ORTHOGONALITY IN LINEAR METRIC SPACES

-- -
BY GARRETT BIRKHOFF
1. Statement of main theorem. Let B be any linear metric space of three
dimensions, whose points we shall suppose mapped linearly onto those of
ordinary space.
It is natural to call a vector issuing from a point p of B "perpendicular"
to a second such vector [in symbols, ---] if and only if there is no point on
the extended line through nearer to q than p.
Remark. Since translations of space are isometric, and uniform expansions
about the origin multiply all distances by a constant factor of proportionality,
p-Ip-7 implies that any vector parallel or anti-parall,el to is perpendicular
to any vector issuing from the same point and parallel or antiparallel to
Therefore it is legitimate to say that the direction of is perpendicular to the
-.
direction of -.
The main purpose of this paper is to prove
THEOREM 1. If -J.-7 implies prJ.pq, and if there is at most one perpendicular
from a given line to a point not on that line, then B is "equivalent" to cartesian
space (i.e., isometric with it under a linear transformation).

2. Outline of proof. The proof of Theorem 1 involves such simple ideas that
it is sufficient to sketch it.
First, let us fix on a particular linear representation of B in ordinary space.
It is clear that the metric of B is determined by the "unit pseudo-sphere" S of
points whose absolute values (in the terminology of yon Neumann) are unity.
It is also clear that S is a convex surface.
The argument then proceeds in two main steps. First it is shown that rela-
tive to any choice of cylindrical coSrdinates, the equation defining S is of the form
(1) r f(z).g(O).
Then it is shown (in effect) that any plane section of such a surface is an ellipse,
essentially completing the proof.
To establish equation (1), let us first note that the radius from the origin
o to any point s on S is perpendicular to every line in any plane of support of S
at s. Hence by the uniqueness and reciprocity of perpendicularity, S can have
at most one plane of support at s.
Received March 10, 1935.
As defined for instance by J. von Neumann, On complete topological spaces, Trans.
Am. Math. Soc., vol. 37 (1935), pp. 3-4. The reader’s attention is called to the definition
of orthogonality in B. D. Roberts’ On the geometry of abstract vector spaces, TShoku Math.
Jour., vol. 39 (1934), pp. 42-59, which is essentially different.
169
170 GARRETT BIRKHOFF

Consequently by an elementary (and furthermore obvious) lemma on convex

surfaces, S is a smooth convex surface having a unique continuously varying
tangent plane, whose direction maps S onto the sphere continuously. It follows
in,cidentally that the perpendiculars to any given direction form a plane, and
that all the perpendiculars in any given plane are perpendicular to a suitable
line--which is unique by assumption.
It is hence possible to pick the z-axis in an entirely arbitrary direction in B,
and then to pick x- and y-axes such that all three are mutually perpendicular.
Let us suppose that this has been done, and that in additicn cylindrical
coSrdinates have been set up, satisfying
(2) x rcos, y-- rsin, z z.
Consider the plane curves of intersection ,
of S with the planes z const.
Because S is a smooth convex surface, they must be smooth convex curves.
Besides, the tangential direction T(z, 0) to- at e 0 must be perpendicular
,
to (A) the radius to at 0, and (B) the z-axis [since it lies in the (x, y)-
plane]..Therefore the plane of directions perpendicular to T(z, 0) must be the
plane 0, and independent of z.
,
It follows that the curves are all similar, and that the equation for S may
be written in the form (1). The rest of the argument is based exclusively on
this fact.
First one considers the "great pseudo-circles" fl of intersection of S with the
planes c’onst. By (1) any one of them can be rotuted into ny other
under a linear transformation
(3) z’ z, O’ + const., r’ g(O’)/g(O).r
between the planes on which they lie leaving the z-axis fixed.
Then one recalls that the choice of the z-axis (of rotation) is rbitrary. There-
fore by iterated use of the construction just described, one can rotate .the
(x, z)-plane P first about the z-axis into the (y, z)-plane, then about the y-axis
into the plane through the y-axis and any line a :x Xz, y 0 in the (x, z)-
plane, and finally about ax back into P. Moreover, the following statements
are true.
(a) All three transformations are linear, nd hence their product is also.
(b) The product preserves orientation in P.
(c) The intersection 0 of S with P goes into itself.
(d) The intersection of S with the z-axis goes into the intersection of S
with ax.
The last two paragraphs show that there exists an orientation-preserving
homogeneous linear transformation of P into itself leuving 0 invariant and
carrying the z-axis into any straight line x ),z. But on the other hand, since
Proved for instance in T. Bonnesen and W. Fenchel, Theorie der konvexen KSrper,
Berlin, 1934, p. 13, bottom.
By integrating a homogeneous linear differential equation of the form dr/dO T(O),
where T(O) is continuous.
 ORTHOGONALITY IN LINEAR METRIC SPACES 171

0 is compact and does not contain the origin, the group G of all orientation-
preserving homogeneous linear transformations of P into itself leaving 0 in-
variant, is bounded--and therefore equivalent after a homogeneous linear trans-
formation h to a subgroup S of the orthogonal group. But since S is transitive
on the lines through the origin, it must be the entire orthogonal group--th.t is,
the entire group of rigid rotations.
Consequently the set of the transforms under S of any given point p of
is a circle, and, h(/0) being a simple closed curve containing this set, h(0) is a

space, the "absolute value" of a point (x, 0, z) is (x

theorem is true.
,
circle. Therefore under a suitable linear representation of B in coSrdinate
z) and the pythagorean
But perpendicularity and absolute value in B are intrinsic; hence the pythag-
orean theorem is true under all representations.
The conclusion of Theorem 1 now follows by decomposing (i) any vector into
its y-component and perpendicular (x, z)-component; (ii) the (x, z)-component
into perpendicular x- and z-components.

3. Applications. We can apply Theorem 1 immediately to the classification

of complete linear metric spaces ["B-spaces" of Banach] by proving

sets (xl,
x (xl, x) and y (yl,
real number, then
(a) x y is the element (x -k yl,
(b) Xx is the element (Xx, Xx);
--
THEOREM 2. Any complete linear metric space L of three or more dimensions,
in which perpendicularity is reciprocal and unique, is characterized by its (finite
or transfinite) cardinal "dimension-number" n. Its elements correspond to the
xn) of n real numbers satisfying x < in such a way that if
y) are any two elements

x yn);
(c) the distance between x and y is [ (x y)].
of L, and X is any

To prove Theorem 2, it is sufficient to observe that by Theorem 1 we can

define "inner products" having the properties usually attached to inner products
in Hilbert space. For these involve at most three independent elements---and
so can be proved by use of Theorem 1 within a 3-space containing them. Once
this has been done, the rest of the argument is familiar.
Another application yields an interesting geometrical definition of differential
geometries. If one defines "pseudo-differential geometries" as varieties in
which distance is given by integrating a differential form

(4) f (x, x,; dx/dt, dxn/dt)dt f ds

J. von Neumann, Almost periodic functions in a group, Trans. Amer. Math. Soc., vol. 36
(1934), p. 465, Theorem 19.
M. H. Stone, Linear transformations in Hilbert space, New York, 1932, pp. 3-4.
Cf. the mimeographed lecture notes of J. yon Neumann on Operator theory, Theorem
12.27, Princeton, 1934.
172 GARRETT BIRKHOFF

such that for fixed x, (x; dx/dt) behaves like an "absolute value" for dx/dt,
a condition which amounts to requiring that the coSrdinates be non-singular
and that smooth geodesics exist, then Theorem 2 can be restated as
THEOREM 3. A pseudo-differential geometry is a differential geometry [that is,
for fixed x, (x; dx/dt) is the square root of a positive definite quadratic function of
the dx/dt] if and only if local perpendicularity is unique and reciprocal.
If in a Finsler space F transversality (= orthogonality) is unique and recip-
rocal, it is easy to show that $(x; dx/dt) (x; dx/dt), so that F is a
"pseudo-differential" geometry. Hence. a Finsler space is a differential geom-
etry if and only if transversality is unique and reciprocal.
A final application yields a system of postulates for the geometry of ordinary
space, more geometrical than the purely arithmetical definition of cartesian
geometry, and simpler than the approach of Euclid.

4. Necessity of three dimensions. It is interesting that Theorem 1 is

definitely false without the restriction that B have three dimensions. In fact,
THEOREM 4. There exist infinitely many intrinsically different metric linear
spaces of two diaensions in which perpendicularity is reciprocal and unique.
Without going into detail, one may observe that a smooth convex simple
closed curve 2 in coSrdinate space is the image of the "unit pseudo-circle" of
elements of absolute value unity of a metric linear plane, if and only if the
diameter drawn through the origin parallel to the tangent at any point p of 2
cuts 2 in points where the tangents to 2 are parallel to the diameter through p.
This is a .property related to conjugacy in the diameters of an ellipse.
But any smooth convex arc 2 defined over the first quadrant, symmetrical
about the half-ray 0 /4, and perpendicular to the x-axis at its intersection
with it, can be uniquely continued by this very condition through the other
three quadrants. And by the symmetry about 0 /4, this will lead to a
simple closed curve carried into itself under reflection through the origin, and
hence into a satisfactory "unit pseudo-circle".
SOCIETY OF FELLOWS HARVARD UNIVERSITY.