the con-
vergence in (15). Moreover we can pick a Cauchy subsequence so that without
loss of generality p p

0 for some p P. Clearly we must have

p = p which proves (i).
(ii) The assertion is again obvious where all three values p(x), q(x) and are dis-
tinct. When there is a tie between two we may invoke the same argument as in
1 3
A continuous framework for open pit mine planning 39
Fig. 2 Illustration of level excavations for different levels between two given proles
(i) and then extend it to a three way tie. Since at all x
|q

q

| = | max{ p, min{q, }} max{ p, min{q, }}|
| min{q, } min{q, }| | |
we have in fact Lipschitz continuity of q

w.r.t. R. Consequently, P is
path connected in C() since any two p, q Pcan be transformed into each
other via min{ p, q} or max{ p, q} with the help of a path generated by a varying
level in (12).
(iii) Follows from (i) and (ii).
The path in (ii) is not continuous w.r.t.
Li p
. Take for instance p 0 and
q(x) =
1
4
(x
1
2
)
2
on = (0, 1) and sufciently large, so that we have p, q P
and q

(x) = max{0, min{q(x), }} belongs to P, but q

q
0

Li p
= q

Li p
= 1
for all > 0.
The prole modications applying intermediate level proles used in Proposition
3(ii) will be referred to as level excavations. They are depicted in Fig. 2 and make
practical sense as material is taken away in horizontal layers. While that does not
mean optimality when gains are discounted as discussed in Sect. 4 we note that any
monotonic chain of feasible proles p
0
< p
1
< < p
m
P can be extended
to a feasible path from p
0
to p
m
by level excavation between successive proles
p
j
p
j +1
. The next result on level excavations shows that we can regain feasibility
from any bounded q p P, so there is no danger of getting trapped away from
the admissible set during an iterative optimization process.
Proposition 4 Under (6) we have that if P p q L

() then q

(),
and the set V

{x | L
q

(x) > (x, q

(x))} is monotonically growing w.r.t. .

Proof The proof will be obtained from a case study for the bounds
q limsup
xx
q( x) q liminf
xx
q( x) p(x)
at any particular point x .
1 3
40 F. Alvarez et al.
If q > q we must have L
q
(x) = and also L
q

(x) = as long as > q

which in turn means x V

for > q. On the other hand it follows for < q that

min{, q( x)} = for all x near x so that clearly q

(x) = max{ p(x), } and thus

x / V

. Thus we have monotonicity whether or not x V

for = q.
Now suppose q = q which means that q(x) is continuous at x. If > q(x) then
for x x we have that q

( x) = q( x) and if < q(x) then for x x we have that

q

( x) = max{ p( x), }. Hence we have again x / V

if < q(x) and for > q(x) we

have x V

L
q
(x) > (x, q(x)). Now the only case left to consider is L
q
(x)
(x, q(x)) where we have to exclude that x V

for = q(x). However it follows

exactly as in (14) that for = q(x) we obtain L
r
(x) max{L
p
(x), L
q
(x)} = L
q
(x)
so that we have monotonicity in all cases.
2.5 Effort constraints and gain objective function
In addition to the modeling of the open pit problem relies on two other given real
valued functions, namely
e(x, z) e
0
> 0 and g(x, z) R for (x, z) Z. (16)
which are only assumed to be uniformly bounded, i.e.
e, g L

( Z) (17)
Hence it is allowed that e and g have jumps due the existence of different types of
material in the ore body. For any two given proles q p the integral
E([ p, q])
_

q(x)
_
p(x)
e(x, z)dzdx
represents the effort to excavate all the material between prole p and q, which is
expected to be bounded by the capacity of the mine operation. On the other hand, the
function
G([ p, q])
_

q(x)
_
p(x)
g(x, z)dzdx
represents the total value or gain of the material between p and q (without consider-
ing a discount rate). Notice that the function g(x, z) may take negative values. When
p = p
0
we abbreviate G(q) G([ p
0
, q]) and E(q) E([ p
0
, q]). For an ordered
triplet p q r with p, q, r P we have additivity in the sense that
G([ p, r]) =G([ p, q]) + G([q, r]) and E([ p, r]) = E([ p, q]) + E([q, r]) (18)
1 3
A continuous framework for open pit mine planning 41
Now we can give the basic properties of G and E as follows.
Proposition 5 As a consequence of (17) we have that
(i) E( p) and G( p) are Lipschitz continuous on C() with constants e

|| and
g

|| respectively, where || denotes the area of .

(ii) E( p) and G( p) are Gteaux differentiable at all p C() \ A where A is a
meager set in the sense of Preiss and Zaj cek (2001).
(iii) If e (resp. g) is continuous on Z then E( p) (resp. G( p)) is everywhere
Frchet differentiable. In particular, for any p C() we have that
E( p) p =
_

e(x, p(x))p(x)dx. (19)

(iv) E is convex (resp. G is concave) if e is monotonically increasing (resp. g is
monotonically decreasing) w.r.t. z.
Proof We consider throughout only E without making use of the positivity assump-
tion on e. Thus the results apply analogously to g(x, z).
(i) Considering two proles p, p P we get
|E( p) E( p)|
_

| p(x) p(x)| sup

zZ
|e(x, z)|dzdx || e

p p

(ii) By Preiss and Zaj cek (2001, Theorem 12), a generalization of Rademachers
theorem, every Lipschitz mapping on an open subset G of a separable Banach
space X to a Banach space Y with the RadonNikodym property is Gteaux
differentiable at all points of G except those belonging to a meager set A. In our
case X = C() is separable and Y = R has the RadonNikodym property, i.e.
every Lipschitz map R Y is differentiable almost everywhere.
(iii) We claim that under continuity of e and g there is an explicit representation for
the Frchet derivative, which in the case of E is given by (19). To establish this
we rewrite the difference between E and its proposed linearization as follows:
1
p

_
_
E( p +p) E( p)
_

e(x, p(x))p(x)dx
_
_

=
1
p

_
_
_
_
p(x)+p(x)
_
p(x)
e(x, z)dz
_

_
e(x, p(x))p(x)
_

_
dx

=
1
p

e(x, p(x) +p(x))p(x) e(x, p(x))p(x)dx

1 3
42 F. Alvarez et al.
where = (x, p(x), p(x)) [0, 1] is obtained by the rst mean value the-
orem for integration. Since e(x, z) is uniformly continuous on the compact set
Z there exist for any > 0 a bound > 0 such that p

< implies
|e(x, p + p) e(x, p)| and thus the last expression on the right hand
side is bounded above by ||, which completes the proof.
(iv) The monotonicity assumption implies that the antiderivative e(x, z)
_
z
p
0
(x)
e(x,
z)d z is convex w.r.t. z. Hence we nd for two proles p, p P
E((1 ) p + p) =
_

(1) p(x)+ p(x)

_
p
0
(x)
e(x, z)dzdx
=
_

e(x, (1 ) p(x) + p(x))dx

(1 ) e(x, p(x)) + e(x, p(x))dx

=
_

_
_
_
(1 )
p(x)
_
p
0
(x)
e(x, z)dz +
p(x)
_
p
0
(x)
e(x, z)dz
_
_
_
dx
= (1 )E( p) +E( p)
For g decreasing we obtain G((1 ) p + p) (1 )G( p) + G( p)
analogously.
Note that any of the results of the last proposition apply analogously to E([ p, q])
and G([ p, q]) when p = p
0
. The assumption that the effort rate e is monotonically
increasing w.r.t. the depth z is natural and realistic; but that the gain rate g be mono-
tonically decreasing w.r.t. z would only apply to very particular deposits. Therefore,
in general we cannot expect the gain function G to be concave. Without continuity
w.r.t z no global Frchet differentiability is attainable even in terms of
Li p
.
3 Optimal stationary proles
Using the properties of P, G and E we have derived in the previous section we can
establish existence results for proles that are optimal in various senses. The con-
tinuous formulation we propose for the Final Open Pit problem mentioned in the
introduction is the following:
(FOP) max{G( p) | p P}
Similarly, the continuous Capacitated FOP problem is:
(CFOP) max{G( p) | p P, E( p) E}
1 3
A continuous framework for open pit mine planning 43
The sets of optimal solutions (global maximizers) for these problems are denoted by
S(FOP) and S(CFOP), respectively. The following result establishes a property of
the gain and effort functions which in particular is useful to investigate the structure
of S(FOP).
Lemma 2 For all admissible proles p, q P, whether ordered, optimal or not,
we have that G(max{ p, q}) = G( p) + G(q) G(min{ p, q}) and E(max{ p, q}) =
E( p) + E(q) E(min{ p, q}).
Proof We have that
G([ p
0
, p]) + G([ p
0
, q]) = G([ p
0
, min{ p, q}]) + G([ p
0
, min{ p, q}])
+G([min{ p, q}, p]) + G([min{ p, q}, q])
. , .
=G([min{ p,q},max{ p,q}])
Here everything is done by the decomposition formula for ordered triplet (18) and the
fact, that from the minimum of two proles both excavation and gain are obtained on
disjoint areas when we go on excavating to one or the other except the sets where they
are the same and thus equivalent to the minimum.
Proposition 6 Under the conditions (6) and (17) we have
(i) S(FOP) is nonempty and contains unique minimal and maximal elements p
g

p
g
so that p S(FOP) p
g
p p
g
.
(ii) For any bound E > 0 there exists at least one global optimizer of (CFOP).
Proof (i) The maximum G

on P of G is attained due to the continuity of G and

the compactness of P by Proposition 1. On the other hand, Lemma 2 implies
that the maximum p
g
and minimum p
g
over all globally optimal proles are also
optimal. The nal assertion follows directly from Proposition 3(i).
(ii) The existence follows again from the compactness of Pand the continuity of E
and G.
Proposition 3(i) yields that P is a complete lattice as there is a maximal element p
and a minimal element p
0
, which indeed proofs that it is a bounded lattice. By the
existence of (13) for each subset as shown in Proposition 3(i) we have that all subsets
have a joint and a meet. In particular the solution set of FOP is a sublattice of P. On
the other hand, S(FOP) need not to be connected nor convex unless g is decreasing
w.r.t. z.
We can use the previous result to obtain a path of optimal proles subject to exca-
vation constraints. For each 0 the combined function
G

( p) G( p) E( p)/
satises all the assumptions we made on G( p) so far. It is in fact concave if this is
true for G( p) and E( p) is convex, which follows from e(x, z) and g(x, z) being
monotonically growing. Hence G

( p) has global minimizers on P just like G and

1 3
44 F. Alvarez et al.
we obtain a whole path. The next result, for which we provide a proof for the sake
of completeness, follows from the general theory of parametric lattice programming;
see Topkis (1978).
Proposition 7 Under (6) and (17) there exists a path of proles
p

arg min{E( p) | p arg max{G

( p) | p P}} (20)
so that 0 < p

and p

arg max{G( p) | p P E( p) E( p

)}.
Proof The existence of the p

follows fromProposition 5(i), namely G and E are Lips-

chitz. Let p

be dened as the unique minimal element among the global optimizers

of G

existing by the fact that P is a complete lattice. To prove the monotonicity

consider q = min{ p

, p

} for . By (18) we get for p

0
q p

( p

) = G( p

) E( p

)/
= G(q) E(q)/ + G([q, p

]) E([q, p

])/
= G

(q) + G

([q, p

]). (21)
Hence, by optimality of p

, we have
0 G

([q, p

]) = G([q, p

]) E([q, p

])/
G([q, p

]) E([q, p

])/ = G

([q, p

])
and using again the disjointness of [q, p

] and [q, p

] we nd that
G(max{ p

, p

}) E(max{ p

, p

})/ = G

(q) + G

([q, p

]) + G

([q, p

])
= G

( p

) + G

([q, p

])
G

( p

).
By optimality of p

for G

we obtain G

([q, p

]) = 0 G

([q, p

]) = 0. Hence
Eq. (21) becomes G

( p

) = G

(q) and by minimality of p

we derive p

= q. Hence
p

. The last assertion can be checked easily by contradiction.

It should be noted that the path p

established in Proposition 7 is in general not

continuous. That can only be expected in nice cases where G is strictly concave and
E is strictly convex.
As Pis bounded the same is true for its image I (E( p), G( p))
pP
R
2
in the
conguration space. E( p) ranges between E( p
0
) =0 and the maximal capacity E( p
u
)
that can be sensibly utilized. Here p
u
is the ultimate prole dened in Proposition 3(i).
The corresponding gain G( p
u
) will typically be positive but might be zero like G( p
0
)
in exceptional cases. Every point (E( p), G( p)) can be reached from the origin by
the level excavation path according to Proposition 3(ii) and similar (E( p

), G( p

)),
where p

denotes an element of the solution set of (FOP), can be reached from it

by another level excavation. The slopes of all the paths in the conguration space are
bounded by the following result:
1 3
A continuous framework for open pit mine planning 45
Proposition 8 For any pair p q P with p = q
|G( p) G(q)|
E(q) E( p)
sup
_
|g(x, z)|
e(x, z)
|(x, z) Z
_

e
0
where e
0
is the lower bound on the effort density function as introduced in (16).
Proof We have that
|G(q) G( p)| =

q(x)
_
p(x)
g(x, z)dzdx

q(x)
_
p(x)

g(x, z)
e(x, z)

e(x, z)dzdx
sup
_
|g(x, z)|
e(x, z)
|(x, z) Z
__

q(x)
_
p(x)
e(x, z)dzdx
= (E(q) E( p)),
which proves the estimate.
Geometrically represents a Lipschitz constant on G( p()) along any monotone
path parametrized such that for > exactly E( p()) E( p( )) = . In
particular bounds the slope of the boundary I , wherever that can be dened
at all. Figure 3 illustrates how the pairs (E, G) may be distributed in congu-
ration space according to Propositions 6, 7 and 8. For each capacity bound E
there exists a global solution G

of the equality constraint problem max G( p) s.t

E( p) = E which represents the upper boundary of the range of feasible cong-
urations (E( p), G( p)) for all p P with E( p) = E. The existence of these
global optima is ascertained analogous to Proposition 6(ii). The vertical dashed
lines depict ve such pairs at the bounds labeled E
1
, E
2
, E
3
, E

, and E
4
. The
Fig. 3 Illustration of the behavior of the pairs (E( p), G( p)) for all the admissible proles p
1 3
46 F. Alvarez et al.
one at E

represents the global solution of (FOP), whose existence is ascertained

by Proposition 6(i). That particular pair (E

, G

) and also (E
1
, G
1
) but not the
other three belong also to the curve of solutions (E( p

), G( p

)) with p

dened as
the global minimizer of G

. Its existence and monotonicity is guaranteed by Prop-

osition 7. The images (E( p

), G( p

)) form the Northwestern border of the con-

vex hull of the set of congurations I. Let 1/
1
> 1/
4
denote the slopes of the
two convexifying dashed lines. Note that this numbering is independent of that of
the previous ve pairs. Consequently, for <
1
the points p

are the unique global

minimizers of G

( p) and vary continuously as (0;

1
]. Here G

1
has at least two
global maximizers p

1
and p

2
which are quite some distance apart and connected
with the rst convexifying dashed line. For (
2
,
3
] the optimal solutions p

move
continuously along the boundary from p

2
to p

3
and afterwards for in [
3
,
4
] the
prole p

stays constant until there is another jump to p

4
. Then there is another con-
tinuous variation as tends to innity and hence the slope to zero, which is reached at
p

. Note that the original solutions (E

2
, G
2
), (E
3
, G
4
) and (E
4
, G
4
) of the equality
constraint problem are not reached by the path p

. More specifically the pair.

(E
1
, G
1
) represents a global maximum of G subject to the constraint E( p) E
1
which occurs also on the path p

. The pair (E
2
, G
2
) represents also a sensible global
maximum of (CFOP) but it can not be reached along the path p

. The pair (E
3
, G
3
)
represents at best a local maximum of G( p) s.t. E( p) E
3
but not the global max-
imum. The pair (E
4
, G
4
) may look like a global maximum but one can do better by
simplygoingtothe ultimate gainpit p

whichrenders the effort constraint E( p) E

4
inactive for this specic illustration.
4 Dynamic trajectory planning
In the previous section we have established the existence of optimal gain proles with-
out and with excavation constraints. Rather than solving just this stationary problems
one is really interested in an trajectory of proles that gets to the valuable material
as fast as possible. In other words we are interested in maximizing the present value
based with a certain discount function for future earnings.
We consider paths P : [0, T] P that are monotonic, i.e. s, t [0, T], with
s t imply for p = P(t ) and q = P(s) that q(x) p(x) for x . Naturally,
the function E(P(t )) must be also monotonically increasing. We assume that there
exists an absolutely continuous function C : [0, T] R
+
with
C(t )
t
_
0
c()d
representing the mining capacity in the time interval [0, t ], with density c L

(0, T)
and c 0. Finally, we impose the capacity condition on P
1 3
A continuous framework for open pit mine planning 47
E([P(s), P(t )]) = E(P(t )) E(P(s)) C(t ) C(s)
=
t
_
s
c()d for s t. (22)
Now we introduce the set of feasible excavation paths:
U = { P C([0, T]; P) |
p
0
P(s) P(t ), E([P(s), P(t )]) C(t ) C(s) for 0 s t T}.
Proposition 9 Under assumption (22), for any P U we have that
P(t ) P(s)

_
c

e
0

+2
_
(t s)
1/3
. (23)
Hence the elements of U are Hlder equicontinuous and U is compact in C([0, T]; C
()) which is identied with C([0, T] ) endowed with the uniform norm

.
Proof For the proof of compactness we will apply again the Arzela-Ascoli theorem.
The closedness of U is direct from its definition together with the continuity of E on
C([0, T]; P) endowed with the norm

. In fact, as for all t [0, T] we have

{P(t )}
PP
P with P being compact by Proposition 2, it only remains to prove
that U is equicontinuous. To this end, x x
0
and take s, t [0, T] with s < t .
Let P U be arbitrary. For any x , as P(s), P(t ) P, by global Lipschitz
continuity of each prole we have that
P(t )(x) P(s)(x) 2x x
0
+ P(t )(x
0
) P(s)(x
0
).
Multiplying by the lower bound e
0
> 0 on e(x, z) and integrating with respect to x
on the ball |x x
0
| for some > 0 sufciently small which we will choose later
on, we get
e
0

2
[2 + P(t )(x
0
) P(s)(x
0
)]

P(t )(x)
_
P(s)(x)
e(x, z)dzdx
= E(P(t )) E(P(s)),
where we have supposed that is an open subset of R
n
with n = 2 (the case n = 1
is similar) and that the ball B

(x
0
) is contained in . Therefore
e
0

2
[2 + P(t )(x
0
) P(s)(x
0
)]
t
_
s
c()d c

(t s),
1 3
48 F. Alvarez et al.
Hence we deduce that for all > 0 small enough, we have
0 P(t )(x
0
) P(s)(x
0
)
c

e
0

t s

2
+2L.
If t s > 0 is small then we can take for instance
3
= t s to obtain
0 P(t )(x
0
) P(s)(x
0
)
_
c

e
0

+2
_
(t s)
1/3
,
and since x
0
is arbitrary, we get (23). This proves the equicontinuity of U .
Concerning the internal structure of U , we rstly note that for arbitrary P, Q U ,
neither max{P, Q} nor min{P, Q} need to be elements of U for all t , as both may
violate the dynamic capacity constraint (22). In fact, it is easy to construct simple
examples where such a situation occurs. As a consequence, the lattice structure is
lost in the dynamic case. Nevertheless, U is path connected as a consequence of the
following Lemma.
Lemma 3 Let P U and p Pbe given. Then min{ p, P(t )} U max{ p, P(t )}
for all t [0, T]. Hence any P(t ) is connected to the trivial path p
0
via the path
P
r
(t ) min{ p
0
+r, P(t )} for r [0, z z].
Proof We have p
0
min{ p, P(s)} and p
0
max{ p, P(t )}. As the path P is mono-
tonic, both estimations are true for all s, t [0, T] as well. Now consider the path
min{ p, P(t )}. We choose s t [0, T] arbitrary and dene the following subsets of
:

1
{x | p(x) < P(s)(x) P(t )(x)},
2
{x | P(s)(x) p(x)
P(t )(x)},
3
{x | P(s)(x) P(t )(x) < p(x)}. Of course =
1

3
.
For (22) we have
_

min{ p,P(t )}(x)

_
min{ p,P(s)}(x)
e(x, z)dzdx =
_

1
p(x)
_
p(x)
e(x, z)dzdx
. , .
=0
+
_

2
p(x)
_
P(s)(x)
e(x, z)dzdx
. , .

P(t )(x)
_
P(s)(x)
e(x,z)dzdx
+
_

3
P(t )(x)
_
P(s)(x)
e(x, z)dzdx

P(t )(x)
_
P(s)(x)
e(x, z)dzdx
t
_
s
c()d
The last estimation holds because P is a feasible path. For the max operation the
argument is analogous, so the proof is complete.
Let C
1
(0, T) be a monotonically decreasing discount function starting from
(0) = 1 and ending at (T) < 1 for some xed time period [0, T]. The standard
1 3
A continuous framework for open pit mine planning 49
choice is (t ) = e
t
for some discount rate > 0. For paths P(t, x) P(t )(x) that
are smooth in time we dene the present value of the gain as

G(P)
T
_
0
(t )
_

g(x, P(t, x)) dx dP(t ) =

_

T
_
0
(t )g(x, P(t, x))P

(t )dt dx
where the notation dP suggest that P(t, x) must be differentiable w.r.t. t . Integrating
by parts we can avoid this requirement and obtain

G(P) =
_
_
_
(t )
_

P(t,x)
_
P(0,x)
g(x, z)dzdx
_

_
T
0
+
_

T
_
0
[

(t )]
P(t,x)
_
P(0,x)
g(x, z)dzdt dx
= (T)
_
_
_

g(x, P(T, x)) g(x, P(0, x))dx

_
_
+
_

T
_
0
[

(t )]
_
g(x, P(t, x)) g(x, P(0, x))
_
dt dx (24)
where g(x, z)
_
z
p
0
(x)
g(x, )d is the antiderivative based on the initial prole
which is already familiar fromthe proof of Proposition 5(iv). This formis feasible even
if the Path P does not satisfy P(0, ) = p
0
(). We note that for feasible excavationpaths
the rst term in (24) is the value of the total excavated material of the path discounted
by (T). The last term is a correction due to the variation of (with

(t ) > 0).
This representation of

G(P) is well dened for every path in C([0, T] ).
The optimization problem in the dynamic trajectory planning case, the so called
Capacitated Dynamic Open Pit Problem, is the following one
(CDOP) max{

G(P) | P U , P(0) = p
0
}
For the objective function

G we obtain the following results.
Proposition 10 For arbitrary paths P, Q U we have
(i)

G(P)

G(Q)

2|| g

P Q

(ii) If g(x, z) is decreasing w.r.t. z then

G is concave on the feasible set of (CDOP).
(iii) For any p P we have

G(P) =

G(min{P, p}) +

G(max{P, p}) for any path
of the feasible set of (CDOP).
Proof (i) Since g

is a Lipschitz constant for g, with respect its second

argument we can bound the difference in the rst term of

G in (24) by
2(T)|| g

P Q

as g(x, P(0, x)) is not necessarily zero. Since

(t ) 0 by assumptions we can similarly bound

1 3
50 F. Alvarez et al.

T
_
0
[

(t )]
_
_
_
P(t,x)
_
Q(t,x)
g(x, z)dz
P(0,x)
_
Q(0,x)
g(x, z)dz
_
_
_
dt dx

T
_
0
[

(t )] (| P(t, x)Q(t, x)|g

+| P(0, x)Q(0, x)|g

) dt dx

T
_
0
[

(t )]2P Q

dt dx
=
_

2 g

P Q

((0) (T))dx
This is equal to (1(T))2|| g

P Q

. Summing the terms cancels

out (T) and the Lipschitz constant becomes completely independent of the
discount function.
(ii) This is analogous to the proof of Proposition 5(ii) by taking into account that
for any feasible path of (CDOP) g(x, P(0, x)) = 0 holds. We omit the details.
(iii) Let
p
(t ) {x | p P(t )} be the subset of the domain where the path
is pointwise greater than the given prole. We have

G(P) = (T)
_
_
_
_
_

p
(T)
_
_
_
g(x, p(x)) +
P(T,x)
_
p(x)
g(x, )d
_

_
dx
+
_

p
(T)
g(x, P(T, x))dx
_
_
_
_
+
T
_
0
[

(t )]
_
_
_
_
_

p
(t )
g(x, p(x))dx +
_

p
(t )
P(t,x)
_
p(x)
g(x, )d dx
+
_

p
(t )
g(x, P(T, x))dx
_
_
_
_
dt
This yields to

G(P) = (T)
_
_
_
_
_

p
(T)
g(x, p(x))dx +
_

p
(T)
g(x, P(T, x))dx
_
_
_
_
1 3
A continuous framework for open pit mine planning 51
+
T
_
0
[

(t )]
_
_
_
_
_

p
(t )
g(x, p(x))dx +
_

p
(t )
g(x, P(T, x))dx
_
_
_
_
dt
+(T)
_

p
(T)
P(T,x)
_
p(x)
g(x, )d dx+
T
_
0
[

(t )]
_

p
(t )
P(t,x)
_
p(x)
g(x, )d dxdt
=

G(min{P, p}) +

G(max{P, p})
The rst two lines of the sum indeed represent the discounted gain of the path
P stopped at the prole p.
By combining Propositions 9 and 10 we conclude immediately that

G attains global
maximizer on U . Without any additional assumption and for a limited time horizon
T < one can quite easily construct examples where the set of global minimizers
is disconnected and may jump around violently when the gain density is slightly per-
turbed. One only has to think of two areas with high gain in separate places that have
nearly identical values and excavation costs.
Due to the Lipschitz continuity of

G on the Banach space C([0, T] ), it has
similar differentiability properties as those stated in Proposition 5 for G. However as
it is not possible to guarantee that the optimizers are not elements of the meager set
of exceptional points where one may not even have Gteaux differentiability, opti-
mality conditions cannot be formulated in terms of classical derivatives. For example
the minimization of the maximum of nitely many distinct afne functions in one
variable will almost certainly lead to a point where two of them tie. There zero is in
the subdifferential, i.e. the convex hull of the two adjacent slopes but is not a proper
derivative. Only where the maximum of the afne functions happens to be attained by
a constant function can there be differentiability at a minimizer.
We conclude with a relation to the stationary problem (FOP). Recall that p

is a
solution of FOP accordingly with (20).
Proposition 11 Let C
1
([0, T]) be monotonically decreasing. Then
(i) If

G(P) > 0 for some P U with P(0) = p
0
then G

G( p

) > 0.
(ii) If

G

is the optimal value of (CDOP) then

G

= max{

G(P) | P U

}, where
U

U C([0, T]; P

) for P

{ p P | p p

}.
Proof (i) Suppose G

= G( p

) 0 and choose P U arbitrary. Then we

have for any t also G(P(t )) =
_

g(x, P(t, x))dx 0. Otherwise we would

have G(P(t )) > G

which contradicts the definition of G

. Substituting
this relation into the integrated form (24) immediately yields

G(P) 0 which
completes the proof of (i).
(ii) Consider a P U with
p

(t ) = from a certain time t < T. By Proposi-

tion 10
1 3
52 F. Alvarez et al.
(iii) we know we can decompose the objective in the following way.

G(P) =

G(min{P, p

}) +(T)
_

p
(T)
P(T,x)
_
p

(x)
g(x, )d dx
+
T
_
0
[

(t )]
_

p
(t )
P(t,x)
_
p

(x)
g(x, )d dxdt
The last two terms, representing the gain the path yields between p

and
P(T), are an integral over the part where the path is pointwise larger than p

.
Hence, because p

is the solution of (FOP), they can not be positive and for

the stopped path we have

G(min{P, p

})

G(P). Thus any maximizer of

G
is bounded by the solutions of (FOP).
By virtue of Prop. 11(ii) for maximizing the discounted gain it sufces to consider
feasible proles such that p p

. This is a continuous analogue of a well known

property of standard binary formulations for these type of problems, namely that the
optimal Final Open Pit is an upper bound on the region to be considered of interest
for the nested sequence of proles, a property which is used to reduce the size of the
original block model.
5 Concluding remarks
As we mentioned in the introduction, to the best of our knowledge, the rst attempts to
give rigorous mathematical formulations of open pit mine planning problems based on
continuous functions date back to the works by Matheron in the 1970s. For instance, in
Matheron (1975) is developed a general measure-theoretic approach to a parametrized
version of the CFOP problem, obtaining the analogous to Propositions 3(i), 6(i) and
7 of this paper. No significant theoretical contribution seems to haven been made. A
complete comparison with Matherons approach is out of the scope of this paper but
will be addressed in a forthcoming work currently under preparation.
In our approach, the optimization problems are posed in functional spaces and a
complete existence theory holds. In fact, under realistic assumptions, the feasible set
of proles in the stationary case as well as the set of paths of them in the dynamical
case are compact sets in suitable Banach spaces of real valued functions, which implies
the existence of solutions as the objective functions and constraints are proven to be
(Lipschitz) continuous. In addition, we have provided structural properties of feasible
sets and optimal solutions in the stationary case, and we have developed a parametric
qualitative analysis of the behavior in the effort-gain conguration space. We have
also obtained sufcient regularity conditions for differentiability of effort and gain
functions, under the assumption that the effort and gain densities are continuous.
The optimization problems formulated in this paper are examples of variational
problems in Banach spaces with a special structure. We have focused on the existence
1 3
A continuous framework for open pit mine planning 53
theory, and there remain plenty of interesting open questions. The rst one is how to
obtain useful necessary optimality conditions under weak regularity assumptions. As
the reference Banach space is non-reexive, optimality conditions as those presented
in Bazaraa et al. (1976) are not valid. Also we lack Slater-type conditions, which
require the feasible set to have interior points. To avoid these requirements one has
to generalize the concepts of convexity and interior points; see, for instance, Borwein
and Goebel (2003), Borwein and Lewis (1992) while a short introduction can be found
in Daniele et al. (1994).
The approximate numerical solution of the continuous problems is beyond the
scope of this paper. Instead of trying to resolve optimality conditions on function
space once they are formulated, a natural idea is to consider a so-called direct method,
i.e. to discretize rst. For example one may approximate the admissible proles by
piecewise linear-afne functions or more general polynomials, thus obtaining nite
dimensional NLP problems. One complication is that due to the presumed discontinu-
ity of the data one cannot restrict the approximations to the solution to be continuous
either. On the other hand, if the original innite dimensional problem is convex, a
suitable discretization scheme may preserve this property so that the resulting nite
dimensional optimization problem may be solved by algorithms for convex NLP (see
Boyd and Vandenberghe 2004). Moreover, one might exploit some of the established
structural properties for algorithmic purposes. Indeed, the parametric characterization
of the stationary optimal proles given in (20) leads naturally to a nite dimensional
Bilevel Programming problem. Finally, if one were able to design a discretization
method which preserves the lattice structure of the stationary admissible set, one might
adapt to this setting some efcient algorithms coming from discrete optimization (see
McCormick 2006). We plan to investigate these possibilities in future research.
Finally, we would like to make clear that the continuous approach proposed here is
not intended to substitute the more traditional discrete techniques, but to supplement
them with additional tools coming from continuous optimization in functional spaces.
On one hand, as the underlying discrete block models become larger in terms of the
number of unitary extraction blocks, the continuous formulations may be viewed as
a sort of limiting averaged model, which should provide qualitative and quantitative
information about the behavior of optimal solutions froma macroscopic point of view.
In particular, continuous optimization is potentially well suited for investigating the
sensibility w.r.t. discount rates or extraction effort capacities, which are very inter-
esting issues for future research. On the other hand, the approximate resolution of
a macroscopic continuous framework might be useful to obtain insight on how to
construct good starting points for discrete formulations based on namely microscopic
block models.
The connections between the continuous and discrete approaches, in terms of the-
oretical and algorithmic aspects, and the question on how we might use them from
a practical point of view, are clearly key points that will be addressed in our future
research.
Acknowledgments The authors are grateful to Maurice Queyranne for his comments on an earlier version
of this paper which help us to improve the presentation, and specially for pointing out to us the works by
Matheron (1975) and Topkis (1978). The authors are also indebted to amanonymous referee for pointing out
1 3
54 F. Alvarez et al.
some inconsistencies and suggesting modications to improve the readability. The research of the rst two
authors was supported by FONDEF grant D06-I-1031 and FONDAP-BASAL programs from CONICYT.
The rst author thanks also the Institute on Complex Engineering Systems (ICM: P-05-004-F, CONICYT:
FBO16). The third and fourth author acknowledge the support from the DFG Research Center Matheon
Mathematics for Key Technologies, Berlin.
References
Bazaraa M, Shetty C, Goode J, Nashed M (1976) Nonlinear programming without differentiability in
Banach spaces: necessary and sufcient constraint qualication. Appl Anal 5:165173
Boland N, Fricke C, Froyland G (2006) A strengthened formulation for the open pit mine production
scheduled problem. Preprint, University of Melbourne, Parkville, VIC 3010, Australia
Borwein J, Goebel R (2003) Notions of relative interior in Banach spaces. J Math Sci 115:25422553
Borwein J, Lewis A (1992) Partially nite convex programming, Part I: quasi relative interiors and duality
theory. Math Progr 57:1548
Boyd S, Vandenberghe L (2004) Convex optimization. Cambridge University Press, Cambridge
Cacetta L, Hill SP (2003) An application of branch and cut to open pit mine scheduling. J Glob Optim
27:349365
Cacetta L (2007) Application of optimisation techniques in open-pit mining. In: Weintraub A et al (ed)
Handbook of operations research in natural resources. Springer, New York
Daniele P, Giuffr S, Idone G, Maugeri A (1994) Innite dimensional duality and applications. Math Ann
339:221239 (2007)
Denby B, Schoeld D, Open pit design and scheduling by use of genetic algorithms. Trans Inst Min Met
(Sec. A: Min. Industry) 103:A2A26
Evans L (1998) Partial differential equations. Graduate Studies in Mathematics, vol 19. American Mathe-
matical Society, Providence
Ferland J, Amaya J, Djuimo MS (2007) Particle swarm procedure for the capacitated open pit mining
problem. In: Autonomous Robot and Agents. Studies in Computational Intelligence, Book Series.
Springer, Berlin
Guzmn JI (2008) Ultimate Pit Limit determination: a new formulation for an old (and poorly specied)
problem. Communication in the Workshop on Operations Research in Mining, December 1012, Via
del Mar, Chile
Hochbaum D, Chen A (2000) Performance analysis and best implementation of old and new algorithms
for open-pit mining problem. Oper Res 48:894914
Hustrulid W, Kuchta M (2006) Open pit mine planning and design, vol 1: Fundamentals. Taylor & Francis,
London
Johnson TB, Sharp WR (1971) A Three-dimensional dynamic programming method for optimal ultimate
open pit design. Report of Investigation 7553, U.S. Bureau of Mines
Lerchs H, Grossman IF (1965) Optimum design of open pit mines. Trans CIM 58:4754
Matheron G (1975) Parametrage de contours optimaux, Note Geostat. No. 128. Centre de Geostatistique,
Fontainebleau
McCormick ST (2006) Submodular function minimization, Chap. 7. In: K Aardal, G Nemhauser,
R Weismantel (eds) The handbook on discrete optimization. Elsevier, Amsterdam pp 321391
Morales N (2002) Modelos Matemticos para planificacin minera. Engineering Thesis, Universidad de
Chile, Santiago, Chile
Preiss D, Zaj cek L (2001) Directional derivatives of Lipschitz functions. Isr J Math 125:127
Topkis DM (1978) Minimizing a submodular function on a lattice. Oper Res 26:305321
Wilke FL, Wright EA (1984) Determining the optimal ultimate pit design for hard rock open pit mines
using dynamic programming. Erzmet 37:139144
1 3