Abstract
Weighted quasi-arithmetic means on two-dimensional regions are demonstrated, and risk averse conditions are discussed by the corresponding utility functions. For two utility functions on two-dimensional regions, we introduce a concept that decision making with one utility is more risk averse than decision making with the other utility. A necessary condition and a sufficient condition for the concept are demonstrated by their utility functions. Several examples are given to explain them.
Access provided by CONRICYT-eBooks. Download conference paper PDF
Similar content being viewed by others
1 Introduction
Weighted quasi-arithmetic means are important concept for mathematical theory such as the mean value theorems, and it is a fundamental tool for subjective estimation regarding information in management science, artificial intelligence and so on. Weighted quasi-arithmetic means of an interval are given mathematically by aggregation operations (Kolmogorov [4], Nagumo [6] and Aczél [1]). Bustince et al. [2] discussed aggregation operations on two-dimensional OWA operators, and Labreuche and Grabisch [5] demonstrated Choquet integral for aggregation in multicriteria decision making, and Torra and Godo [7] studied continuous WOWA operators for defuzzification. In micro-economics, subjective estimations with preference relations are formulated as utility functions (Fishburn [3]). From the view point of utility functions, Yoshida [8, 9] have studied the relations between weighted quasi-arithmetic means on an interval and decision maker’s behavior regarding risks. In one-dimensional cases, for twice continuously differentiable strictly increasing functions \(\varphi , \psi : [a,b] \mapsto {\mathbb {R}}\) as decision makers’ utility functions and a continuous function \(\omega : [a,b] \mapsto (0,\infty )\) as a weighting function, weighted quasi-arithmetic means \(\mu \) and \(\nu \) on a closed interval [a, b] are real numbers satisfying
in the mean value theorem for integration. Then it is said that decision making with utility function \(\varphi \) is more risk averse than decision making with utility function \(\psi \) if \(\mu \le \nu \) for all closed intervals [a, b]. Its equivalent condition is
on \({\mathbb {R}}\) (Yoshida [10, 11]).
Yoshida [12] introduced weighted quasi-arithmetic means on two-dimensional regions, which are related to multi-object decision making. In this paper, using decision makers’ utility functions we discuss relations between risk averse/risk neutral/risk loving conditions and the corresponding weighted quasi-arithmetic means on two-dimensional regions. In this paper we compare two decision makers’ behaviors regarding risks by the weighted quasi-arithmetic means on two-dimensional regions and we give a characterization by their utility functions.
In Sect. 2 we introduce weighted quasi-arithmetic means on two-dimensional regions and we discuss their risk averse conditions. For two utility functions f and g on two-dimensional regions, we introduce a concept that decision making with utility f is more risk averse than decision making with utility g. Further we derive a necessary condition where decision making with utility f is more risk averse than decision making with utility g on two-dimensional regions, and we investigate the condition by several examples. In Sect. 3 we give sufficient conditions for the results in Sect. 2 when utility functions are quadratic.
2 Weighted Quasi-arithmetic Means on Two-Dimensional Regions
Let \({\mathbb {R}}= (- \infty , \infty )\) and let a domain D be a non-empty open convex subset of \({\mathbb {R}}^2\), and let \(\mathcal{R}(D)\) be a family of closed convex subsets of D. Denote by \(\mathcal{L}\) a family of twice continuously differentiable functions \(f : D \mapsto {\mathbb {R}}\) which is strictly increasing, i.e. \(f_x>0\) and \(f_y>0\) on D, and denote by \(\mathcal{W}\) a family of continuous functions \(w : D \mapsto (0,\infty )\). For a closed convex set \(R \in \mathcal{R}(D)\), weighted quasi-arithmetic means on region R with utility \(f \in \mathcal{L}\) and weighting \(w \in \mathcal{W}\) are given by a subset \(M^{f}_{w}(R)\) of region R as follows.
Then we have \(M^{f}_{w}(R) \ne \emptyset \) since f is continuous on R and
We introduce the following natural ordering on \({\mathbb {R}}^2\).
Definition 2.1
(A partial order \(\preceq \) on \({\mathbb {R}}^2\)).
-
(i)
For two points \((\underline{x},\underline{y}), (\overline{x},\overline{y}) (\in {\mathbb {R}}^2)\), an order \((\underline{x},\underline{y}) \preceq (\overline{x},\overline{y})\) implies \(\underline{x} \le \overline{x} \ \text{ and } \ \underline{y} \le \overline{y}\).
-
(ii)
For two points \((\underline{x},\underline{y}), (\overline{x},\overline{y}) (\in {\mathbb {R}}^2)\), an order \((\underline{x},\underline{y}) \prec (\overline{x},\overline{y})\) implies \((\underline{x},\underline{y}) \preceq (\overline{x},\overline{y})\) and \((\underline{x},\underline{y}) \ne (\overline{x},\overline{y})\).
-
(iii)
For two sets \(A,B (\subset {\mathbb {R}}^2)\), an order \(A \preceq B\) implies the following (a) and (b):
-
(a)
For any \((\underline{x},\underline{y}) \in A\) there exists \((\overline{x},\overline{y}) \in B\) satisfying \((\underline{x},\underline{y}) \preceq (\overline{x},\overline{y})\).
-
(b)
For any \((\overline{x},\overline{y}) \in B\) there exists \((\underline{x},\underline{y}) \in A\) satisfying \((\underline{x},\underline{y}) \preceq (\overline{x},\overline{y})\).
-
(a)
Let a closed convex region \(R \in \mathcal{R}(D)\) and let a weighting function \(w \in \mathcal{W}\). We define a point \((\overline{x}_R,\overline{y}_R)\) on region R by the following weighted quasi-arithmetic means:
Hence, \((\overline{x}_R,\overline{y}_R)\) is called an invariant risk neutral point on R with weighting w (Yoshida [12]). We separate the space \({\mathbb {R}}^2\) as follows. Let \(R_{w,-}^{(\overline{x}_R,\overline{y}_R)} = \{ (x,y) \in {\mathbb {R}}^2 \mid (x,y) \prec (\overline{x}_R,\overline{y}_R) \} = \{ (x,y) \in {\mathbb {R}}^2 \mid x \le \overline{x}_R, \ y \le \overline{y}_R, (x,y) \ne (\overline{x}_R,\overline{y}_R) \}\) and \(R_{w,+}^{(\overline{x}_R,\overline{y}_R)} = \{ (x,y) \in {\mathbb {R}}^2 \mid (\overline{x}_R,\overline{y}_R) \prec (x,y) \} = \{ (x,y) \in {\mathbb {R}}^2 \mid x \ge \overline{x}_R, \ y \ge \overline{y}_R, (x,y) \ne (\overline{x}_R,\overline{y}_R) \}\). Then \(R_{w,-}^{(\overline{x}_R,\overline{y}_R)}\) denotes a subregion of risk averse points and \(R_{w,+}^{(\overline{x}_R,\overline{y}_R)}\) denotes a subregion of risk loving points. Let \(R_w^{(\overline{x}_R,\overline{y}_R)} = R_{w,-}^{(\overline{x}_R,\overline{y}_R)} \cup R_{w,+}^{(\overline{x}_R,\overline{y}_R)} \cup \{ (\overline{x}_R,\overline{y}_R) \}\). Now we introduce the following relations between decision maker’s behavior and his utility.
Definition 2.2
Let a utility function \(f \in \mathcal{L}\) and let a rectangle region \(R \in \mathcal{R}(D)\).
-
(i)
Decision making with utility f is called risk neutral on R if
$$\begin{aligned} f(\overline{x}_R,\overline{y}_R) \iint _{R} w(x,y) \, dx \, dy = \iint _{R} f(x,y) w(x,y) \, dx \, dy \end{aligned}$$(2.4)for all density functions w.
-
(ii)
Decision making with utility f is called risk averse on R if
$$\begin{aligned} f(\overline{x}_R,\overline{y}_R) \iint _{R} w(x,y) \, dx \, dy \ge \iint _{R} f(x,y) w(x,y) \, dx \, dy \end{aligned}$$(2.5)for all density functions w.
-
(iii)
Decision making with utility f is called risk loving on R if
$$\begin{aligned} f(\overline{x}_R,\overline{y}_R) \iint _{R} w(x,y) \, dx \, dy \le \iint _{R} f(x,y) w(x,y) \, dx \, dy \end{aligned}$$(2.6)for all density functions w.
Example 2.1
Let a domain \(D=(-0.5,1.25)^2\) and a region \(R=[0, 1]^2\), and let a weighting function \(w(x,y)=1\) for \((x,y) \in D\). Then an invariant neutral point is \((\overline{x}_R,\overline{y}_R)=(0.5,0.5)\) and \(R_{w,-}^{(\overline{x}_R,\overline{y}_R)} = [0,0.5]^2 \setminus \{ (0.5,0.5) \}\) and \(R_{w,+}^{(\overline{x}_R,\overline{y}_R)} = [0.5,1]^2 \setminus \{ (0.5,0.5) \}\). Let us consider two utility functions \(f(x,y)=-x^2-y^2+3x+3y\) and \(g(x,y)=2x^2+2y^2-5x-5y\) for \((x,y) \in D\). Then by Yoshida [12, Example 3.1(i), Lemma 2.2] decision making with utility function f is called risk averse on R with weighting w, and decision making with utility function g is also called risk loving on R with weighting w. Hence the corresponding weighted quasi-arithmetic means \(M^{f}_{w}(R)\) and \(M^{g}_{w}(R)\) are ordered by the order \(\preceq \) in a restricted subregion \(R_{w}^{(\overline{x}_R,\overline{y}_R)} = R_{w,-}^{(\overline{x}_R,\overline{y}_R)} \cup R_{w,+}^{(\overline{x}_R,\overline{y}_R)} \cup \{ (\overline{x}_R,\overline{y}_R) \}\). However they can not be ordered on a subregion \(R \setminus R_{w}^{(\overline{x}_R,\overline{y}_R)}\) (Fig. 1).
It is natural that the order \(\preceq \) should be given between weighted quasi-arithmetic means \(M^{f}_{w}(R)\) of risk averse utility f and weighted quasi-arithmetic means \(M^{g}_{w}(R)\) of risk loving utility g in Example 3.1. Therefore when we compare weighted quasi-arithmetic means \(M^{f}_{w}(R)\) and \(M^{g}_{v}(R)\), we discuss it on the meaningful restricted subregion \(R_w^{(\overline{x}_R,\overline{y}_R)}\). Hence we introduce the following definition regarding the comparison of utility functions.
Definition 2.3
Let \(f,g \in \mathcal{L}\) be utility functions on D. Decision making with utility f is more risk averse than decision making with utility g if it holds that
for all weighting functions \(w \in \mathcal{W}\) on D and all closed convex regions \(R \in \mathcal{R}(D)\).
Example 2.2
Let a domain \(D=(-0.5,1.25)^2\) and a region \(R=[0, 1]^2\), and let a weighting function \(w(x,y)=1\) for \((x,y) \in D\). Then an invariant neutral point is \((\overline{x}_R,\overline{y}_R)=(0.5,0.5)\) and \(R_{w,-}^{(\overline{x}_R,\overline{y}_R)} = [0,0.5]^2 \setminus \{ (0.5,0.5) \}\) and \(R_{w,+}^{(\overline{x}_R,\overline{y}_R)} = [0.5,1]^2 \setminus \{ (0.5,0.5) \}\). Let us consider two utility functions \(f(x,y)=-x^2-y^2+3x+3y\) and \(g(x,y)=-2x^2-2y^2+5x+5y\) for \((x,y) \in D\). Then decision making with utility f is more risk averse than decision making with utility g as we see the relation (2.7) in Fig. 2.
Now we give a necessary condition for (2.7), i.e. decision making with utility f is more risk averse than decision making with utility g.
Theorem 2.1
Let \(f,g \in \mathcal{L}\) be utility functions on D. If decision making with utility f is more risk averse than decision making with utility g, then it holds that
on D for all positive numbers h and k and all real numbers r satisfying \(-1 \le r \le 1\).
From Theorem 2.1 we can easily obtain the following result, which is corresponding to [12, Theorem 3.1(i)].
Corollary 2.1
Let \(f,g \in \mathcal{L}\) be utility functions on D. If decision making with utility f is more risk averse than decision making with utility g, then it holds that
Equation (2.8) in Theorem 2.1 gives a detailed relation between f and g rather than (2.9). A parameter r in necessary condition (2.8) depends on the shapes of closed convex regions \(R \in \mathcal{R}(D)\). Now we investigate several examples with different shapes of regions R.
Example 2.3
(Rectangle regions). Let h and k be positive numbers.Let rectangle regions
for \((a,b) \in D\) and \(t>0\). Denote a family of rectangle regions by \(\mathcal{R}^\mathrm{Rect}_{h,k}(D) = \{ R^\mathrm{Rect}_{h,k}(a,b,t) \mid R^\mathrm{Rect}_{h,k}(a,b,t) \subset D, \, (a,b) \in D, \ t>0 \} (\subset \mathcal{R}(D))\), (Fig. 3).
Corollary 2.2
If utility functions \(f,g \in \mathcal{L}\) satisfy \(M^{f}_{w}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)} \preceq M^{g}_{v}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)}\) for all weighting functions \(w \in \mathcal{W}\) on D and all rectangle regions \(R \in \mathcal{R}^\mathrm{Rect}_{h,k}(D)\), then it holds that
on D.
Example 2.4
(Oval regions). Let h and k be positive numbers. Let oval regions
for \((a,b) \in D\) and \(t>0\). Denote a family of oval regions by \(\mathcal{R}^\mathrm{Oval}_{h,k}(D) = \{ R^\mathrm{Oval}_{h,k}(a,b,t) \mid R^\mathrm{Oval}_{h,k}(a,b,t) \subset D, \, (a,b) \in D, \ t>0 \} (\subset \mathcal{R}(D))\), (Fig. 3).
Corollary 2.3
If utility functions \(f,g \in \mathcal{L}\) satisfy \(M^{f}_{w}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)} \preceq M^{g}_{v}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)}\) for all weighting functions \(w \in \mathcal{W}\) on D and all oval regions \(R \in \mathcal{R}^\mathrm{Oval}_{h,k}(D)\), then it holds that
on D.
Example 2.5
(Triangle regions). Let h and k be positive numbers. Let triangle regions
for \((a,b) \in D\) and \(t>0\). Denote a family of triangle regions by \(\mathcal{R}^\mathrm{Tri}_{h,k}(D) = \{ R^\mathrm{Tri}_{h,k}(a,b,t) \mid R^\mathrm{Tri}_{h,k}(a,b,t) \subset D, \, (a,b) \in D, \ t>0 \} (\subset \mathcal{R}(D))\), (Fig. 4).
Corollary 2.4
If utility functions \(f,g \in \mathcal{L}\) satisfy \( M^{f}_{w}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)} \preceq M^{g}_{v}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)} \) for all weighting functions \(w \in \mathcal{W}\) on D and all triangle regions \(R \in \mathcal{R}^\mathrm{Tri}_{h,k}(D)\), then it holds that
on D.
Example 2.6
(Parallelogram regions). Let h and k be positive numbers. Let parallelogram regions
for \((a,b) \in D\) and \(t>0\). Denote a family of parallelogram regions by \(\mathcal{R}^\mathrm{Para}_{h,k}(D) = \{ R^\mathrm{Para}_{h,k}(a,b,t) \mid R^\mathrm{Para}_{h,k}(a,b,t) \subset D, \, (a,b) \in D, \ t>0 \} (\subset \mathcal{R}(D))\), (Fig. 4).
Corollary 2.5
If utility functions \(f,g \in \mathcal{L}\) satisfy \(M^{f}_{w}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)} \preceq M^{g}_{v}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)}\) for all weighting functions \(w \in \mathcal{W}\) on D and all parallelogram regions \(R \in \mathcal{R}^\mathrm{Para}_{h,k}(D)\), then it holds that
on D.
Example 2.3 (Rectangle regions) and Example 2.4 (Oval regions) are cases where \(r=0\) in (2.8), and Example 2.5 (Triangle regions) and Example 2.6 (Parallelogram regions) are cases where \(r=-\frac{1}{2}\) and \(r=\frac{3}{10}\) respectively in (2.8).
3 A Sufficient Condition
Let \(f,g \in \mathcal{L}\) be utility functions on an open convex domain D. Theorem 2.1 gives a necessary condition that decision making with utility f is more risk averse than decision making with utility g. In this section, we discuss its sufficient condition. For a utility function \(f \in \mathcal{L}\), its Hessian matrix is written by
for \((x,y) \in D\). The the following proposition gives a sufficient condition for (2.8) in Theorem 2.1.
Proposition 3.1
Let \(f,g \in \mathcal{L}\) be utility functions on D. Then the following (i) and (ii) hold.
-
(i)
Matrices
$$\begin{aligned} \frac{1}{f_x(x,y)} H^f(x,y) {-} \frac{1}{g_x(x,y)} H^g(x,y) \ {and} \ \frac{1}{f_y(x,y)} H^f(x,y) - \frac{1}{g_y(x,y)} H^g(x,y) \end{aligned}$$(3.2)are negative semi-definite for all \((x,y) \in D\) if and only if a matrix
$$\begin{aligned} \frac{1}{h f_x(x,y) + k f_y(x,y)} H^f(x,y) - \frac{1}{h g_x(x,y) + k g_y(x,y)} H^g(x,y) \end{aligned}$$(3.3)is negative semi-definite for all \((x,y) \in D\) and all positive numbers h and k.
-
(ii)
If (3.2) are negative semi-definite at all \((x,y) \in D\), then (2.8) holds on D for all positive numbers h and k and all real numbers r satisfying \(-1 \le r \le 1\).
From Proposition 3.1 implies that the condition (3.2) is stronger than the condition (2.8), however (3.2) is easier than (2.8) to check in actual cases. In this paper, utility functions \(f (\in \mathcal{L})\) are called quadratic if the second derivatives \(f_{xx}\), \(f_{xy}\) and \(f_{yy}\) are constant functions. When utility functions are quadratic, the following theorem gives a sufficient condition for what decision making with utility f is more risk averse than decision making with utility g.
Theorem 3.1
Let utility functions \(f, g \in \mathcal{L}\) be quadratic on D. If
are negative semi-definite at all \((x,y) \in D\), then decision making with utility f is more risk averse than decision making with utility g, i.e.
for all weighting functions \(w \in \mathcal{W}\) and all closed convex regions \(R \in \mathcal{R}(D)\).
Now we give an example for Theorem 3.1.
Example 3.1
(Quadratic utility functions). Let a domain \(D=(-0.5,1.5)^2\) and a region \(R= [0, 1]^2\), and let a weighting function \(w(x,y)=1\) for \((x,y) \in D\). Then an invariant neutral point is \((\overline{x}_R,\overline{y}_R)=(0.5,0.5)\) and \(R_{w,-}^{(\overline{x}_R,\overline{y}_R)} = [0,0.5]^2 \setminus \{ (0.5,0.5) \}\) and \(R_{w,+}^{(\overline{x}_R,\overline{y}_R)} = [0.5,1]^2 \setminus \{ (0.5,0.5) \}\). Let us consider two quadratic utility functions \(f(x,y)=-2x^2-2y^2+2xy+8x+8y\) and \(g(x,y)=-x^2-y^2+xy+5x+5y\) for \((x,y) \in D\). Then f and g are increasing on D, i.e. \(f_x(x,y)=-4x+2y+8>0\), \(f_y(x,y)=2x-4y+8>0\), \(g_x(x,y)=-2x+y+5>0\) and \(g_y(x,y)=x-2y+5>0\) on D. Their Hessian matrices are
Let A(x, y) and B(x, y) by \(A(x,y)=\frac{1}{f_x(x,y)} H^f(x,y) - \frac{1}{g_x(x,y)} H^g(x,y)\) and \(B(x,y)=\frac{1}{f_y(x,y)} H^f(x,y) - \frac{1}{g_y(x,y)} H^g(x,y)\) for \((x,y) \in D\), and then we have
We can easily check A(x, y) and B(x, y) are negative definite for all \((x,y) \in D\). From Theorem 3.1, decision making with utility f is more risk averse than decision making with utility g on R and it holds that \(M^{f}_{w}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)} \preceq M^{g}_{v}(R) \cap R_w^{(\overline{x}_R,\overline{y}_R)}\) for all weighting functions \(w \in \mathcal{W}\) (Fig. 5).
Concluding Remark. When utility functions are quadratic, Theorem 3.1 gives a sufficient condition where decision making with utility f is more risk averse than decision making with utility g. It is an open problem whether (3.2) is a sufficient condition when utility functions are not quadratic but more general.
References
Aczél, J.: On weighted mean values. Bulletin of the American Math. Society 54, 392–400 (1948)
Bustince, H., Calvo, T., Baets, B., Fodor, J., Mesiar, R., Montero, J., Paternain, D., Pradera, A.: A class of aggregation functions encompassing two-dimensional OWA operators. Inf. Sci. 180, 1977–1989 (2010)
Fishburn, P.C.: Utility Theory for Decision Making. Wiley, New York (1970)
Kolmogoroff, A.N.: Sur la notion de la moyenne. Acad. Naz. Lincei Mem. Cl. Sci. Fis. Mat. Natur. Sez. 12, 388–391 (1930)
Labreuche, C., Grabisch, M.: The Choquet integral for the aggregation of interval scales in multicriteria decision making. Fuzzy Sets Syst. 137, 11–26 (2003)
Nagumo, K.: Über eine Klasse der Mittelwerte. Japan. J. Math. 6, 71–79 (1930)
Torra, V., Godo, L.: On defuzzification with continuous WOWA operators. In: Aggregation Operators, pp. 159–176. Springer, New York (2002)
Yoshida, Y.: Aggregated mean ratios of an interval induced from aggregation operations. In: Torra, V., Narukawa, Y. (eds.) MDAI 2008. LNCS, vol. 5285, pp. 26–37. Springer, Heidelberg (2008). doi:10.1007/978-3-540-88269-5_4
Yoshida, Y.: Quasi-arithmetic means and ratios of an interval induced from weighted aggregation operations. Soft Comput. 14, 473–485 (2010)
Yoshida, Y.: Weighted quasi-arithmetic means and conditional expectations. In: Torra, V., Narukawa, Y., Daumas, M. (eds.) MDAI 2010. LNCS, vol. 6408, pp. 31–42. Springer, Heidelberg (2010). doi:10.1007/978-3-642-16292-3_6
Yoshida, Y.: Weighted quasi-arithmetic means and a risk index for stochastic environments. Int. J. Uncertainty Fuzziness Knowl. Based Syst. 16(suppl), 1–16 (2011)
Yoshida, Y.: Weighted quasi-arithmetic means on two-dimensional regions and their applications. In: Torra, V., Narukawa, Y. (eds.) MDAI 2015. LNCS, vol. 9321, pp. 42–53. Springer, Cham (2015). doi:10.1007/978-3-319-23240-9_4
Acknowledgments
This research is supported from JSPS KAKENHI Grant Number JP 16K05282.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2017 Springer International Publishing AG
About this paper
Cite this paper
Yoshida, Y. (2017). Comparison of Risk Averse Utility Functions on Two-Dimensional Regions. In: Torra, V., Narukawa, Y., Honda, A., Inoue, S. (eds) Modeling Decisions for Artificial Intelligence. MDAI 2017. Lecture Notes in Computer Science(), vol 10571. Springer, Cham. https://doi.org/10.1007/978-3-319-67422-3_2
Download citation
DOI: https://doi.org/10.1007/978-3-319-67422-3_2
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-67421-6
Online ISBN: 978-3-319-67422-3
eBook Packages: Computer ScienceComputer Science (R0)