Bi-Objective Optimization: A Pareto Method with Analytical Solutions ()
1. Introduction
Many real life optimization problems require that two or more objectives under analysis be optimized simultaneously. Frequently, these objectives conflict with each other, and it is not possible to find a single platform that maximizes all objectives simultaneously. Among them are the cases with two conflicting objectives such as: inflation and unemployment, risk and returns, environmental preservation and national income, current enjoyment and future education, and short-term profit and future growth, etc. Studies in bi-objective optimization constitute a non-trivial part in multi-objective analyses. For instance, Zhou et al. [1], Kukkonen and Deb [2], Pinto-Varela et al. [3], Lath et al. [4], Pereyra et al. [5], Garg [6], Futrell et al. [7], Hirpa et al. [8], Liu et al. [9], Wang et al. [10], Cheraghalipour et al. [11], Ho-Huu et al. [12], Yeh [13], Liu et al. [14], Nagamanjula and Pethalakshmi [15], Xu et al. [16], Diao et al. [17], Mohammadi et al. [18], Kparib et al. [19], Kparib et al. [20], Gulben and Orhan [21], Zaninudin and Paputungan [22], and Stutzle and Hoos [23]. Studies proposing multi-objective optimization techniques and solution can be found in Messac [24], Das and Dennis [25], Deb [26], Messac et al. [27], Messac and Mattson [28], Kim and Weck [29], Zhang and Li [30], Chinchuluun and Pardalos [31], Mueller-Gritschneder et al. [32], Pereyra et al. [5], Pérez-Fernández et al. [33], Marler and Arora [34], Gunantara [35], Orths et al. [36], Collette and Siarry [37], Ehrgott [38], Eskelinen et al. [39], Fonseca and Fleming [40], Alaa et al. [41], Subhamoy and Sugata [42], Wilfried and Blum (2014) [43], Caramia and Dell’Olmo (2008) [44], Rohilla (2020) [45], Engau and Wiecek [46], Obayashi et al. [47], Lagarias et al. [48], Miettinen [49], Zhang and Li [30], Bendsoe et al. [50], and Chankong and Haimes [51].
In these studies, several methods are commonly used for constructing aggregation functions, they include the weighted sum, Tchebycheff inequality, the normal boundary intersection, the normal constraint method, the Physical Programming method, Goal Programming, the epsilon constraints and Directed Search Domain to approximate the preference of the decision-maker. Often very lengthy computational efforts have to be invested and may end up with insufficient number of Pareto optimal points to be considered.
A crucial goal of a multi-objective optimization problem is to construct the Pareto optimal front (POF), which depicts the best trade-offs among the objectives to be optimized. The POF can be approximated as the solution of a series of scalar optimization subproblems in which the objective is an aggregation of the objectives. This paper presents a new Pareto Method for bi-objective optimization yielding the POF in the form of analytical solutions. An analytical solution involves framing the problem in a well-understood form and deriving exact solution. The analytical method is often preferred because its solution is in exact closed form.
Analytical solutions have three important advantages:
1) Transparency: Analytical solutions are presented as mathematical expressions, they make the effects of variables and their interactions with each other explicit.
2) Efficiency: Usually, algorithms and models expressed with analytical solutions are more efficient for manipulation and analysis than numerical analysis. Specifically, it is often faster, more accurate, and more convenient to evaluate an analytical solution than to perform an equivalent numeric implementation.
3) Mathematical Rigor: Analytical methods are rigorous and provide exact solutions with high tractability.
This paper is organized as follows. Bi-objective optimization problem is formulated in Section 2. Derivation of POF with equality constraints is provided in Section 3. Section 4 presents different analytical Pareto solutions with equality constraints. Section 5 derives the POF in cases with equality and inequality constraints. Analytical Pareto solutions under equality and inequality constraints are examined in Section 6. An Illustrative example is given in Section 7. Extension and conclusion are provided in Section 8 concludes.
2. Bi-Objective Optimization Problem
Consider a bi-objective optimization in which the decision-maker faces two objectives:
and
. The problem becomes
,
subject to
. (2.1)
where
is a set of decision variables which values are to be chosen in the optimization problem.
The feasible set of decision variables
is implicitly determined by a set of equality constraints and a set of inequality constraints,
and
, (2.2)
where
is a m-dimensional vector of functions, and
is a τ-dimensional vector of functions.
The objectives
and
are functions which measures the effects of the decision variables x on the objectives
and
. The function
represents the ranking preference of different combinations of
and
. It can take various functional forms, contingent upon the preference or targets fixed by the decision-maker.
The problem defined in (2.1)-(2.2) belongs to the class of constrained multi-objective optimization problems. There are a number of methods designed to assist the decision maker to arrive at the best compromise solution.
1) Scalarization: The most commonly used methods adopt schemes to convert the multiple objectives into a single scalar objective and apply standard scalar optimization algorithms to generate an optimal solution. Various weighted schemes to scalarize the multiple objectives into a scalar function are available, such as weighted global methods, weighted sum methods and exponential weighted criterion. One of the problems in scalarization is the existence of conflicting objectives.
2) Utility-Based Optimization: Another solution for multi-objective optimization is to explicitly consider the possible trade-offs between conflicting objective functions. Such trade-offs can be analyzed on the basis of the utility that these compromises have for the decision-maker. Many studies considered the utility-based optimization should be a common standard in multi-objective optimization.
3) Axiomatic Solution: Often the decision-maker cannot concretely define what he prefers. Axiomatic solutions like the Nash arbitration scheme can be chosen. Based on predetermined axioms of fairness, the solution suggests an arbitration yielding the maximum (over a convex compact set of points) of the product of the players’ utilities. In this case, the utility functions always have non-negative values and have a value of zero in the absence of cooperation. It can also be generalized to become weighted product methods. Similarly, the Kalai-Smorodinsky solution is another solution to bargaining problems of utility maximizing players. In multi-objective problems, players’ utilities are replaced by objectives that the decision-maker aims to maximize simultaneously.
4) Goal Programming Method: Finally, the decision-maker may consider a goal programming solution. In particular, the decision-maker aims to reach or getting as close as possible to a goal or a vector of targets.
In Section 4, we consider five methods with analytical solutions. Specifically, they are Nash arbitration and objective product method, target-attainment method, Kalai-Smorodinsky bargaining solution, scalarization method with weighted-sum and utility-based method.
3. Derivation of POF with Equality Constraints
We first consider the case where there are only equality constraints in the decision variables as a bench mark. (This corresponds to the case where the inequality constraints are either absent or inactive). A way to obtain Pareto efficient strategies in the bi-objective optimization problem is through the weighted-sum method. Such approach is also employed in identifying the players’ cooperative strategies belonging to the Pareto optimal set in non-transferrable utility games (see [52] [53] [54] ). In particular, the POF can be traced out by identifying the Pareto efficient strategies through systematically changing the weights among the objective functions. Therefore, the decision-maker considers the problem:
, for
,
subject to
. (3.1)
The corresponding Lagrange function can be expressed as:
(3.2)
where
is the set of Lagrange multipliers, and
and
, for
, are the weight for the objective 1 and objective 2 respectively.
First-order conditions for a maximum yield
for
,
, for
. (3.3)
If the system of
first-order conditions in (3.3) satisfies the implicit function theorem, one can express the optimal decision variables
and the corresponding Lagrange multipliers
as functions the exogenous parameter
, that is
, for
,
, for
. (3.4)
Substituting the optimal decision variables
from (3.4) into the objectives
and
, we can obtain the optimal objectives under
as:
and
, (3.5)
where
.
In the case where
, it generates the anchor point where the best of objective
is obtained, that is
. In the case where
, it generates the anchor point where the best of objective
is obtained, that is
. The Pareto optimal frontier (POF) can be obtained as
, for
, (3.6)
which is analytically tractable.
An increase in the value of
signifies an increase in the weight for objective
and a decrease in the weight for objective
. Hence the POF is downward sloping in the
space.
The point
is an anchor point at which the objective
reaches its maximum. Similarly, the point
is an anchor point at which the objective
reaches its maximum. The point
is the utopia (ideal) point at which
reaches its maximum and
reaches its maximum simultaneously.
In addition, if there exist minimum levels of the objectives,
and
, that the optimal solution have to fulfilled, then the range of the POF has to be restricted to be above
and above
. The corresponding restriction on the weight can be obtained as
, where
and
.
The point
is called the nadir point. The point
becomes an anchor point at which the objective
reaches its maximum. The point
becomes an anchor point at which the objective
reaches its maximum. The point
becomes the utopia point.
The POF is inside the rectangle bounded the nadir point, the utopia point and the two anchor points.
The part inside the area bounded by the nadir point, two anchor points and the curve of the POF in Figure 1 are dominated points. The part inside the area bounded by the utopia point, two anchor points and the curve of the POF in Figure 1 are unreachable points. Since the decision-maker would not choose a
Figure 1. POF under equality constraints.
dominated point and could not reach unreachable points, any optimal solution chosen by the decision-maker would be on the POF.
4. Analytical Pareto Solutions with Equality Constraints
In this section, we consider various solution methods via the analytical solution of the POF derived in Equation (3.6) under equality constraints only.
4.1. Nash Arbitration and Objective Product Method
The Nash objective Product maximization seeks a solution which yields the maximum of the product of the objectives in the feasible decision region. The idea is derived from Nash [55] and applied by Davis [56] in multi-objective optimization. Consider Figure 1, the feasible decision region is the POF bounded by the vertical line
and the horizontal line
. The maximization of the product of the relevant objectives can be expressed as:
. (4.1)
Performing the maximization operator in (4.1) we obtain the condition
(4.2)
The weight
that satisfies (4.2) yields the solution to the objective product maximization method can be obtained as
.
4.2. Target-Attainment Method
In the target-attainment method, the decision-maker aims to reach a target or a vector of targets. For instance, the target for
is to reach
and the target for
to is reach
. The objective is to minimize the deviation of the solution from the targets. One can depict the explicitly derived POF and compared to the target
.
If the target point
is outside the POF, the problem becomes minimizing the distance between the POF and the point
indicated by the dotted line in Figure 2, that is
. (4.3)
The solution to (4.3) will be characterized by the condition
(4.4)
We can derive the weight
that satisfies (4.4), and obtain the solution
.
Consider the case that the target is
must be attained. We first identify the weight
such that
. (4.5)
The solution is then
.
4.3. Kalai-Smorodinsky Bargaining Solution
Aboulaich et al. [57] and Oukennou et al. [58] applied the Kalai-Smorodinsky Bargaining Solution [59] for solving multi-objective optimization problems. The Kalai-Smorodinsky solution is a solution to bargaining problems of utility maximizing players. In multi-objective problems, players’ utilities are replaced by objectives that the decision-maker aims to maximize simultaneously. The main advantage of the solution is that it yields a concrete criterion to select one and only one unique point along the POF. Mathematically, it is the intersection of the POF and the line segment connecting the nadir point and the utopia point.
The nadir point is
. To obtain the utopia point, we first identify the
that satisfies
, and denote it by
. The point
is the top anchor point of the POF. Similarly, we identify the
that satisfies
and denote it by
. The point
is the bottom anchor point of the POF. Using the top anchor point and the bottom anchor point of the POF, we can obtain the utopia point as
.
The slope of the line segment connecting the nadir point and the utopia point can be obtained as
, which denote by
. To obtain the Kalai-Smorodinsky solution, we trace the
satisfying
. (4.6)
Let
denote the
that satisfies (4.6). The Kalai-Smorodinsky solution can be obtained as
. Graphically, the Kalai-Smorodinsky bargaining solution is the point of intersection of the POF and the line joining the Nadir point and the Utopia point in Figure 3.
Figure 3. Kalai-Smorodinsky bargaining solution.
4.4. Scalarization Method with Weighted-Sum
The scalarization method makes the multi-objective function create a single solution and the weight is determined before the optimization process. The scalarization method incorporates multi-objective functions into scalar fitness function as in the following equation [60].
. (4.7)
The weight of an objective function determines the solution and reveals the performance priority [61]. A large weight that is given to an objective function that has a higher priority compared to the ones with a smaller weight. Normalizing
the weights
and
, we can obtain
and
. The solution can be obtained as
.
4.5. Utility-Based Method
Rădulescu et al. [62] considered utility-based analysis to be the standard paradigm for studying multi-objective problems. In particular, they argued that compromises between competing objectives in MOMAS should be analyzed on the basis of the utility that these compromises have for the users of a system, where an agent’s utility function maps their payoff vectors to scalar utility values. The utility of different combinations of objectives is given by the utility function
. It represents a scalarization of the objectives into a preference ranking index. It can be linear or nonlinear. If the utility function is linear, it resembles a scalarization of the objectives with weighted-sum of the objectives. Very often, nonlinear utility function
yields a set of indifference (level) curves of preferences which are convex, showing diminishing marginal rate of substitution between the objectives. Such utility functions represent a nonlinear scalarization of the objectives.
Consider the case where the utility function
. It yields in difference (level) curves which are convex and showing diminishing marginal rate of substitution between the objectives.
The maximization of the utility function
can be expressed as:
. (4.8)
Perform the maximization operator in (4.8) we obtain the condition
(4.9)
The weight
that satisfies (4.9) yields the solution of maximizing
with
and
. The point where the POF and the indifference curve are tangent to each other demonstrates the utility-based solution Figure 4.
4.6. Performance of Pareto Method with Analytical Solution
In general, multi-objective optimization requires huge computational effort. Frequently an insufficient number of Pareto optimal points will be found. Pareto methods usually require less complicated mathematical equations. The solution using the Pareto method is a performance indicators component that produces a compromise solution and can be displayed on the Pareto optimal front. Obtaining a Pareto optimal solution set is preferable to a single solution. It provides a basis upon which to make value judgment’s in order to settle on a final solution.
Pareto method with analytical solution involves the framing the problem in a well-understood form and deriving exact solution. The method is often more preferred because its solution is in exact closed form. A wide range of the POF can be traced out analytically with relevant mathematical expressions. The method is efficient for manipulation and analysis than numerical analysis. Specifically, it is often faster, more accurate, and more convenient to evaluate an analytical solution than to perform an equivalent numeric implementation. In addition, the effects of variables and their interactions with each other and parameter changes are highly tractable. Finally, the availability of the POF (or its relevant parts) in closed form allows the decision-maker to compare solutions under different criteria for multi-objective optimization.
5. POF with Equality and Inequality Constraints
To complete the analysis, we consider the case where there are equality and inequality constraints in the decision variables.
5.1. Pareto Efficient Strategies
Again, we identify the Pareto efficient strategies by systematically changing the weights among the objective functions. Specifically, the decision-maker considers the problem:
, for
, (5.1)
subject to
and
. (5.2)
To solve the optimization with equality and inequality constraints, we invoke the Karush-Kuhn-Tucker conditions and use the Lagrange multipliers approach with the corresponding Lagrange function:
(5.3)
where
and
are the sets of Lagrange multipliers. Necessary conditions for a maximum include:
for
,
, for
,
, for
; (5.4)
, for
, and
, for
. (5.5)
In the case where
, the inequality constraint is binding with
being held and acts as an active constraint. In the case where
, the condition
does not have to hold and the constraint is inactive.
Equation system (5.4) gives rise to
equations for n decision variables
, m Lagrange multipliers
, and
Lagrange multipliers
.
Moreover, any admissible solution has to satisfy (5.5). If (5.5) is not satisfied, it means that the solution satisfying the first order conditions is either not in the region fulfilling the constraints, or has a negative Lagrange multiplier, which is not allowed for a maximum.
If condition (5.5) fulfilled and the first order conditions (5.4) for an interior solution satisfy the implicit theorem, one can express the optimal decision variables
and the corresponding Lagrange multipliers
and
as functions the exogenous parameter
, that is
, for
,
, for
,
, for
. (5.6)
5.2. The Corresponding POF
Substituting the optimal decision variables
from (5.6) into the objectives
and
, we obtain the optimal objectives under
as:
and
. (5.7)
The Pareto optimal frontier (POF) at the point which corresponds to the adoption of objective weight
can be obtained as
, for
, (5.8)
which is again analytically tractable.
Theoretically, the frame of the POF with both equality and inequality constraints can be delineated by computing the Pareto strategies for different values of
between 0 and 1. Note that the Pareto optimal point
may have to be calculated point by point for different values of
, because the set of active inequality constraints may vary as
changes. The POF with both equality and inequality constraints is bounded by the POF with equality constraint only. Unlike the case with equality constraints, we have to track down the corresponding point of the POF with individual values of
, and there exists the possibility that the solution satisfying the first order conditions is not in a feasible region bounded by the constraints. Therefore, the POF may have broken ranges as shown in Figure 5.
6. Analytical Pareto Solutions under Equality and Inequality Constraints
Note that various solution methods via the analytical solution of the POF derived in Section 4 yield a unique solution
. If the solution is in an area where all inequality constraints are inactive, the solution would be the same as that in section 4.1. If the solution is in an area where some inequality constraints are active, we first solve the first-order conditions in (5.4) for
in an area near
identified in Section 4. Specifically, we obtain
, for
,
, for
,
, for
, for
. (6.1)
The corresponding point of the POF can be expressed as
, for
. (6.2)
Then, we check whether the point derived in (6.2) with active inequality constraints still fulfills the optimality condition. If not, we have to identify some points on the POF in the adjacent area and search for the optimal solution.
For instance, consider the target-attainment method in Section 4.2 in which the target for
is to reach
and the target for
to is reach
. The decision maker seeks to minimize the deviation of the solution from the target
. We first identify the POF points for
under equality constraint only given in Section given in Section 4.2. Then we verify whether there exist active inequality constraints. If some inequality constraints are active in the solution point
, we have to consider some POF points at
in a neighborhood near
. We follow (5.3)-(5.5) and solve the problem with equality and inequality constraints under the weight
to obtain the Pareto efficient strategies and the corresponding POF. We let
, for
and
(6.3)
denote the optimal decision variables with the presence of active inequality constraints. We then calculate the distance between the target
and
, that is
, for
. (6.4)
Finally, the point
which yields the shortest distance in (6.4) is the solution for the target-attainment method.
7. An Illustrative Example
Consider a bi-objective optimization in which the decision-maker faces two objectives:
, (7.1)
and
. (7.2)
There is an equality constraint
, (7.3)
and an inequality constraint
. (7.4)
7.1. POF with Equality Constraint Only
We first consider as a bench mark the case with the equality constraint only. To obtain Pareto efficient strategies in the bi-objective optimization problem (7.1)-(7.4), the decision-maker considers the problem:
(7.5)
subject to (7.3).
The corresponding Lagrange function can be expressed as:
(7.6)
The Pareto efficient strategies of the problem of maximizing (7.5) subject to equality constraint (7.3) can be solved as:
Proposition 7.1.
The Pareto efficient strategies of the problem of maximizing (7.5) subject to equality constraint (7.3) are:
(7.7)
Proof: See Appendix A.
The relationship between the Pareto efficient strategies
,
and
can be obtained as follows.
Proposition 7.2.
,
,
.
Proof: See Appendix B.
Substituting the Pareto efficient strategies into the objective functions (7.1)-(7.2) yields the POF as:
(7.8)
In addition, if there exist minimum levels of the objectives,
and
, that the optimal solution have to fulfilled, then the range of the POF has to be restricted to be above
and above
. We denote the corresponding restriction on the weight as
. The values of
can be obtained by solving
. (7.9)
The values of
can be obtained by solving
. (7.10)
The point
(7.11)
becomes an anchor point at which the objective
reaches its maximum.
The point
(7.12)
becomes an anchor point at which the objective
reaches its maximum.
The point
(7.13)
becomes the utopia point.
7.2. POF with Equality and Inequality Constraints
Now, we consider the case under both the equality constraint and the inequality constraint. Invoking (7.4), one can observe that the inequality constraint will be active if
. To depict the POF, we first check whether the inequality constraint is active at
and
. If
, then the inequality constraint is active at
. Similarly, if
, then the inequality constraint is active at
. Since
is monotonically decreasing in
, the inequality constraint is active in the entire POF.
To obtain the Pareto efficient strategies, the decision-maker considers the problem:
(7.14)
subject to (7.3) and (7.4).
The corresponding Lagrange function can be expressed as:
(7.15)
The Pareto efficient strategies of the problem of maximizing (7.14) subject to equality constraint (7.3) and inequality constraint (7.4) can be solved as:
Proposition 7.3.
The Pareto efficient strategies of the problem of maximizing (7.14) subject to the constraints (7.3)-(7.4) are:
(7.16)
Proof: See Appendix C.
The values of
can be obtained by solving
. (7.17)
The values of
can be obtained by solving
. (7.18)
Substituting the Pareto efficient strategies from (7.13) into the objectives
and
, we can obtain the POF as
(7.19)
The corresponding anchor points and utopia point can be derived accordingly.
Finally, consider the case where
and
, we search the point at which the inequality constraint turns active, that is
. (7.20)
Solving (7.20) yields
. (7.21)
Therefore, the POF will be the same as that without inequality constraint in the range of
, and be the same as that with inequality constraint in the range of
. The actual POF is the solid line in Figure 6.
7.3. The Case of Kalai-Smorodinsky Solution
With the analytical solution of POF completely depicted, we can solve the solutions in Section 4. Consider the case of using the Kalai-Smorodinsky solution for solving multi-objective optimization problems. The solution is the intersection of the POF and the line segment connecting the nadir point and the utopia point. We first obtain the bench-mark POF with equality constraint only. Then, when check the anchor points in (7.11) and (7.12). If
in both anchor points, then the POF will be the same as that with equality constraint only. If
in both anchor points, then the POF will be the same as that with equality constraint and active inequality constraint. If
in the anchor point (7.11) and
in the anchor point (7.12) the inequality constraint is active, then the point
has to be identified as
(see (7.21)).
Given the above information, the relevant utopia point can be identified
and the nadir point is
. The slope of the line segment linking the nadir point and the utopia point in the area bounded by the nadir point and the utopia point is
. (7.22)
If there exist a
that satisfies
, and
, (7.23)
then, the Kalai-Smorodinsky solution is given by
(7.24)
If there exist a
that satisfies
, and
. (7.25)
then the Kalai-Smorodinsky solution is given by
(7.26)
Remark 7.1.
Note that with the part of POF under equality constraint and the part under inequality constraint as indicated explicitly in (7.17)-(7.21), we can characterize the solutions to the Nash arbitration, target-attainment method, scalarization method with weighted-sum and utility-based method in a similar way as that for the characterization of the Kalai-Smorodinsky bargaining solution.
8. Extension and Conclusion
The analysis can be extended to the case with more than two objectives separated into two competing/conflicting types of objectives. In particular, the type A objectives include
, and the type B objectives include
. A normalized weight is attached to each objective within a type, reflecting the relative importance of the objective in that group of objectives. The weighted sum of objectives within a type signifies the scalarized preference of the decision-maker for that type of objectives. The problem becomes
(8.1)
subject to
and
, (8.2)
where
,
,
and
.
We identify the Pareto efficient strategies by systematically changing the weights among the objective functions
and
. Specifically, the decision-maker considers the problem:
, for
, (8.3)
subject to (8.2).
Invoking Karush-Kuhn-Tucker conditions, we can express the corresponding Lagrange function as:
(8.4)
Following the analysis in Section 5, we attempt to establish an analytical path of the POF which would be used for obtaining the solution under different methods that are mentioned in Section 4.
Finally, this paper presents a new Pareto Method for bi-objective optimization yielding the POF in the form of analytical solutions. Analytical methods enjoy the advantages of being transparent, efficient and rigorous. These advantages are extremely useful in deriving accurate, exact and well-understood solutions, especially for policy design. The possibility to provide an extension for multi-objective optimization by separating the objectives into two types allows wider applicability of the developed results. This paper does not claim superiority of the analytical Pareto method over other methods of multi-objective optimization, rather the method is a novel addition to the growing pursuit of Pareto generators, with potential advantages of being handy for analysis. Further theoretical development and applications are expected.
Appendix
Appendix A: Proof of Proposition 7.1
First-order conditions for a maximum from the Lagrange function (7.6) yield
,
,
,
(A.1)
Solving (A.1) yields the Pareto efficient strategies and the Lagrange multiplier with equality constraint only
,
,
,
. (A.2)
Hence,
and
(A.3)
Appendix B: Proof of Proposition 7.2.
Differentiating
in Proposition 7.1 with respect to
yields
. (B.1)
Invoking the constraint
in (7.3) and the first two equations in (A.2), we have
, (B.2)
which shows that
.
Hence,
. (B.3)
In a similar manner, we can show that
. (B.4)
Finally,
. (B.5)
Appendix C: Proof of Proposition 7.3
First-order conditions for a maximum for the problem of maximizing (7.11) subject to (7.3)-(7.4) yield
,
,
,
,
,
and
. (C.1)
Solving (C.1) yields the Pareto efficient strategies and Lagrange multipliers:
,
,
,
,
. (C.2)
Substituting
and
into
and
in (C.2) yields:
and
. (C.3)