Let $E\neq \{0\}$ be a Banach space. For each $p\in[1,\infty), $ we define $$E\oplus_p E = \{(x,y): x\in E, y\in E, \|(x,y)\| = \sqrt[p]{\|x\|^p + \|y\|^p}\}.$$ Let $F$ be another Banach space. By $E\cong F,$ I mean that $E$ and $F$ are isometrically isomorphic.

Question: Suppose that $p,q\in [1,\infty).$ If $$E\oplus_p E \cong E\oplus_q E\,$$ then is it true that $p=q$?

If $E$ is of finite-dimensional, then the question is affirmative. However, I do not know what will happen if $E$ is of infinite-dimensional. I would be glad to see a proof if it is true or a counterexample if it is false.

We say that a norm $\phi:\mathbb{R}^2\to\mathbb{R}$ is normalized if $$\phi(0,1) = \phi(1,0) = 1.$$

Also, $\phi$ is monotone if for $|a_1|\leq |b_1|$ and $|a_2|\leq |b_2|,$ then $$\phi(a_1,a_2) \leq \phi(b_1,b_2).$$

We define $$E\oplus_\phi E = \{(x,y): x\in E, y\in E, \|(x,y)\| = \phi(\|x\|, \|y\|) \}.$$

A more general question:

Suppose that $\phi,\psi:\mathbb{R}^2\to \mathbb{R}$ are norms that satisfy normalization and monotonicity. Assume that $\phi$ and $\psi$ are not $\ell^\infty$ norm. If $$E\oplus_\phi E \cong E\oplus_\psi E,$$ then is it true that $\phi = \psi?$?

  • 1
    $\begingroup$ Objection to your final statement since in the case where $E$ is the real line. then the cases $1$ and $\infty$ are isometric. Rather too silly to be worth mentioning but for the fact that it suggests looking at the extreme points of the unit balls. If they are what I suspect (i.e. situated on the axes and the diagonals---except for $p=2$), then this would prove your conjecture. $\endgroup$
    – cevasix
    Jul 31, 2018 at 13:25
  • $\begingroup$ As you might have noticed, but just in case, I do not assume that $p$ and $q$ are holder conjugate. $\endgroup$
    – Idonknow
    Jul 31, 2018 at 13:49
  • $\begingroup$ In the infinite dimensional case, the unit balls need not have any extreme points. $\endgroup$
    – user114263
    Jul 31, 2018 at 17:24
  • $\begingroup$ @DuchampGérardH.E. If $E = \mathbb{R},$ then I guess we can look at their set of extreme points and deduce that $p=q?$ Because for two norms to be the same, they must have the same set of extreme points.I think this is a characterization, that is, two norms are the same iff they have the same set of extreme points. $\endgroup$
    – Idonknow
    Aug 2, 2018 at 7:05
  • 1
    $\begingroup$ Since $E\cong F$ implies $E\oplus_\psi E \cong F\oplus_\psi F$, you might as well phrase the question like this: does $E\oplus_\phi E \cong E\oplus_\psi E$ imply $\phi = \psi$? (Nitpick: and you may want to assume $E\neq 0$, since otherwise the answer is trivially no.) $\endgroup$ Aug 3, 2018 at 12:46

2 Answers 2


It is a sketch of the proof in the case of $p\ne q$ (a complete proof on these lines is rather lengthy). Something similar can be done in more general case, possibly with some exceptions.

Our plan is the following: We assume that $p,q\in[1,\infty)$, $p\ne q$ and $W=E_1\oplus_p E_2=F_1\oplus_q F_2$ (isometrically), where $E_1,E_2,F_1,F_2$ are all isometric to a Banach space $E$, and get a contradiction.

By $S(X)$ we denote the unit sphere of a Banach space $X$, if $Z$ and $Y$ are subspaces of $X$, we set $\delta(Y,Z)=\inf\{||y-z||:~ y\in S(Y), z\in Z\}$ and call it an inclination of $Y$ to $Z$.

If we can find $i\ne j\in\{1,2\}$ such that $F_i$ has zero inclination to $E_1$ and $F_j$ has zero inclination to $E_2$, we get a contradiction by observing that it means that two-dimensional $\ell_p$-sphere approximates $\ell_q$-sphere with an arbitrary precision, which is not true ($p$ and $q$ are fixed).

[This paragraph is added on 8/8/18 according to the request below] Assume without loss of generality that $\delta(F_1,E_1)=\delta(F_2,E_2)=0$. This implies that there are sequences $\{x_i\}\in S(F_1)$, $\{y_i\}\in S(E_1)$, $\{z_i\}\in S(F_2)$, and $\{w_i\}\in S(E_2)$, such that $$\lim_{i\to\infty}||x_i-y_i||=0=\lim_{i\to\infty}||z_i-w_i||=0$$ Observe that our assumptions imply that the subspace $A_i$ spanned by $\{x_i,z_i\}$ is isometric to $\ell_q^2$ and the subspace $B_i$ spanned by $\{y_i,w_i\}$ is isometric to $\ell_p^2$. Therefore, using the argument of Proposition 1.a.9 in Lindenstrauss-Tzafriri, Classical Banach spaces, v. I we see that there should be isomorphisms $T_i$ between $\ell_p^2$ and $\ell_q^2$ mapping unit vector basis of $\ell_p^2$ onto the unit vector basis of $\ell_q^2$ and being arbitrarily close to the identity. This is clearly false if $p\ne q$.

It remains to consider the case where both $F_1$ and $F_2$ have nonzero inclination to $E_1$ (or to $E_2$, the cases are similar).

By $P_1$ and $P_2$ we denote projections on $W$ corresponding to the decomposition $E_1\oplus_p E_2$. Nonzero inclination to $E_1$ implies that the restriction of $P_2$ to both $F_1$ and $F_2$ are isomorphic embeddings. If there are nonzero points $y_1$ and $y_2$ in $F_1$ and $F_2$ which have the same image in $E_2,$ we get a contradiction by considering the space spanned by $y_1$ and $y_2$: on one hand its unit sphere is $\ell_p$-sphere, and on the other hand - $\ell_q$-sphere.

In a similar way (but using approximations) we get a contradiction in the case where the inclination of $P_2(F_1)$ to $P_2(F_2)$ is zero. Finally, if the inclination of $P_2(F_1)$ to $P_2(F_2)$ is nonzero, we get a contradiction because in this case $P_2$ would be an isomorphic embedding of $W$ into $E_2$, which is obviously false.

  • $\begingroup$ +1 but I'd like to know more about your argument in [if we can find ... ($p,q$ are fixed).] or a reference. $\endgroup$ Aug 8, 2018 at 7:26
  • $\begingroup$ @ Duchamp Gérard H. E. I have inserted the corresponding argument. $\endgroup$ Aug 8, 2018 at 15:31
  • 2
    $\begingroup$ @Idonknow It is a strange and impolite approach: to open a bounty and to pay no attention to an answer. $\endgroup$ Aug 8, 2018 at 20:15
  • $\begingroup$ Thanks. I suppose that you note $\ell_p^2$, the space $\ell_p\oplus \ell_p$. $\endgroup$ Aug 8, 2018 at 20:25
  • $\begingroup$ @ Duchamp Gérard H. E. I use the notation $\ell_p^2$ for two-dimensional $\ell_p$, it also can be written as $\mathbb{R}\oplus_p\mathbb{R}$. $\endgroup$ Aug 8, 2018 at 21:10

It was proved by E. Behrends in Studia Math. 55, 71-85 (1976) that apart from $E=\mathbb{R}^2$ with the sup norm, which is isometric to $E$ with the $\ell_1$-norm, a Banach space $E$ admits a decomposition $E_1\oplus_p E_2$ for at most one value of $p$. This theorem is what Misha has outlined above.

  • $\begingroup$ @ Dirk Werner. Welcome to MO! I did not know about this paper of Ehrhard, otherwise I would prefer to give your answer rather than mine. I was inspired to prove it myself because the proposer of the question was eager to know the answer and proposed a 100-point bounty. $\endgroup$ Aug 19, 2018 at 0:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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