Characterizing norm triangle inequalites via convexity

TL;DR

I walk through a cool and possibly less known result connecting convexity and the triangle inequalities for norms. Using this result, typical proofs of the triangle inequality for a proposed norm function are significantly simplified. This exposition is based on (Chapter 3, Robinson 2020) ¹.

(Robinson 2020) is a wonderful functional analysis book, highly recommend it 💯!

Background - Norms

Experienced readers can freely skip this Background section.

Normed linear spaces are a natural setting for much applied mathematics and statistics. These are vector spaces, $V$, endowed with a norm function, $\Vert \cdot {\Vert }_V$. Intuitively, norms give us a “yardstick” to measure the “lengths” of individual vectors in the given vector space space. A standard definition of a norm is as follows:

Definition 1 (Norms in vector spaces) For a given vector space $V$, a norm $\Vert \cdot {\Vert }_V:V\to \mathbb{R}$, is a function satisfying the following three properties.

1. Positive definiteness: For a $\mathbf{x}\in V$, if $\Vert \mathbf{x}\Vert =0$ then $\mathbf{x}={\mathbf{0}}_V$.
2. Absolute homogeneity: $\Vert \lambda \mathbf{x}\Vert =|\lambda |\Vert \mathbf{x}\Vert$, for a $\mathbf{x}\in V,\lambda \in \mathbb{R}$.
3. Triangle inequality: $\Vert \mathbf{x}+\mathbf{y}\Vert \le \Vert \mathbf{x}\Vert +\Vert \mathbf{y}\Vert$, for a $\mathbf{x},\mathbf{y}\in V$.

We often use many norms on a given vector space, depending on which seems “meaningful” for the given purpose. See here for common examples.

Remarks

Remark (Derived properties from Definition 1). We note that a norm, per Definition 1, in fact, implies the following properties:

1. In Definition 1, we can always replace positive definiteness with the stronger claim, namely that $$\text{For a }\mathbf{x}\in V\text{, if }\Vert \mathbf{x}\Vert =0⟺\mathbf{x}={\mathbf{0}}_V\text{.}$$ In short, we want to show that the reverse implication to positive definiteness always holds, i.e., $\mathbf{x}={\mathbf{0}}_V⟹\Vert \mathbf{x}\Vert =0$. To prove this observe that using absolute homogeneity in Definition 1, we have: $$\Vert \mathbf{x}\Vert =\Vert {\mathbf{0}}_V\Vert =\Vert 0({\mathbf{0}}_V)\Vert =|0|\Vert {\mathbf{0}}_V\Vert =0$$ As required.

2. We also have that $\Vert \mathbf{x}\Vert \ge 0$, for a $\mathbf{x}\in V$. To see this, observe that for a $\mathbf{x}\in V$ $$\begin{array}{rl}0& =\Vert {\mathbf{0}}_V\Vert \text{(by previous part of this remark)}\\ & =\Vert \mathbf{x}+(-\mathbf{x})\Vert \\ & \le \Vert \mathbf{x}\Vert +\Vert -\mathbf{x}\Vert \text{(by the triangle inequality)}\\ & =\Vert \mathbf{x}\Vert +|-1|\Vert \mathbf{x}\Vert \text{(by absolute homogeneity)}\\ & =2\Vert \mathbf{x}\Vert \\ ⟹\Vert \mathbf{x}\Vert & \ge 0\end{array}$$ In effect this means the co-domain can always be changed from $\Vert \cdot {\Vert }_V:V\to \mathbb{R}$ to $\Vert \cdot {\Vert }_V:V\to {\mathbb{R}}_{\ge 0}$.

3. Since these can always be derived directly from Definition 1, as shown, we can keep Definition 1 in its minimal form as noted here.

4. These ideas work for seminorms as well, see here for more details.

Main theorem

This is based on (Lemmas 3.3-3.4 Robinson 2020).

Theorem 1 (Characterization of norm triangle inequality) Let $N:V\to {\mathbb{R}}_{\ge \mathbb{0}}$, be a function satisfying the following two properties².

1. Positive definiteness: For a $\mathbf{x}\in V$, if $N(\mathbf{x})=0$ then $\mathbf{x}={\mathbf{0}}_V$.
2. Absolute homogeneity: $N(\lambda \mathbf{x})=|\lambda |N(\mathbf{x})$, for a $\mathbf{x}\in V,\lambda \in \mathbb{R}$.

We then have that:

$$\begin{array}{}\text{(1)}& N(\mathbf{x}+\mathbf{y})\le N(\mathbf{x})+N(\mathbf{y})\text{, for each }\mathbf{x},\mathbf{y}\in V⟺\mathbb{B}:=\{\mathbf{z}\in V|N(\mathbf{z})\le 1\}\text{ is convex}\end{array}$$

In simple terms, the importance of Theorem 1 (as captured by Equation 1) can be summarized as follows:

Let $N:V\to [0,\mathrm{\infty })$ be a function satisfying positive definiteness and absolute homogeneity. Then $N$ satisfies the triangle inequality if and only if the unit ba induced by $N$, i.e., $\mathbb{B}:=\{\mathbf{z}\in V|N(\mathbf{z})\le 1\}$, is a convex set.

Remarks

Remark. In Theorem 1, we note the following:

1. The function $N:V\to {\mathbb{R}}_{\ge 0}$, is a norm-like function, and only becomes a valid norm per Definition 1 once we establish the triangle inequality, i.e., $N(\mathbf{x}+\mathbf{y})\le N(\mathbf{x})+N(\mathbf{y})$.
2. To prove the triangle inequality for $N:V\to {\mathbb{R}}_{\ge 0}$, the necessary condition of Theorem 1 to establish is: $$\begin{array}{}\text{(2)}& \mathbb{B}:=\{\mathbf{z}\in V|N(\mathbf{z})\le 1\}\text{ is convex}\end{array}$$ which will imply the triangle inequality for $N$ - huzzah!
3. The nice thing is, proving the convexity of $\mathbb{B}$ can be much easier to show than trying to prove the triangle inequality property of $N$ directly, as we will soon see.
4. Subtle point: note that here we had to assume that the co-domain of $N$ is non-negative (not $\mathbb{R}$), i.e., $N:V\to {\mathbb{R}}_{\ge 0}$. This is because in a typical norm, which satisfies the triangle inequality, is always shown to be non-negative (see remark below Definition 1 for more details). Here we impose non-negativity of $N$ as an additional constraint to establish the triangle inequality property for $N$. This is not an issue, since one would always first check the non-negativity of a candidate norm-like function $N$.

Applications: Minkowski inequalities

Before getting into the details of the proof, let’s just see what Theorem Theorem 1 can do! We’ll consider two related applications taken from (Lemma 3.6, Example 3.13 Robinson 2020), respectively.

Application 1: ${\ell }_p$-norm triangle inequality in ${\mathbb{F}}^n$

Example 1 (Minkowski inequality in finite dimensions) Let us consider $({\mathbb{F}}^n,\mathbb{F})$, where $\mathbb{F}=\mathbb{R}\text{ or }\mathbb{C}$. We then define the norm-like function $N_{{\ell }^p}:{\mathbb{F}}^n\to {\mathbb{R}}_{\ge 0}$:

$$\begin{array}{}\text{(3)}& N_{{\ell }^p}(\mathbf{x}):={\left(\sum _{j=1}^n{\left|x_j\right|}^p\right)}^{1/p},1\le p<\mathrm{\infty }\end{array}$$

One can show that $N_{{\ell }^p}$ satisfies positive definiteness and absolute homogeneity. To show that $N_{{\ell }^p}$ is a norm function we need to prove the triangle inequality. We will use Theorem 1. Let us define $\mathbb{B}:=\{\mathbf{z}\in {\mathbb{F}}^n|N_{{\ell }^p}(\mathbf{z})\le 1\}=\{\mathbf{z}\in {\mathbb{F}}^n|N_{{\ell }^p}^p(\mathbf{z})\le 1\}$. We now need to show that $\mathbb{B}$ is convex. We will need to use the fact that for each $t\in \mathbb{R},t\mapsto |t{|}^p$ is convex. Let $\mathbf{x},\mathbf{y}\in \mathbb{B}$, we then have that for $\lambda \in [0,1]$:

Note: $t\mapsto |t{|}^p$, is convex, for each $t\in \mathbb{R}$ and $p\in [1,\mathrm{\infty })$. Sketch: We have, $|t{|}^p=:h(t):=(g\circ f)(t)$. Here, for each $x\in \mathbb{R}$, we observe $g:x\mapsto x^p$ is increasing and convex, and $f:x\mapsto |x|$ is convex. Thus, $h(t):=|t{|}^p$ is convex. $◼$

$$\begin{array}{rl}{N}_{{\ell }^p}^p(\lambda \mathbf{x}+(1-\lambda )\mathbf{y})& =\sum _{j=1}^n|\lambda |x_j|+(1-\lambda )|y_j|{|}^p\text{(by definition)}\\ & \le \sum _{j=1}^n\lambda {\left|x_j\right|}^p+(1-\lambda ){\left|y_j\right|}^p\text{(since }t\mapsto |t{|}^p\text{ is convex for each }t\in \mathbb{R}\text{)}\\ & =\lambda \sum _{j=1}^n{\left|x_j\right|}^p+(1-\lambda )\sum _{j=1}^n{\left|y_j\right|}^p\\ & \le 1\text{(since }\mathbf{x},\mathbf{y}\in \mathbb{B}\text{.)}\end{array}$$

It follows that $N_{{\ell }^p}(\lambda \mathbf{x}+(1-\lambda )\mathbf{y})\le 1$, and so $\lambda \mathbf{x}+(1-\lambda )\mathbf{y}\in \mathbb{B}$, as required $◼$.

This triangle inequality is proved in just four lines 🔥!

In fact, since it $N_{{\ell }^p}$ satisfies the three conditions for a norm per Definition 1 we can now denote it using the conventional ${\ell }_p$-norm form, i.e., $\parallel \mathbf{x}{\parallel }_{{\ell }^p}:=N_{{\ell }^p}(\mathbf{x})$

Application 2: $L_p$-norm triangle inequality

We can also similarly prove the triangle inequality norms involving integrals efficiently. This is seen in the next example.

Here $C([0,1])$ is the space of continuous functions on the interval $[0,1]$.

Let us consider $(C([0,1]),\mathbb{R})$. We then define the norm-like function $N_{L^p}:C([0,1])\to {\mathbb{R}}_{\ge 0}$:

$$\begin{array}{}\text{(4)}& N_{L^p}(\mathbf{x}):={\left({\int }_0^1{|f(x)|}^p\right)}^{1/p},1\le p<\mathrm{\infty }\end{array}$$

Let us define $\mathbb{B}:=\{h\in C([0,1])|N_{L^p}(h)\le 1\}=\{h\in C([0,1])|N_{L^p}^p(h)\le 1\}$. We now need to show that $\mathbb{B}$ is convex. Let $f,g\in \mathbb{B}$, we then have that for $\lambda \in [0,1]$:

Once again, one can show that $N_{{\ell }^p}$ satisfies positive definiteness and absolute homogeneity.

$$\begin{array}{rl}{N}_{L^p}^p(\lambda f+(1-\lambda )g)& ={\int }_0^1|\lambda f(x)+(1-\lambda )g(x){|}^{p}dx\text{(by definition)}\\ & \le {\int }_0^1\lambda |f(x){|}^p+(1-\lambda )|g(x){|}^{p}dx\text{(since }t\mapsto |t{|}^p\text{ is convex for each }t\in \mathbb{R}\text{)}\\ & =\lambda {\int }_0^1|f(x){|}^{p}dx+(1-\lambda ){\int }_0^1|g(x){|}^{p}dx\\ & \le 1\text{(since }f,g\in \mathbb{B}\text{.)}\end{array}$$

It follows that $N_{L^p}(\lambda f+(1-\lambda )g)\le 1$, and so $\lambda f+(1-\lambda )g\in \mathbb{B}$, as required $◼$.

Again, we can now denote $N_{L^p}$ using the conventional $L_p$-norm form, i.e., $\parallel f{\parallel }_{L^p}:=N_{L^p}(f)$.

Punchline: what did Theorem 1 buy us?

We just saw that applying Theorem 1 enabled us to write very short proofs of Minkowski’s inequality in ${\mathbb{F}}^n$ and $C([0,1])$.

To appreciate this approach, note that proving Minkowski’s inequality typicay requires one to first prove Young’s inequality and then Hölder’s inequality. Moreover these need to be done separately in ${\mathbb{F}}^n$ and $C([0,1])$. Using Theorem 1 aowed us to achieve both of these goals using near identical style of proofs 🎉!

Note: The proofs of Young’s and Hölder’s inequality are beautiful and studying them is also insightful.

Proof of Theorem 1

Assuming $N:V\to {\mathbb{R}}_{\ge \mathbb{0}}$ satisfies the two properties in Theorem Theorem 1, we need to prove both implications in Equation Equation 1.

Proof - easy direction

Assume that $N(\mathbf{x}+\mathbf{y})\le N(\mathbf{x})+N(\mathbf{y})$, for each $\mathbf{x},\mathbf{y}\in V$. Let $\lambda \in [0,1]$ be arbitrary. We need to show that this implies for each $\mathbf{x},\mathbf{y}\in \mathbb{B}$ that the expression $\lambda \mathbf{x}+(1-\lambda )\mathbf{y}\in \mathbb{B}$ holds. This implies the convexity of $\mathbb{B}$.

Proof

Proof. ($⟹$) We proceed directly.

Assume that $N(\mathbf{x}+\mathbf{y})\le N(\mathbf{x})+N(\mathbf{y})$, for each $\mathbf{x},\mathbf{y}\in V$. Let $\lambda \in [0,1]$ be arbitrary. We need to show that this implies for each $\mathbf{x},\mathbf{y}\in \mathbb{B}$ that the expression $\lambda \mathbf{x}+(1-\lambda )\mathbf{y}\in \mathbb{B}$ holds. This implies the convexity of $\mathbb{B}$.

We observe that for $\lambda \in \{0,1\}$ our required expression is equal to either $\mathbf{x}$ or $\mathbf{y}$ which are both in $\mathbb{B}$, by assumption. Now fix $\lambda \in (0,1)$. We then note:

$$\begin{array}{rl}N(\lambda \mathbf{x}+(1-\lambda )\mathbf{y})& \le N(\lambda \mathbf{x})+N((1-\lambda )\mathbf{y})\text{(}N\text{ satisfies triangle inequality)}\\ & =|\lambda |N(\mathbf{x})+|1-\lambda |N(\mathbf{y})\text{(by absolute homogeneity of }N\text{)}\\ & =\lambda N(\mathbf{x})+1-\lambda N(\mathbf{y})\text{(since }\lambda >0\text{)}\\ & \le (\lambda )(1)+(1-\lambda )(1)\text{(since }N(\mathbf{z})\le 1\text{, for }\mathbf{z}\in \mathbb{B}\text{)}\\ & =1\\ ⟹\lambda \mathbf{x}+(1-\lambda )\mathbf{y}& \in \mathbb{B}\end{array}$$

Which implies the convexity of $\mathbb{B}$, as required. $◼$

Proof - interesting direction

Assume $\mathbb{B}$ is a convex set. We need to show that this implies that $N(\mathbf{x}+\mathbf{y})\le N(\mathbf{x})+N(\mathbf{y})$, for each $\mathbf{x},\mathbf{y}\in V$.

Tip: In linear algebra proofs, it is best to first deal with the zero vector, ${\mathbf{0}}_V$, separately.

Proof

Proof. ($⟸$) We proceed directly.

Assume $\mathbb{B}$ is a convex set. We need to show that this implies that $N(\mathbf{x}+\mathbf{y})\le N(\mathbf{x})+N(\mathbf{y})$, for each $\mathbf{x},\mathbf{y}\in V$.

Let $\mathbf{x},\mathbf{y}\in V$. We will consider four cases.

Case 1: Let $\mathbf{x}=\mathbf{y}={\mathbf{0}}_V$. Then $N(\mathbf{x})=N(\mathbf{y})=N({\mathbf{0}}_V)=N(0{\mathbf{0}}_V)=|0|N({\mathbf{0}}_V)=0$, by absolute homogeneity of $N$. Indeed we then have that $N(\mathbf{x}+\mathbf{y})=N({\mathbf{0}}_V)=0=N(\mathbf{x})+N(\mathbf{y})$, as required.

Case 2: Let $\mathbf{x}={\mathbf{0}}_V,\mathbf{y}\in V\setminus \{{\mathbf{0}}_V\}$. Then $N(\mathbf{y})=0$, and it follows that $N(\mathbf{x}+\mathbf{y})=N(\mathbf{x}+{\mathbf{0}}_V)=N(\mathbf{x})=N(\mathbf{x})+0=N(\mathbf{x})+N(\mathbf{y})$, as required.

Case 3: Let $\mathbf{x}\in V\setminus \{{\mathbf{0}}_V\},\mathbf{y}={\mathbf{0}}_V$. Same as Case 2, with the roles of $\mathbf{x},\mathbf{y}$ reversed.

Case 4: Let $\mathbf{x},\mathbf{y}\in V\setminus \{{\mathbf{0}}_V\}$. It then follows that $N(\mathbf{x}),N(\mathbf{y})>0$, since $N(\mathbf{z})\ge 0$, for each $\mathbf{z}\in V$, and $N(\mathbf{z})=0⟺\mathbf{z}={\mathbf{0}}_V$. Moreover we then have that $|N(\mathbf{x})|=N(\mathbf{x})>0$ and $|N(\mathbf{y})|=N(\mathbf{y})>0$. So we can safely divide by these quantities. Let us then define $\tilde{\mathbf{x}}:=\frac{\mathbf{x}}{N(\mathbf{x})},\tilde{\mathbf{y}}:=\frac{\mathbf{y}}{N(\mathbf{y})}$. We then have by absolute homogeneity of $N$ that, $N(\tilde{\mathbf{x}}):=N\left(\frac{\mathbf{x}}{N(\mathbf{x})}\right)=\left|\frac{1}{N(\mathbf{x})}\right|N(\mathbf{x})=1⟹\tilde{\mathbf{x}}\in \mathbb{B}$. Similarly $\tilde{\mathbf{y}}\in \mathbb{B}$. Let us denote $\lambda :=\frac{N(\mathbf{x})}{N(\mathbf{x})+N(\mathbf{y})}\in (0,1)$, and $\mathbf{z}:=\frac{\mathbf{x}+\mathbf{y}}{N(\mathbf{x})+N(\mathbf{y})}$. We then have: $$\begin{array}{rl}\mathbf{z}& :=\frac{\mathbf{x}+\mathbf{y}}{N(\mathbf{x})+N(\mathbf{y})}\\ & =\left(\frac{N(\mathbf{x})}{N(\mathbf{x})+N(\mathbf{y})}\right)\left(\frac{\mathbf{x}}{N(\mathbf{x})}\right)+\left(\frac{N(\mathbf{y})}{N(\mathbf{x})+N(\mathbf{y})}\right)\left(\frac{\mathbf{y}}{N(\mathbf{y})}\right)\\ & =\left(\frac{N(\mathbf{x})}{N(\mathbf{x})+N(\mathbf{y})}\right)\tilde{\mathbf{x}}+\left(\frac{N(\mathbf{y})}{N(\mathbf{x})+N(\mathbf{y})}\right)\tilde{\mathbf{y}}\\ & =\lambda \tilde{\mathbf{x}}+(1-\lambda )\tilde{\mathbf{y}}\\ & \in \mathbb{B}\end{array}$$

By the assumed convexity of $\mathbb{B}$. We then have that $\mathbf{z}:=\frac{\mathbf{x}+\mathbf{y}}{N(\mathbf{x})+N(\mathbf{y})}\in \mathbb{B}⟹N(\mathbf{z})\le 1$. We then observe: $$\begin{array}{rl}N(\mathbf{z})& \le 1\text{(since }\mathbf{z}\in \mathbb{B}\text{.)}\\ ⟺N\left(\frac{\mathbf{x}+\mathbf{y}}{N(\mathbf{x})+N(\mathbf{y})}\right)& \le 1\text{(by definition of }\mathbf{z}\text{.)}\\ ⟺\left|\frac{1}{N(\mathbf{x})+N(\mathbf{y})}\right|N(\mathbf{x}+\mathbf{y})& \le 1\text{(absolute homogeneity of }N\text{.)}\\ ⟺\frac{1}{N(\mathbf{x})+N(\mathbf{y})}N(\mathbf{x}+\mathbf{y})& \le 1\text{(since }N(\mathbf{x}),N(\mathbf{y})>0\text{.)}\\ ⟺N(\mathbf{x}+\mathbf{y})& \le N(\mathbf{x})+N(\mathbf{y})\end{array}$$

As required. $◼$

Recap

In this article we learned the following about Theorem 1:

• It gives an alternative way to characterize the triangle inequality for norm-like functions.
• Using this characterization we can prove the triangle inequality for such norm-like functions using the convexity of the unit ba induced by such functions.
• This is usuay easier since we have lots of tools from convex analysis to help us prove the convexity of $\mathbb{B}$.
• We saw this in action since Theorem 1 enabled us to write very short proofs of Minkowski’s inequality in ${\mathbb{F}}^n$ and $C([0,1])$.

In summary, if you have a norm-like function for which you are trying to establish the triangle inequality, try out Theorem 1 💯!

Acknowledgements

I thank Prof. James Robinson for providing several technical clarifications on Theorem 1. I thank Mikhail Popov for creating the wikipediapreview R package, which enable an easy interface for Wikipedia Context Cards in Rmd files. These Context Cards enable the hover over preview for Wikipedia articles. I thank Jewel Johnson for providing this helpful guide to enable fixed TOC for this article. I thank Dr. Joel Nitta for providing these instructions to enable me to switch to the giscus comments system. Much of these distill site improvements were brought to our attention due to the excellent distillery site run by Prof. John Paul Helveston.

References

Robinson, James C. 2020. An Introduction to Functional Analysis. Cambridge University Press.

Footnotes

1. Note: The presentation in this post is intentionay verbose. The goal is to give lots of intuition of the key result and its usefulness, and ensure that the proofs are rigorous. It is written with an empathetic mindset to newcomers, and to myself for future reference.

2. We refer to such an $N:V\to {\mathbb{R}}_{\ge \mathbb{0}}$ satisfying these properties, as a norm-like function.

Reuse

https://creativecommons.org/licenses/by/4.0/

Citation

BibTeX citation:

@online{shrotriya2022,
  author = {Shamindra Shrotriya},
  editor = {},
  title = {Characterizing Norm Triangle Inequalites via Convexity},
  date = {2022-02-12},
  url = {https://www.shamindras.com/posts/2021-12-31-shrotriya2021normtriconvexity},
  langid = {en}
}

For attribution, please cite this work as:

Shamindra Shrotriya. 2022. “Characterizing Norm Triangle Inequalites via Convexity.” February 12, 2022. https://www.shamindras.com/posts/2021-12-31-shrotriya2021normtriconvexity.