Abstract
In Bayesian rule an unknown parameter is thought to be a quantity whose variation can be characterized by a prior distribution. Then some data are observed from a population whose distribution function is indexed by the unknown parameter and then the prior distribution is updated according to the observed data. The updated prior distribution is named as the posterior distribution. Based on uncertainty theory, this paper first makes a connection between posterior uncertainty distribution and likelihood function, and proposes a new method to obtain the posterior uncertainty distribution from the prior uncertainty distribution with given observed data. Some examples with special uncertainty distributions are employed to explain the calculation. Furthermore, an uncertain urn problem is provided to illustrate the application of the new method.
Similar content being viewed by others
1 Introduction
The fundamental concepts in uncertainty theory are uncertain measure (Liu, 2007) and product uncertain measure (Liu, 2009). Based on the measure, another fundamental concept, uncertain variable (Liu, 2007), was proposed to represent the uncertain quantities. To characterize the uncertain variable, the concept of uncertainty distribution was presented by Liu (2007). Following that, the sufficient and necessary condition for a real-valued function being an uncertainty distribution was proved by Peng and Iwamura (2010) and revised by Liu and Lio (2020).
To obtain an uncertainty distribution based on expert’s experimental data, Liu (2010) commenced the discussion of uncertain statistics by designing a questionnaire survey for the domain expert, and Chen and Ralescu (2012) showed that the questionnaire survey works. Moreover, Liu (2010) proposed the principle of least squares to estimate the unknown parameters in an uncertainty distribution with a known functional form, while Wang and Peng (2014) suggested the method of moments. For the case with multiple domain experts, Delphi method was revised by Wang et al. (2012) to determine the uncertainty distributions according to the experts’ experimental data.
Essentially, uncertain statistics is a set of mathematical tools to collect, analyze and interpret data based on uncertainty theory, and the techniques can also be applied to dealing with the observed data. In order to determine whether or not a statistical hypothesis is appropriate via observed data, uncertain hypothesis test was commenced by Ye and Liu (2021, 2022). Another main topic of uncertain statistics is uncertain regression analysis, which is a series of techniques that apply uncertainty theory to exploring the relationship between the response variables and explanatory variables. The discussion of uncertain regression was initialized by Yao and Liu (2018), and Lio and Liu (2018) further suggested an approach to estimate the uncertainty distribution of the disturbance term in an uncertain regression model. For predicting future values according to the previously observed data, uncertain time series was first studied by Yang and Liu (2019). Furthermore, many methods were proposed to make a connection between observed data and uncertain differential equation, for example, method of moments (Liu & Liu, 2021; Yao & Liu, 2020), minimum cover estimation (Yang et al., 2020), least squares estimation (Sheng et al., 2020), generalized moment estimation (Liu, 2021), and maximum likelihood estimation (Liu & Liu, 2022).
Bayesian rule is a usual method for connecting observed data and the distribution function by updating the prior distribution to the posterior distribution. It is known that the estimated distribution function got in most practical case is not close enough to the frequency in the world of eternal change and should be treated as an uncertainty distribution. How could a prior uncertainty distribution become a posterior probability distribution? That means, the classical method is not available when the prior distribution is regarded as an uncertainty distribution. Therefore, this paper will propose a new method to obtain the posterior uncertainty distribution from the prior uncertainty distribution with given observed data. The rest of the paper is organized as follows: Sect. 2 is a preliminary about some properties for likelihood function in the sense of uncertainty theory. Section 3 presents the method for updating prior uncertainty distribution to the posterior uncertainty distribution based on likelihood function and observed data. From Sects. 4 to 7, the theorems and examples for calculating the posterior uncertainty distributions from some special uncertainty distributions are provided. Section 8 introduces an application for the posterior uncertainty distribution to the uncertain urn problem and compares it to the posterior probability distribution. Finally, a concise conclusion is made in Sect. 9.
2 Preliminary
To define the posterior uncertainty distribution, the preliminary about the likelihood function is first presented in this section.
Theorem 1
[Lio and Liu (2020), Likelihood Function] Suppose \(\eta _{1},\eta _{2},\ldots ,\eta _{n}\) are iid uncertain variables with uncertainty distribution \(F(y\left| \theta \right. )\) where \(\theta \) is an unknown parameter, and have observed values \(y_{1},y_{2},\ldots ,y_{n}\), respectively. If \(F(y\left| \theta \right. )\) is differentiable at \(y_{1},y_{2},\ldots ,y_{n}\), then the likelihood function associated with \(y_{1},y_{2},\ldots ,y_{n}\) is
Example 1
Suppose \(\eta _{1},\eta _{2},\ldots ,\eta _{n}\) are iid uncertain variables which are normal uncertain variables \(\text{ N }(e,\sigma )\) with unknown parameters e and \(\sigma \), and have observed values \(y_{1},y_{2},\ldots ,y_{n}\), respectively. Then the likelihood function associated with \(y_{1},y_{2},\ldots ,y_{n}\) is
Example 2
(Linear Likelihood) Suppose \(\eta _{1},\eta _{2},\ldots ,\eta _{n}\) are iid uncertain variables which are linear uncertain variables \(L (a,b)\) with unknown parameters a and b, and have observed values \(y_{1},y_{2},\ldots ,y_{n}\), respectively. Then the likelihood function associated with \(y_{1},y_{2},\ldots ,y_{n}\) is
3 Likelihood function and posterior uncertainty function
A prior uncertainty distribution \(\Phi (x)\), which can be regarded as the uncertainty distribution of an uncertain variable \(\xi \), is available. Some uncertain variables \(\eta _1,\eta _2,\ldots ,\eta _n\) with uncertainty distribution \(F(y | \xi )\) are obtained, and they have observed values \(y_1,y_2,\ldots ,y_n\), respectively. What is the posterior uncertainty distribution \(\Psi (x|y_1,y_2,\ldots ,y_n)?\)
To answer this question, the likelihood function proposed by Lio and Liu (2020) should be applied. Essentially, the likelihood function represents the likelihood of taking value at x. Note that the derivative of a posterior uncertainty function can also represent the likelihood. Moreover, since the posterior uncertainty function is updated from the prior uncertainty distribution, in the sense of uncertainty theory, it is natural to regard that the derivative of the posterior uncertainty distribution should be in proportion to the minimum of likelihood function and the derivative of prior uncertainty distribution, i.e.,
Unfortunately, if an uncertainty distribution is defined by
then for any \(x\in \Re \), the distribution function is always strictly smaller than 1. It seems that we have no alternative but to divide Eq. (5) by
Hence the definition of posterior uncertainty function is given as follows:
Definition 1
Suppose \(\xi \) is an uncertain variable with prior uncertainty distribution \(\Phi (x)\), and \(\eta _1,\eta _2,\ldots ,\eta _n\) are iid uncertain variables from a population with uncertainty distribution \(F(y | \xi )\). Suppose \(\Phi '(x)\) and \(F'(y | \xi )\) can be obtained, and \(\eta _1,\eta _2,\ldots ,\eta _n\) have observed values \(y_1,y_2,\ldots ,y_n\), respectively. Then the posterior uncertainty distribution is defined by
It is clear that if
then the posterior uncertainty distribution defined by Eq. (6) is a continuous monotone increasing function satisfying
It was proved by Peng and Iwamura (2010) and Liu and Lio (2020) that Eq. (6) is indeed an uncertainty distribution.
4 Linear prior distribution and linear population
A definition was provided by Liu (2010) that a linear uncertain variable has an uncertainty distribution
denoted by \(L (a,b)\) where a and b are real numbers with \(a<b\). This section illustrates the calculation for the posterior uncertainty distribution when the prior distribution is assumed to be linear and the data are observed from a population whose distribution function is also linear.
Theorem 2
Suppose \(\xi \) is an uncertain variable with linear prior uncertainty distribution \(L (a,b)\), and \(\eta _1,\eta _2,\cdots ,\eta _n\) are iid uncertain variables from a population with linear uncertainty distribution \(L (\xi -c,\xi +d), \ c,d \ge 0\) and observed values \(y_1,y_2,\ldots ,y_n\), respectively. If it is assumed that
then the posterior uncertainty distribution is
Proof
It follows from Example 2, Definition 1 and
that the posterior uncertainty distribution satisfies
Then we have
That means, the posterior uncertainty distribution is
and the theorem is proved. \(\square \)
Example 3
Suppose \(\xi \) is an uncertain variable with linear prior uncertainty distribution \(L (171,175)\), and \(\eta _1,\eta _2,\eta _3\) are iid uncertain variables from a population with linear uncertainty distribution \(L (\xi -1,\xi +2)\) and observed values \(y_1=173, y_2=172, y_3=174\), respectively. Then it follows from Theorem 2 that the posterior uncertainty distribution is \(L (172,173)\), i.e.,
5 Normal prior distribution and linear population
A definition was provided by Liu (2010) that a normal uncertain variable has an uncertainty distribution
denoted by \(\text{ N }(e,\sigma )\) where e and \(\sigma \) are real numbers with \(\sigma >0\). This section illustrates the calculation for the posterior uncertainty distribution when the prior distribution is assumed to be normal and the data are observed from a population whose distribution function is linear.
Theorem 3
Suppose \(\xi \) is an uncertain variable with normal prior uncertainty distribution \(\text{ N }(e,\sigma )\), and \(\eta _1,\eta _2,\ldots ,\eta _n\) are iid uncertain variables from a population with linear uncertainty distribution \(L (\xi -c,\xi +d), \ c,d\ge 0\) and observed values \(y_1,y_2,\ldots ,y_n\), respectively. If it is assumed that
then the posterior uncertainty distribution is
Proof
It follows from Example 2, Definition 1 and
that the posterior uncertainty distribution satisfies
Since
the theorem is verified. \(\square \)
Example 4
Suppose \(\xi \) is an uncertain variable with normal prior uncertainty distribution \(\text{ N }(174,1)\), and \(\eta _1,\eta _2\) are iid uncertain variables from a population with linear uncertainty distribution \(L (\xi -2,\xi +1)\) and observed values \(y_1=174, y_2=173\), respectively. Then it follows from Theorem 3 that the posterior uncertainty distribution is
In fact, by taking
the equation
can be simplified as
Hence we can suppose \(x^*\) is the analytic solution of Eq. (9) satisfying \(x^* \le 174\), i.e., \(x^* \approx 173.3725\). Then for any \(173 < x \le 175\), it can be further obtained that
where \(\Phi \) is the uncertainty distribution \(\text{ N }(174,1)\). Moreover, for any \(173 < x \le 175\) (otherwise, \(\Psi ' (x | \ y_1,y_2)=0\)), we have
where \(\Phi \) is the uncertainty distribution \(\text{ N }(174,1)\). The posterior uncertainty distribution \(\Psi \) and its derivative \(\Psi '\) are revealed in Figs. 1 and 2, respectively.
6 Linear prior distribution and normal population
This section illustrates the calculation for the posterior uncertainty distribution when the prior distribution is assumed to be linear and the data are observed from a population whose distribution function is normal.
Theorem 4
Suppose \(\xi \) is an uncertain variable with linear prior uncertainty distribution \(L (a,b)\), and \(\eta _1,\eta _2,\ldots ,\eta _n\) are iid uncertain variables from a population with normal uncertainty distribution \(\text{ N }(\xi ,\sigma )\) and observed values \(y_1,y_2,\ldots ,y_n\), respectively. Then the posterior uncertainty distribution is
Proof
It follows from Definition 1 and
that the posterior uncertainty distribution is
The theorem follows from Example 1 immediately. \(\square \)
Example 5
Suppose \(\xi \) is an uncertain variable with linear prior uncertainty distribution \(L (171,178)\), and \(\eta _1,\eta _2\) are iid uncertain variables from a population with normal uncertainty distribution \(\text{ N }(\xi ,1)\) and observed values \(y_1=173, y_2=175\), respectively. Then it follows from Theorem 4 that the posterior uncertainty distribution is
Since the function
is symmetric about \(s=(173+175)/2=174\), reaches its maximum at \(s=174\), and decreases as s goes away from \(s=174\), it can be supposed that \(x^*\) is the analytic solution of equation
satisfying \(x^* \le 174\), i.e., \(x^*\approx 173.6983\). Then for \(171 < x \le 178\), we have
where \(\Phi _{1}\) and \(\Phi _{2}\) are the uncertainty distributions \(\text{ N }(173,1)\) and \(\text{ N }(175,1)\), respectively. Moreover, for \(171 < x \le 178\) (otherwise, \(\Psi ' (x | \ y_1,y_2)=0\)), it can be obtained that
where \(\Phi _{1}\) and \(\Phi _{2}\) are the uncertainty distributions \(\text{ N }(173,1)\) and \(\text{ N }(175,1)\), respectively. The posterior uncertainty distribution \(\Psi \) and its derivative \(\Psi '\) are revealed in Figs. 3 and 4, respectively.
7 Normal prior distribution and normal population
This section illustrates the calculation for the posterior uncertainty distribution when the prior distribution is assumed to be normal and the data are observed from a population whose distribution function is also normal.
Theorem 5
Suppose \(\xi \) is an uncertain variable with normal prior uncertainty distribution \(\text{ N }(e,\sigma )\), and \(\eta _1,\eta _2,\ldots ,\eta _n\) are iid uncertain variables from a population with normal uncertainty distribution \(\text{ N }(\xi ,\sigma )\) and observed values \(y_1,y_2,\ldots ,y_n\), respectively. Then the posterior uncertainty distribution is
where \(\Phi _M\) and \(\Phi _m\) are the uncertainty distributions \(\text{ N }(M,\sigma )\) and \(\text{ N }(m,\sigma )\), respectively, and
Proof
Since the function
reaches its maximum at \(z=0\), decreases as z goes away from 0 and is symmetric about 0, it follows from Example 1, Definition 1 and
that the posterior uncertainty distribution is
For any \(s \le (m+M)/2\) and \(s>(m+M)/2\), it is clear that
and
respectively. That means
Similarly, it is clear that
Then the theorem follows immediately. \(\square \)
Example 6
Suppose \(\xi \) is an uncertain variable with normal prior uncertainty distribution \(\text{ N }(174,1)\), and \(\eta _1,\eta _2\) are iid uncertain variables from a population with normal uncertainty distribution \(\text{ N }(\xi ,1)\) and observed values \(y_1=171, y_2=175\), respectively. Then it follows from Theorem 5 that the posterior uncertainty distribution is
where \(\Phi _{1}\) and \(\Phi _{2}\) are the uncertainty distributions \(\text{ N }(171,1)\) and \(\text{ N }(175,1)\), respectively. It can be further obtained that
where \(\Phi _{1}\) and \(\Phi _{2}\) are the uncertainty distributions \(\text{ N }(171,1)\) and \(\text{ N }(175,1)\), respectively. The posterior uncertainty distribution \(\Psi \) and its derivative \(\Psi '\) are revealed in Figs. 5 and 6, respectively.
8 Application to uncertain urn problem
Assume Jason has filled an uncertain urn with 5 balls which are either red or black, and the distribution function is completely unknown. In this case, Laplace criterion makes you assign equal likelihood to the possible number of the red balls in the uncertain urn. That means, the number of red balls should be represented by
Next, one ball is drawn from the urn, and the color of the drawn ball can be represented by
According to the drawn ball’s color, the likelihood function is revealed in the following Table 1.
By probability theory, if the drawn ball is red, then the posterior probability distribution is
if the drawn ball is black, then the posterior probability distribution is
By uncertainty theory, if the drawn ball is red, then the posterior uncertainty distribution is
if the drawn ball is black, then the posterior uncertainty distribution is
Obviously, the posterior probability distribution and posterior uncertainty distribution are different. Which result is more reasonable? To answer this question, let us consider three options as follows:
A: You receive $5 if the total number of red balls is 5 and lose $2 otherwise;
B: You receive $5 if the total number of black balls is 5 and lose $2 otherwise;
C. Don’t bet.
What should be the best choice among A, B and C? If the drawn ball is red and probability theory is used, then the expected incomes of A and B are
and
respectively. Since the income of C is always 0, probability theory tells you choose A in this case; if the drawn ball is black and probability theory is used, then the expected incomes of A and B are
and
respectively. Hence probability theory tells you choose B. If the drawn ball is red and uncertainty theory is used, then the expected incomes of A and B are
and
respectively. Since the income of C is always 0, uncertainty theory tells you choose C in this case; if the drawn ball is black and uncertainty theory is used, then the expected incomes of A and B are
and
respectively. Again, uncertainty theory tells you choose C.
To compare the results, now let us see how Jason filled the uncertain urn. First Jason chose a distribution function,
which is essentially the constant 3. Next a number k is got from the distribution function \(\Upsilon \) and filled k red balls and \(5-k\) black balls to the urn. Then there are 3 red balls and 2 black balls in the urn. No matter what color the drawn ball is, you would lose if probability theory is applied (you would win if and only if the number of the red balls is 0 or 5). Could you believe that the posterior distribution obtained by uncertainty theory is better? The core reason for the failure of probability theory in this case is that our uniform prior function is not close enough to the real frequency of the number of red balls (i.e., the distribution function chosen by Jason). In this case, the prior distribution should be regarded as uncertainty distribution and posterior uncertainty distribution can be obtained based on the method proposed in this paper.
9 Conclusion
This paper made a connection between posterior uncertainty distribution as well as likelihood function and suggested a new method to get the posterior uncertainty distribution from the prior uncertainty distribution based on given observed data. Some examples with special uncertainty distributions and an uncertain urn problem were provided to illustrate the application of the method.
References
Chen, X., & Ralescu, D. A. (2012). B-spline method of uncertain statistics with applications to estimate travel distance. Journal of Uncertain Systems, 6(4), 256–262.
Lio, W., & Liu, B. (2018). Residual and confidence interval for uncertain regression model with imprecise observations. Journal of Intelligent & Fuzzy Systems, 35(2), 2573–2583.
Lio, W., & Liu, B. (2020). Uncertain maximum likelihood estimation with application to uncertain regression analysis. Soft Computing, 24(13), 9351–9360.
Liu, B. (2007). Uncertainty theory (2nd ed.). Berlin: Springer.
Liu, B. (2008). Fuzzy process, hybrid process and uncertain process. Journal of Uncertain Systems, 2(1), 3–16.
Liu, B. (2009). Some research problems in uncertainty theory. Journal of Uncertain Systems, 3(1), 3–10.
Liu, B. (2010). Uncertainty theory: A branch of mathematics for modeling human uncertainty. Springer.
Liu, Z. (2021). Generalized moment estimation for uncertain differential equations. Applied Mathematics and Computation, 392, 125724.
Liu, Y., & Lio, W. (2020). A revision of sufficient and necessary condition of uncertainty distribution. Journal of Intelligent & Fuzzy Systems, 38(4), 4845–4854.
Liu, Y., & Liu, B. (2021). Residual analysis and parameter estimation of uncertain differential equations. Fuzzy Optimization and Decision Making. https://doi.org/10.1007/s10700-021-09379-4.
Liu, Y., & Liu, B. (2022). Estimating unknown parameters in uncertain differential equation by maximum likelihood estimation. Soft Computing. https://doi.org/10.1007/s00500-022-06766-w.
Peng, Z., & Iwamura, K. (2010). A sufficient and necessary condition of uncertainty distribution. Journal of Interdisciplinary Mathematics, 13(3), 277–285.
Sheng, Y., Yao, K., & Chen, X. (2020). Least squares estimation in uncertain differential equations. IEEE Transactions on Fuzzy Systems, 28(10), 2651–2655.
Wang, X., Gao, Z., & Guo, H. (2012). Delphi method for estimating uncertainty distributions. Information: An International Interdisciplinary Journal, 15(2), 449–460.
Wang, X., & Peng, Z. (2014). Method of moments for estimating uncertainty distributions. Journal of Uncertainty Analysis and Applications, 2, 5.
Yang, X., & Liu, B. (2019). Uncertain time series analysis with imprecise observations. Fuzzy Optimization and Decision Making, 18(3), 263–278.
Yang, X., Liu, Y. H., & Park, G. K. (2020). Parameter estimation of uncertain differential equation with application to financial market. Chaos, Solitons and Fractals, 139, 110026.
Yao, K., & Liu, B. (2018). Uncertain regression analysis: An approach for imprecise observations. Soft Computing, 22(17), 5579–5582.
Yao, K., & Liu, B. (2020). Parameter estimation in uncertain differential equations. Fuzzy Optimization and Decision Making, 19(1), 1–12.
Ye, T., & Liu, B. (2021). Uncertain hypothesis test with application to uncertain regression analysis. Fuzzy Optimization and Decision Making. https://doi.org/10.1007/s10700-021-09365-w.
Ye, T., & Liu, B. (2022). Uncertain hypothesis test for uncertain differential equations. Fuzzy Optimization and Decision Making. https://doi.org/10.1007/s10700-022-09389-w.
Acknowledgements
This work was supported by National Natural Science Foundation of China Grant Nos. 62073009 and 61873329.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Lio, W., Kang, R. Bayesian rule in the framework of uncertainty theory. Fuzzy Optim Decis Making 22, 337–358 (2023). https://doi.org/10.1007/s10700-022-09395-y
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10700-022-09395-y