bonsoon's blog |
| latest | about | random
# Where did these classical means come from? For two positive numbers $x,y > 0$, we are familiar with some ways to calculate their "means", for instance their **arithmetic mean** $$ a(x,y) = \frac{1}{2}(x+y), $$their **geometric mean** $$ g(x,y) = \sqrt{xy}, $$and their **harmonic mean** $$ h(x,y) = \frac{2}{\frac{1}{x}+\frac{1}{y}}. $$But how were these dreamt up, and are there others? These means arise from antiquity Greeks, and to capture this notion of mean of two positive numbers, they looked at **ratios**. Take $0 < x < y$, and say $m$ is some kind of mean for both numbers, satisfying $x < m < y$. Then we can consider the ratios of the two segments $y-m$ to $m-x$, and set them equal to the nine possible ratios you can make with $x,y,m$. That is, we can assign $$ \frac{y-m}{m-x} $$ to one of the nine ratios $$\frac{y-m}{m-x}= \,\,\, \begin{array}{}\displaystyle \frac{x}{x}, & \displaystyle\frac{x}{m}, & \displaystyle\frac{x}{y}, \\ \displaystyle\frac{m}{x}, & \displaystyle\frac{m}{m}, & \displaystyle\frac{m}{y}, \\ \displaystyle\frac{y}{x}, & \displaystyle\frac{y}{m}, & \displaystyle\frac{y}{y}. \end{array} $$ ## Case analyses. Let us analyze all 9 cases, and see what kind of means we get. ### Cases $x/x$, $m/ m$, and $y /y$. Note if we set $$ \frac{y-m}{m-x} = \frac{x}{x} = \frac{m}{m} = \frac{y}{y} = 1, $$then $2m = x+y$, hence we get $$ m = \frac{1}{2}(x+y) = a(x,y) $$the arithmetic mean of $x,y$. ### Case $x / m$. If we set $$ \frac{y-m}{m-x} = \frac{x}{m} $$then we get $my - m^{2} = xm - x^{2}$, then $m^{2} + (x-y)m - x^{2} =0$, so $$ m = \frac{y-x + \sqrt{(y-x)^{2}+4x^{2}}}{2} $$where we have to have the $+$ sign for $m > 0$. We don't have a conventional name for this mean. Let us denote it $r(x,y)$. ### Case $x / y$. If we set $$ \frac{y-m}{m-x} = \frac{x}{y} $$then we get $y^{2}-my = mx - x^{2}$, so $x^{2}+y^{2}=m(x+y)$, so $$ m = \frac{x^{2}+y^{2}}{x+y} $$ We don't have a conventional name for this. Let us denote it $t(x,y)$. ### Case $m / x$ and $y / m$. If we set $$ \frac{y-m}{m-x} = \frac{m}{x} $$then we get $xy - mx = m^{2}-mx$, so $$ m = \sqrt{xy}, $$which is the geometric mean of $x,y$. Similarly, if we have $$ \frac{y-m}{m-x} = \frac{y}{m} $$then we get $my - m^{2} = my-xy$, so $m = \sqrt{xy}$ the geometric mean. ### Case $m / y$. If we set $$ \frac{y-m}{m-x} = \frac{m}{y} $$then we get $y^{2}-my = m^{2}-mx$, so $m^{2} + (y-x)m - y^{2} = 0$, so $$ m = \frac{-(y-x) + \sqrt{(y-x)^{2}+4y^{2}}}{2}. $$We don't have a conventional name for this mean. Let us denote it $s(x,y)$. ### Case $y / x$. If we set $$ \frac{y-m}{m-x} = \frac{y}{x} $$then we get $xy - mx = my - xy$, or $2xy = m(x+y)$, or $$ m = \frac{2xy}{x+y} = \frac{2}{\frac{1}{x}+\frac{1}{y}} = h(x,y) $$which is the harmonic mean of $x,y$. So the means $m$ that we get by setting $\displaystyle \frac{y-m}{m-x} = \frac{p}{q}$ where $p,q \in \{x,m,y\}$, for $0 < x < m < y$, gives the following table $$ \begin{array}{cc|ccc} & & & q \\ & & x & m & y \\ \hline & x & a & r & t \\ p & m & g & a & s \\ & y & h & g & a \end{array} $$where $a(x,y) = \frac{1}{2} (x+y)$, $g(x,y) = \sqrt{xy}$, $h(x,y) = \frac{2}{\frac{1}{x}+\frac{1}{y}}$, $r(x,y) = \frac{y-x + \sqrt{(y-x)^{2}+4x^{2}}}{2}$, $s(x,y) = \frac{-(y-x) + \sqrt{(y-x)^{2}+4y^{2}}}{2}$, and $t(x,y) = \frac{x^{2} +y^{2}}{x+y}$. Observe that $r(x,y)$ and $s(x,y)$ are not symmetric in the arguments, while the rest are symmetric functions. All these means are homogeneous, however, that for all $\lambda > 0$, $$ f(\lambda x, \lambda y) = \lambda f(x,y) $$where $f \in \{t, r, s, a, g, h\}$, $0 < x < y$. ## Inequalities. > **Claim / Conjecture** We claim we have for $0 < x< y$, $$ \max(x,y) \ge t(x,y) \ge r(x,y) \ge s(x,y) \ge a(x,y) \ge g(x,y) \ge h(x,y)\ge \min(x,y). $$ For the means $m$ equal to $t$, $r$, or $s$, note their constructions are such that $$ \frac{y-m}{m-x} < 1 \implies m > \frac{x + y}{2} = a(x,y) $$so we get $t,r,s > a$. And for $0 < x < m < y$, the ratios $$ \frac{x}{m} > \frac{x}{y} \text{ and } \frac{m}{y} > \frac{x}{y} $$so $r < t$ and $s < t$ respectively. **Remark.** By this same vein, it is also very easy to show $a > g > h$, by using their ratio definitions! ### What about $r$ and $s$ ? It remains to show $r > s$ Since these means are all homogeneous, it suffices to show for $x > 1$, $$ r(1,x) \ge s(1, x) $$ Note $$ r(1,x) = \frac{x-1 + \sqrt{(x-1)^{2}+4}}{2} $$and $$ s(1,x) = \frac{-(x-1) + \sqrt{(x-1)^{2}+4x^{2}}}{2} $$ Note $$ \begin{align*} r(1,x) - s(1,x) & = x-1 + \frac{1}{2}(\sqrt{(x-1)^{2}+4}-\sqrt{(x-1)^{2}+4x^{2}}) \\ \end{align*} $$ Is there a nice way to see that this is $\ge 0$ for all $x > 1$?