I am trying to wrap my mind around the variance definition.

Given a set of values S and n = #(S), the variance is defined as:

$$ \operatorname{var}(S) = \frac{\sum_{i=1}^n( S_i - \operatorname{mean}(S) )^2} n $$

And the square root of that (standard deviation) measures how far away the values are on average from the mean.

However, there is a simpler formula that also measures how far away the values are from the mean:

$$ \operatorname{another Possible Def For Var}(S) = \frac{\sum_{i=1}^n|S_i - \operatorname{mean}(S)|}{n} $$

I am trying to understand the reasoning behind the fact we use square root instead of a simpler modules function there. Is there a real reason why variance was defined the first way and not the second way?

# EDIT #

Ok, looks like the given reasons so far are way more advanced than what I was expecting.

The argument of squaring it as opposed to taking the modulus saying the modulus make the math more complicated is valid, but more of a consequence of the definition rather than a reason for it being defined as it is IMHO. Same thing goes for the Central Limit Theorem.

I ended up finding the exact same question at Khan Academy. There, the following reasons were also given:

  1. "Squaring emphasizes larger differences (think of the effect outliers have)." Another comment also points out: "In addition to amplifying large differences from the mean, squaring also MINIMIZES tiny differences from the mean".

These are the most convincing reasons I found so far. The modulus will not emphasize large values, neither will it minimize small values. HOWEVER, the same argument goes to any even power. A power of 4 will also amplify large differences and minimize tiny differences (it will actually do a better job at those). So why not take the power of 4 then? (or any other even number for that matter).

  1. "(...) you can also view the equation as being the Euclidean distance between all the points and the mean of the points"

That's more of a "nice-to-have" than a reason to me. If anything, the modules would give the Manhattan distance. So what?

Having said all that, I am not 100% convinced yet. I believe this question is way deeper than it looks at first glance and judging from the Khan Academy number of upvotes, I am not the only one confused about it.

marked as duplicate by kjetil b halvorsen, Community Apr 29 '17 at 9:08

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • 1
    No, it is the standard deviation which is the square root of the variance that is the unit of distance from the mean. What you are suggesting is called the mean absolute deviation. The standard deviation is used because the variance is a natural parameter for the normal distribution – Michael Chernick Apr 28 '17 at 0:01
  • 1
    Another (somewhat circular?) justification relates to this recent question: If you want to approximate the data by a constant value, and you choose the mean as that value, you are minimizing the variance. If you choose the median as the constant, you are minimizing the mean absolute deviation about the constant. – GeoMatt22 Apr 28 '17 at 0:27
  • 1
    @GeoMatt22 We could also talk about Chebyshev's inequality which involves the variance and applies to all distributions with finite second moments. – Michael Chernick Apr 28 '17 at 0:38
  • 1
    Why do you think the modulus function is easier than the square? This is not a trivial question, btw – Aksakal Apr 28 '17 at 1:37
  • 1
    Your alternative is more difficult in the sense of not being differentiable at the origin, whereas variance is infinitely differentiable everywhere. – Mark L. Stone Apr 28 '17 at 4:06

Let $\mu=\operatorname{E}(X).$

The main reason for using $\sqrt{\operatorname{var}(X)} = \sqrt{\operatorname{E}((X-\mu)^2)}$ as a measure of dispersion, rather that using the mean absolute deviation $\operatorname{E}(|X-\mu|),$ is that if $X_1,\ldots,X_n$ are independent, then $$ \operatorname{var}(X_1+\cdots+X_n) = \operatorname{var}(X_1)+\cdots+\operatorname{var}(X_n). \tag 1 $$ Nothing like that works with the mean absolute deviation. For example, try it with $X_1,X_2,X_3,\sim\operatorname{i.i.d.} \operatorname{Bernoulli}(1/2).$

In any problem where you use the central limit theorem, you need this.

For example: What is the standard deviation of the number of heads that appear when a coin is tossed $900$ times? That's easy to find because of $(1).$

  • Central limit theorem does not depend on the definition of variance. It is kind of weak converge of the sum of random variable. Even you do not define variance, the sum of random variable met the certain conditions still converges to normal distribution. – a_statistician Apr 28 '17 at 2:49
  • Agree with your general explanation (+1) but note that actually independence isn't even necessary as this is the more general formula: $\operatorname{var}\left(\sum_i X_i\right)=\sum_i\sum_j\operatorname{cov}\left(X_i,X_j\right)$. When all covariances between variables are 0, this reduces to your equation (1). And there are other simple rules to figure out how variances change under e.g. scaling operations (see en.wikipedia.org/wiki/Variance and en.wikipedia.org/wiki/Propagation_of_uncertainty for more info). – Ruben van Bergen Apr 28 '17 at 12:56
  • @a_statistician : How do you know WHICH normal distribution the sum of i.i.d. random variables approximates? The answer is that you know the variance of the sum because it's the sum of the variances. – Michael Hardy Apr 28 '17 at 22:19
  • The mean absolute deviation (E[|X−μ|]) + mean specify a unique normal distribution, same as variance + mean. Parameters of the distribution can be converted or transformed given the function is 1-1. To study the converge process, maybe the best tool is characteristic function. – a_statistician Apr 29 '17 at 2:17
  • (+1) I agree this is a useful reason! – GeoMatt22 Apr 29 '17 at 2:43

There are already several good answers here, including in the comments. However as the OP requested a "simpler" justification, here I will expand on my comment.

To me this is a very natural distinction between root-mean-square vs. mean-absolute deviations, and why we might prefer one vs. the other when measuring dispersion. (I do not know if it is "simpler"?)


Say you have some data $x_1,\ldots,x_n$, which you want to approximate by a constant $c$, i.e. $$x_i\approx c$$ for all $i$.

How do you choose the constant? A common approach is to minimize some error $E[c]$.

One choice for $E$ is the sum square error $$E_\text{SSE}=\sum_i\big(x_i-c\big)^2$$ the solution will then be $c_\min=\frac{1}{n}\sum x_i$. In other words, we have $$\big[c_\min,E_\min\big]_\text{SSE}=\big[\text{mean}(\mathbf{x}),n\,\text{var}(\mathbf{x})\big]$$ so if you are using the mean as your measure of central tendency, the RMS error is really the "natural" measure of dispersion.

On the other hand, if we choose $E$ to be the sum absolute error $$E_\text{SAE}=\sum_i\big|x_i-c\big|$$ the solution will then be $(c_\min)_\text{SAE}=\text{median}(\mathbf{x})$. So if you want to use mean absolute deviation to measure dispersion, really the "natural" measure of central tendency would be the median.


Summary: If you want to use mean absolute deviation, then arguably you should be measuring dispersion around the median. If you are already using the mean, then arguably standard deviation is the appropriate measure of dispersion. Here "arguably" is justified by optimality (minimum dispersion).

  • This is just a specialization of my answer here to the case of "regression with only an intercept". – GeoMatt22 Apr 29 '17 at 2:47

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