I would like to know if it is possible to calculate the expected value for a continuous random variable as:

$ \mathop{\mathbb{E}}(x)=\int xf(x)d(x)=\frac{1}{N}\sum_{i=1}^{N}X_{i} $

If that is true, how can we convert the integral to the sum? Also, is the sum part $\frac{1}{N}\sum_{i=1}^{N}X_{i}$ used to calculate the discrete uniform distribution?

  • $\begingroup$ Can you explain what is the random variable (or random variables)? What is $x$? What is $X_i$? What is $N$? If $X$ is a random variable with PDF $f_X(x)$ then indeed $E[X] = \int_{-\infty}^{\infty} xf_X(x)dx$. $\endgroup$
    – Michael
    Jul 19 '18 at 17:49
  • $\begingroup$ @Michael I have a series of different values that I want to calculate the expected value for each of them. An example of the series is 5.6, 3.2, 1.9, 6.1, 7.2. $N$ in this case is 5, $X_{i}$ is each value in the series. $\endgroup$
    – Adam Amin
    Jul 19 '18 at 17:56
  • $\begingroup$ @AdamAmin And where is the continuous distribution? $\endgroup$ Jul 19 '18 at 18:03
  • $\begingroup$ @callculus I am not sure. It is confusing me how to calculate the expected value for each value in the series. I need this because I am calculating the autocorrelation for the series I have and part of the equation is to get the expected value. $\endgroup$
    – Adam Amin
    Jul 19 '18 at 18:09
  • 1
    $\begingroup$ This is called monte carlo integration $\endgroup$
    – user3417
    Jul 19 '18 at 18:13

I think you are asking if it is possible to estimate the mean of a continuous distribution by taking the sample mean of a large sample (of size $N$) from that distribution. If so, the answer is Yes, but be aware that the result is an estimate and that the error of the estimate depends on the sample size $N$ (typically, smaller error for larger $N$). Here are some examples in which we sample at random from a few known distributions. (Sampling done using R statistical software.)

1) $N=50$ observations from the distributon $\mathsf{Norm}(\mu = 75,\, \sigma=3).$ The R statement rnorm samples at random from a specified normal population. You will get a somewhat different estimate each time you take a sample and find its mean. In 95% of runs the answer will be within

x = rnorm(50, 75, 3);  mean(x);  2*sd(x)/sqrt(50)
## 74.84001     # aprx E(X) = 75
## 0.7656806    # 95% margin of simulation error

2) $N = 5000$ observations from $\mathsf{Norm}(\mu = 75,\, \sigma=3).$

x = rnorm(5000, 75, 3);  mean(x);  2*sd(x)/sqrt(5000)
## 75.00454
## 0.08390667

3) $N = 1000$ observations from $\mathsf{Unif}(a = 50, b = 100).$ The mean of this population distribution is $\mu = 75.$

y = runif(1000, 50, 100);  mean(y);  2*sd(y)/sqrt(1000)
## 75.54673
## 0.900703

4) $N = 10^6$ observations from an exponential distribution with rate 1/5. This distribution has mean $\mu = 5.$

w = rexp(10^6, 1/5);  mean(w);  2*sd(w)/sqrt(10^6)
[1] 4.997817
[1] 0.00998948

5) $n = 10,000$ observations from the beta distribution with shape parameters $\alpha = 1.2$ and $\beta = 2.8.$ You can look in your text or read about the beta distribution on Wikipedia, if you have not encountered it yet. Its mean is $$\mu = \int_0^1 x\,\frac{\Gamma(\alpha + \beta)}{\Gamma(\alpha)+\Gamma(\beta)}\,x^{\alpha-1}(1-x)^{\beta-1}\,dx = \frac{\alpha}{\alpha+\beta} = 0.3.$$

v = rbeta(10000, 1.2, 2.8);  mean(v);  2*sd(v)/sqrt(10000)
[1] 0.3017991
[1] 0.004096549

Notes: (a) The Law of Large Numbers (LLN) guarantees that the mean of a sample converges 'in probability' to its population mean as $N \rightarrow \infty.$ By definition, that is $\lim_{N \rightarrow \infty}P(|\bar X - \mu| < \epsilon) = 0,$ for any $\epsilon > 0.$

(b) The margins of simulation error depend on the dispersion of the population and the sample size. For most non-normal distributions the same formula I have used in the examples above works because of the Central Limit Theorem (CLT).

(c) As in @Geronimo's Comment (+1), this is one form of 'Monte Carlo Integration.' It is a method often used in applications, but you don't literally "convert the integral into a sum." As remarked above you use the LLN and the CLT to get an approximate answer and to have some confidence how accurate that answer is. (Of course, it's only a 95% confidence interval, so there are some occasions on which the estimate is farther from the target than hoped for.)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.