Sign up ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

If $d(x,y)$ and $e(x,y)$ are metrics then $d(x,y)+e(x,y)$ and $\frac{d(x,y)}{1+d(x,y)}$ are metrics.

If $d_i(x,y)$ for $i=1,\dots,n$ are metrics then so is $\sqrt{\sum_{i=1}^n{d_i^2(x,y)}}$

Are there other interesting ways of constructing new metrics from old metrics?

share|cite|improve this question

7 Answers 7

up vote 8 down vote accepted

Personally, I prefer $\min(d(x,y),\epsilon)$ over the standard $d(x,y)/(1+d(x,y))$ trick if the goal is to turn a metric into an equivalent metric – it's a lot easier to prove the metric axioms for this one.

From that follows one way to turn a countably infinite number of metrics into a new metric: $d(x,y)=\sum\_{n=1}^\infty \min(d\_n(x,y),2^{-n})$. The point being that convergence in the new metric is the same as convergence in all of the original ones. So this can be used to prove stuff like $C(\mathbb{R})$ with the topology of uniform convergence on compact sets is a metrizable space, and similarly for the space of test functions used in distribution theory. (The individual $d\_n$ could even be pseudometrics rather than metrics; it doesn't matter, so long as $x\ne y$ implies $d\_n(x,y)\ne0$ for some $n$.)

share|cite|improve this answer
Does the choice of convergent sequence matter? – Kim Greene Nov 16 '09 at 22:49
You mean convergent series? Nope. As far as generating topologies metrics are very flexible. – Qiaochu Yuan Nov 16 '09 at 22:58
Yes, I mean series. – Kim Greene Nov 17 '09 at 2:10

Generalizing d(x,y)/(1+d(x,y)): Let d(x,y) be a metric, and let $f: R^+ \to R^+$ be a function with $f(0) = 0$, $f^\prime(x) > 0$ for all x, and $f^{\prime\prime}(x) < 0$ for all x. Then I believe h(x,y) = f(d(x,y)) is a metric. For example, the square root of a metric is a metric.

share|cite|improve this answer
Does it always induce the same topology as d? – Qiaochu Yuan Nov 16 '09 at 22:45
I think so. They have the same balls. – Michael Lugo Nov 16 '09 at 23:05
Generalizing slightly more: instead of differentiability assumptions on $f:\mathbb{R}^+\to\mathbb{R}^+$ you only need $f(0)=0$, $f$ is non-decreasing, and $f$ is concave. This then also subsumes the $\min(d,\varepsilon)$ example. – Konrad Swanepoel Nov 17 '09 at 15:39
I would like to add that the $d^\alpha$ metrics, with $0<\alpha<1$, have some interesting properties. For example, every path has infinite length for any of these metrics, $d$ being any metric. It is an open question to determine for which $n,\alpha, N$ one can find a bi-Lipschitz embedding of $(\mathbb{R}^n,d^\alpha_{eucl})$ into $(\mathbb{R}^N,d_{eucl})$. – Benoît Kloeckner Feb 18 '12 at 6:07
Concave defined on all of $\mathbb{R}^{+}$ (and valued in $\mathbb{R}^{+}$) implies non-decreasing. Moreover, one can apply the same process to a bounded d, in which case non-decreasing assumption is not necessary. – Max M Oct 23 at 12:55

If you have a smooth submanifold of a Riemannian manifold, of course you have the subspace metric, but you also have the induced path metric, which is usually the "right" metric to think about.

share|cite|improve this answer

I have to confess to a dislike of the word "interesting" for these constructions! So far, all that's been said is:

  1. Any metric is equivalent to a bounded metric, so "boundedness" is not a topological property.

  2. Countable products of metric spaces are metrisable, so the category of metric spaces has countable products. (Incidentally this construction fails for arbitrary small products.)

To that I can add arbitrary small coproducts: if $x,y$ are in the same component, take $d(x,y) = min(d_X(x,y),1)$ and if $x,y$ are in different components, take $d(x,y) = 1$.

What this shows is that from a topological point of view, it's usually best to have a bounded metric.

But I would say that an interesting construction is one that is not functorial, for example, something where you replace the given metric locally by another one. For example, on a manifold we can always replace a given metric by one which is flat near a point (indeed, flat near quite a large number of points) but we can't necessarily replace it by one which is flat everywhere.

share|cite|improve this answer

This example can be found in Munkres (p. 133). If $X_i$ is a countable collection of metric spaces with metrics $d_i$, let $\bar{d_i} = \text{min}(d_i, 1)$ denote the standard bounded metric (which gives the same topology as $d_i$) and let $d = \sup \frac{ \bar{d_i} }{i}$ be defined on the product $\prod X_i$. Then this metric induces the product topology.

share|cite|improve this answer
I think OP wanted a new metric for a fixed set, say X, out of a metric or collection of metrics for X. – Jose Capco Nov 17 '09 at 18:58

The second example in the original post generalizes a lot. Let $d_i$ be finitely or countably many pseudometrics (it's possible for $d_i(x,y)=0$ even if $x\neq y$) for $i\ge 1$, and assume $d_1$ is a metric. Let $X$ be a space of sequences with a 1-unconditional norm, i.e., $$ \lVert (x_1,x_2,\ldots) \rVert = \lVert (\pm x_1, \pm x_2, \ldots) \rVert$$ for every choice of signs. For example, $X=\ell^p$ works. Then $$d(x,y) = \lVert (d_1(x,y),d_2(x,y),\ldots ) \rVert$$ is a metric.

For example, the metrics on $C^k$ spaces are examples of this kind of construction.

share|cite|improve this answer

Associated to any metric is a length structure. The length of an absolutely continuous path is the integral of it's metric derivative. Associated to any length structure is a (pseudo)-metric. If this is an honest metric, it is also a path metric: for any $x, y\in X$ and any $\varepsilon>0$, there is some $z\in X$ such that $d(x, z)+d(z, y)\le d(x, y)+\varepsilon$.

Taken from Gromov, "Metric structures for Riemannian and non-Riemannian spaces".

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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