Beruflich Dokumente
Kultur Dokumente
47
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
may obtain the cosine of some new “geometrical angle” 1.1 The basic vector orientation of rational
θ between the two given subspaces. In the special case trigonometry
where (u1 , u2 , · · · , uk ) and (v1 , v2 , · · · , vk ) are orthonormal
bases of k-dimensional subspaces P and R of Rn , their for- We begin by reviewing the basics of rational trigonometry,
mula reduces to which was introduced in the planar situation in the author’s
book [15], and to more general situations in [16] and [19].
cos2 θ = det M T M
(1) The main initial idea is to replace “transcendental metrical
in terms of the k × k matrix M = [ui · v j ], although their quantities” in Euclidean geometry with purely algebraic
original more general formulation had an error which was quantities. Thus the “distance” |A, B| between two points
corrected by Gunawan, Neswan and Setya-Budhi in [6]. A and B is replaced by the quadrance Q (A, B), which is
The reason that many of these concepts are put in quotes the sum of squares of the differences between the coordi-
here is that this author is skeptical about the current “the- nates, and the “angle” θ (l, m) between two lines l and m
ory of real numbers” and most claims involving computa- is replaced by the spread s (l, m), which may be thought of
tions that require an “infinite amount” of computing time as the ratio of the opposite quadrance over the hypotenuse
or power. But we do acknowledge that finite truncations quadrance for any right triangle formed by the two lines,
of such procedures can have a very useful applied func- or the cross, which is the ratio of the adjacent quadrance
tion. From our point of view even the usual definition of over the hypotenuse quadrance.
the “angle” between vectors u and v given by More generally and explicitly, if a vector space over a field
u·v is given a symmetric bilinear form, or dot product, u · v,
cos θ =
|u| |v| then we define the quadrance of a vector v to be the num-
ber
is problematic in the realm of pure mathematics, on ac-
count of the “infinite amounts” of work needed to evaluate
Q (v) ≡ v · v.
the square roots implicit in |u| and |v|, and then to evaluate
an inverse circular function.
A vector v is null precisely when Q (v) = 0. The cross be-
We are interested in investigating a framework for pure tween non-null vectors u and v is defined to be the number
mathematics that does not require us to pretend to obtain
outputs from unending algorithms or computer programs. (u · v)2
c (u, v) ≡ .
Now many readers may very well not share this orienta- Q (u) Q (v)
tion, but we do hope that they can see that it might nev-
ertheless be a legitimate logical position, and might ac- The spread between u and v is
tually steer us in natural and fruitful directions. This in
fact leads us to rational trigonometry [15], [16], [17], [18]; (u · v)2
s (u, v) ≡ 1 − c (u, v) = 1 − .
to chromogeometry and associated triangle geometry [9], Q (u) Q (v)
[10] and [19]; and to universal hyperbolic geometry, as in
[20], [21], [22] and [23], so there is plenty of evidence that Since both the cross and spread are invariant under rescal-
the approach has some merit. ing of either or both of the vectors, these quantities ex-
tend to metrical invariants between one-dimensional sub-
To avoid “angles” therefore, we are going to re-frame the
spaces, just by considering spanning or direction vectors.
separation between two k-dimensional subspaces in an n-
So given one-dimensional subspaces l and m with respec-
dimensional dot product space using an extension of a fun-
tive spanning vectors u and v, we may define the respec-
damental formula of rational trigonometry. Notice that we
tive cross and spread between them as c (l, m) ≡ c (u, v)
restrict to this more specific situation here, with subspaces
and s (l, m) ≡ s (u, v). These quantities are numbers in the
of equal dimension. We will investigate some basic prop-
erties of rational invariants obtained from the characteristic given field.
polynomial of a cross matrix determined by two such sub- For the vector space of row vectors of a given dimension
spaces, and in particular will prove a Pythagorean theorem, n, the Euclidean symmetric bilinear form may be written
or Diagonal Rule result, for the special but quite interesting in linear algebraic terms as
case of 2-subspaces in the four-dimensional Euclidean vec-
tor space V 4 of row vectors. The alternate new name for u · v = uvT .
this historically important result reflects the fact that the
Old Babylonian culture had a clear understanding of the A more general symmetric bilinear form is given by
geometry of a right triangle, more than 1000 years before
Pythagoras was born, as discussed for example in [11]. u · v = uMvT
48
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
for some symmetric n × n matrix M, which is usually also 1.2 The affine laws of rational trigonometry
assumed to be non-degenerate, meaning that det M 6= 0.
To enunciate the main laws of rational trigonometry, a
Let us make an important additional observation: that since first step is to extend the above notions from the vec-
u · v = v · u, the (Euclidean) cross may be rewritten in this tor space framework to the corresponding affine frame-
linear algebraic notation as work. We may consider the affine space An , whose points
2 are n-tuples A = [a1 , a2 , · · · , an ] with entries from some
uvT −1 T −1 field, but now with a translational symmetry as well as
c (u, v) = = uvT vvT vu uuT (2) the linear transformational symmetry. To a pair of such
(uuT ) (vvT )
affine points A, B we may associate the displacement vector
−→
while the more general cross may be written as v = AB = (b1 − a1 , b2 − a2 , · · · , bn − an ) in the usual fash-
ion by taking differences of coordinates.
2
uMvT Suppose now that the associated vector space V n of row
c (u, v) =
(uMuT ) (vMvT ) vectors v = (x1 , x2 , · · · , xn ) is given a non-degenerate sym-
−1 −1 metric bilinear form u · v as above. The quadrance be-
= uMvT vMvT vMuT uMuT . (3) tween the affine points A1 and A2 is then correspondingly
defined to be
How do these rational trigonometric definitions relate to −−→ −−→
Q (A1 , A2 ) ≡ A1 A2 · A1 A2 .
the more familiar notions of distance and angle? One may
attempt to “introduce a square-root” of the quadrance Q (v) The affine line A1 A2 determined by two points A1 and A2
and so define a length |v|, but this is problematic on at least is a null line precisely when Q (A1 , A2 ) = 0, that is when a
two counts; if working algebraically it generally requires direction vector for it is null.
an extension field, but analytically, this is a computational The cross between non-null lines l ≡ A1 A2 and m ≡ B1 B2
process that does not generally terminate, and so the os- is the number
tensible outputs cannot be said to be well-defined (hence −−→ −−→2
the quotes). Furthermore for other bilinear forms, as in A1 A2 · B1 B2
Einstein’s special theory of relativity, the quadrance can be c (l, m) ≡
Q (A1 , A2 ) Q (B1 , B2 )
negative, in which case we need also to introduce complex
extension fields. while the spread between l and m is
One may attempt to introduce an “inverse sine of a square −−→ −−→2
root” of the spread, or the “inverse cosine of a square root” A1 A2 · B1 B2
s (l, m) ≡ 1 − c (l, m) = 1 − .
of the cross, to get an “angle”, but again the transcenden- Q (A1 , A2 ) Q (B1 , B2 )
tal aspect of these operations and the functions involved
These are just the cross and spread of the associated direc-
means that in practice only an approximation to an “ideal
tion vectors, and are independent of the choice of points
angle” is obtained after terminating the program to get an
lying on the two lines. If one or both of the lines involved
output in a finite amount of time. And in other geome- are null, then the cross and spread are undefined, and state-
tries, as in Einstein’s relativistic geometry, the quantities ments involving them will be considered empty. Note also
that appear as crosses and spreads need not be in the famil- that we do not require the two lines to be meeting in order
iar interval [−1, 1], so “applying inverse circular functions” for these quantities to be defined.
is not directly appropriate.
Two non-null lines are perpendicular precisely when the
When we move to rational trigonometry as a framework cross between them is 0, or equivalently when the spread
for metrical geometry, many new possibilities for preci- between them is 1. These conditions are just a restate-
sion, clarity and generality open up. The quantities and ment of the orthogonality of corresponding direction vec-
their relations become algebraic and rational; we are not tors with respect to the underlying bilinear form.
forced to assume a “real number framework” involving an Here then are the five main laws of rational trigonom-
algebraic structure that is rarely, if ever, set out logically etry for this metrical affine situation, where we use the
and correctly in its entirety. We may aspire to obtain com- convention that a triangle A1 A2 A3 is a set {A1 , A2 , A3 }
plete and correct results to metrical questions without the of three distinct points with quadrances Q1 ≡ Q (A2 , A3 ),
need to invoke symbolic
√ arithmetic involving un-evaluated Q2 ≡ Q (A1 , A3 ) and Q3 ≡ Q (A1 , A2 ), and spreads
symbols such as π, 2, and cos 5, or resorting to approxi- s1 ≡ s (A1 A2 , A1 A3 ), s2 ≡ s (A2 A1 , A2 A3 ) and s3 ≡
mate values such as 3.1415, 1.4142 or 0.2837 etc. s (A3 A1 , A3 A2 ).
49
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
Theorem 1 (Diagonal Rule or Pythagoras’ theorem) hyperbolic geometry in which many new phenomenon are
The lines A1 A3 and A2 A3 are perpendicular precisely when visible, and which has closer connections to relativistic
physics on account of the fact that we can work uniformly
Q1 + Q2 = Q3 . inside or outside the light cone, as described in [20]. The
essential projective nature of this approach means that el-
Theorem 2 (Triple quad formula) The points A1 , A2 and liptic geometry is captured by the same laws as for hyper-
A3 are collinear precisely when bolic geometry – not just analogs where circular transcen-
dental functions are replaced by hyperbolic ones as in the
(Q1 + Q2 + Q3 )2 = 2 Q21 + Q22 + Q23 .
classical situation.
So we have enlarged geometry many-fold by moving to ra-
Theorem 3 (Spread law)
tional trigonometry for our metrical computations: we can
s1 s2 s3 work over the rational numbers; or over a finite field; we
= = .
Q1 Q2 Q3 can consider arbitrary bilinear forms, including relativistic
geometries; and we can create projective analogs of both
Theorem 4 (Cross law) elliptic and hyperbolic geometries simultaneously in the
same general arena.
(Q1 + Q2 − Q3 )2 = 4Q1 Q2 (1 − s3 ) .
But another vista still beckons by more explicitly adopting
a linear algebraic point of view. How might we extend ra-
Note that the Cross law includes as special cases both the tional trigonometry to higher dimensions, to consider met-
Triple quad formula and the Diagonal Rule when s3 = 0 rical relations in the spirit of Jordan’s “canonical or prin-
and s3 = 1 respectively. The next result is the algebraic cipal angles”, between higher dimensional objects in such
analog to the sum of the angles in a triangle formula. spaces? For example how does a symmetric bilinear form
allow us to define rational metrical relations between k-
Theorem 5 (Triple spread formula) dimensional subspaces of an n-dimensional vector V n or
an associated affine space? And what theorems can we
(s1 + s2 + s3 )2 = 2 s21 + s22 + s23 + 4s1 s2 s3 .
hope to find in these larger settings?
50
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
The det-cross agrees with the square of the “cosine of the So clearly c (P, MR) = c (P, R) and d (P, MR) = d (P, R) .
geometrical angle” of Risteski and Trenčevski as in (1), There is one more important invariance with respect to
which we can see by considering the special case when P these quantities: that they are unchanged under isome-
and R contain orthonormal vectors, so that PPT = RRT = I. tries of the Euclidean space with symmetric bilinear form
But our introduction of the cross matrix yields a richer in- u · v = uvT . Such an isometry is given by an n × n orthog-
variant that potentially gives us all the coefficients of the onal matrix Q, with the property that QQT = I, acting on
characteristic polynomial as invariants. These are some the right on row vectors and so also on k × n matrices.
kinds of analogs of the canonical angles of Jordan.
The cross, spread and det-cross are numbers in the under- Theorem 9 If P, R are both k × n matrices of rank k, and Q
lying field that depend on the matrices P and R. We now is an orthogonal n × n matrix, then C (PQ, RQ) = C (P, R)
demonstrate that they really depend only on the subspaces so that c (MP, R) = c (P, R) and d (MP, R) = d (P, R).
determined by P and R. We first show that if we rearrange
Proof. If Q is an orthogonal n × n matrix then
the rows of P by an invertible linear transformation, then
the new cross matrix is similar to the original, where two C (PQ, RQ) =
matrices X and Y are similar precisely when there is an −1 −1
=(PQ)(RQ)T (RQ)(RQ)T (RQ)(PQ)T (PQ)(PQ)T
invertible matrix M with
−1 −1
= PQQT RT RQQT RT RQQT PT PQQT PT
X = MY M −1 . −1 T −1
= PRT RRT RP PPT
Theorem 7 If P, R are both k × n matrices of rank k, = C (P, R) .
and M is an invertible k × k matrix, then C (MP, R) =
MC (P, R) M −1 so that c (MP, R) = c (P, R) and d (MP, R) = As before it follows that c (PQ, RQ) = c (P, R) and
d (P, R) . d (PQ, RQ) = d (P, R).
51
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
In the vector case, two vectors u and v are perpendicular A Diagonal Rule for 2-subspaces
precisely when s (u, v) = 1, or equivalently in terms of the
cross c (u, v) = 0. We now show that the same holds true The Diagonal Rule, or Pythagoras’ theorem, for vectors
for 2-subspaces in V 4 , assuming we are working over the may be stated as follows.
rational numbers. Note that the general situation, over a
different field, may be different! The computations in the
proof of this theorem are useful independently. Theorem 11 (One-dimensional Diagonal Rule) If P and
R are perpendicular 1-subspaces of a two-dimensional
Theorem 10 (Perpendicular 2-subspace) Suppose that space, that is for which c (P, R) = 0, then for any 1-
the underlying field is the rational numbers. If P is a subspace T we have
2-subspace of V 4 , then the only 2-subspace T for which
c (P, T ) = 0 is the orthogonal subspace T = P⊥ . c (P, T ) + c (R, T ) = 1.
52
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
and we know that c (P, R) = 0. Over the rational numbers Proof. Any two distinct 2-subspaces in V 3 meet in a one-
we know that R is the only 2-subspace with the prop- dimensional subspace. Let us assume, by performing an
erty that c (P, R) = 0, but we do not require that here. orthogonal transformation, that this common 1-subspace
Now any 2-dimensional subspace T , after row reduc- is the span of the vector (1, 0, 0) . By row reduction, the
tion, can be represented by one of the general matrices two 2-subspaces may be taken to be
T12 , T13 , T14 , T23 , T24 or T34 displayed in the proof of the
1 0 0 1 0 0
Perpendicular 2-subspace theorem above. That proof al- P= and R = .
0 a b 0 c d
ready established the crosses between P and each of these
Then
matrices. We now similarly calculate the crosses between
−1 T −1
these matrices and the matrix R: C = C (P, R) ≡ PRT RRT RP PPT
!
1 0
1 2w2 x2 + w2 − 4wxyz + x2 + 2y2 z2 + y2 + z2 = (ac+bd)2
c (R, T12 ) = 0
2 w2 x2 + w2 −2wxyz + x2 + y2 z2 + y2 + z2 + 1 (a2 +b2 )(c2 +d 2 )
1 w2 + 2y2 + 1 with cross
c (R, T13 ) =
2 w2 + y2 + 1 !
1 1 1 (ac + bd)2
c (R, T14 ) = c = c (P, R) = trC (P, R) = 1+ 2
2 2 2 (a + b2 ) (c2 + d 2 )
1 w2 + 2y2 + 1 and spread
c (R, T23 ) =
2 w2 + y2 + 1
1 s = s (P, R) = 1 − c (P, R)
c (R, T24 ) =
2 1 (ad − bc)2
= . (Spread formula in V 3 )
c (R, T34 ) = 1. 2 (a2 + b2 ) (c2 + d 2 )
This is one-half of the spread between the vectors u =
Then using the preceding computations, we get (0, a, b) and v = (0, c, d) .
As an immediate corollary we obtain the following:
c (P, T12 ) + c (R, T12 ) =
1 w2 + x2 + y2 + z2 + 2 Proposition 1 For any 2-subspaces P and R of V 3 over the
rational numbers the cross c (P, R) and the spread s (P, R)
2 w x + w − 2wxyz + x2 + y2 z2 + y2 + z2 + 1
2 2 2
satisfy the inequalities
1 2w2 x2 + w2 − 4wxyz + x2 + 2y2 z2 + y2 + z2
+ =1 1 1
2 w2 x2 + w2 − 2wxyz + x2 + y2 z2 + y2 + z2 + 1 ≤ c (P, R) ≤ 1, 0 ≤ s (P, R) ≤ .
2 2
53
KoG•21–2017 N J Wildberger: Rational Trigonometry in Higher Dimensions and a Diagonal Rule ...
Proof. Over the rational number field, squares are always [12] J. M IAO , A. B EN -I SRAEL, Product cosines of angles
positive. So the inequality for c (P, R) follows immediately between subspaces, Linear Algebra Appl. 237/238
from the formula for it in the previous proof, while the in- (1996), 71–81.
equality for s (P, R) follows from that.
[13] V. R AKO ČEVI Ć , H.K. W IMMER, A variational char-
acterization of canonical angles between subspaces,
Acknowledgements
J. Geometry 78 (2003), 122–124.
The author would like to thank Hendra Gunawan for help- [14] I.B. R ISTESKI , K.G. T REN ČEVSKI, Principal val-
ful discussions, and Michael Reynolds for his help in com- ues and principal subspaces of two subspaces of vec-
puting numerous examples of crosses and spreads in higher tor spaces with inner product, Beitr. Algebra Geom.,
dimensional spaces, as well as the referee for useful sug- Contributions to Algebra and Geometry 42(1) (2001),
gestions. 289–300.
[15] N.J. W ILDBERGER, Divine Proportions: Ratio-
References nal Trigonometry to Universal Geometry, Wild Egg
Books, Sydney, 2005, http://wildegg.com
[1] T.W. A NDERSON, An Introduction to Multivariate
Statistical Analysis, John Wiley & Sons, Inc., New [16] N.J. W ILDBERGER, Affine and projective metrical
York, 1958. geometry, 2007, http://arxiv.org/abs/math/
0701338
[2] A. B J ÖRCK , G.H. G OLUB, Numerical methods for
computing angles between linear subspaces, Math. [17] N.J. W ILDBERGER, A Rational Approach to
Comp. 27 (1973), 579–594. Trigonometry, Math Horiz. 15(2)(2007), 16–20.
[3] C. DAVIES , W. K AHAN, The rotation of eigenvectors [18] N.J. W ILDBERGER, One dimensional metrical ge-
by a perturbation. III, SIAM J. Numer. Anal. 7 (1970), ometry, Geom. Dedicata 128(1) (2007), 145–166.
1–46.
[19] N.J. W ILDBERGER, Chromogeometry and relativis-
[4] Z. D RMA Č, On principal angles between subspaces tic conics, KoG 13 (2009), 43–50.
of Euclidean space, SIAM J. Matrix Anal. Appl. (elec-
tronic) 22 (2000), 173–194. [20] N.J. W ILDBERGER, Universal Hyperbolic Geome-
try I: Trigonometry, Geom. Dedicata 163(1) (2013),
[5] A. G AL ÁNTAI , C S .J. H EGED ŨS, Jordan’s principal 215–274.
angles in complex vector spaces, Numer. Linear Al-
gebra Appl. 13 (2006), 589–598. [21] N.J. W ILDBERGER, Universal Hyperbolic Geometry
II: A pictorial overview, KoG 14 (2010), 3–24.
[6] H. G UNAWAN , O. N ESWAN , W. S ETYA -B UDHI, A
formula for angles between two subspaces of inner [22] N.J. W ILDBERGER, Universal Hyperbolic Geometry
product spaces, Beitr. Algebra Geom., Contributions III: First steps in projective triangle geometry, KoG
to Algebra and Geometry, 46(2) (2005), 311–320. 15, (2011), 25–49.
[7] C. J ORDAN, Essai sur la géométrie à n dimensions, [23] N.J. W ILDBERGER , A. A LKHALDI, Universal Hy-
Bull. Soc. Math. France. 3 (1875), 103–174. perbolic Geometry IV: Sydpoints and Twin Circum-
[8] A.V. K NYAZEV, M.E. A RGENTATI, Principal an- circles, KoG 16 (2012), 43–62.
gles between subspaces in an A-based scalar prod- [24] H.K. W IMMER, Canonical angles of unitary spaces
uct: algorithms and perturbation estimates, SIAM J. and perturbations of direct complements, Linear Al-
Sci. Comput. 23 (2002), 2008–2040. gebra Appl. 287 (1999), 373–379.
[9] N. L E , N.J. W ILDBERGER, Incenter Circles, Chro-
mogeometry, and the Omega Triangle, KoG 18
(2014), 5–18. N J Wildberger
orcid.org/0000-0003-3503-6495
[10] N. L E , N.J. W ILDBERGER, Universal Affine Trian-
e-mail: n.wildberger@unsw.edu.au
gle Geometry and Four-fold Incenter Symmetry, KoG
16 (2012), 63–80. School of Mathematics and Statistics UNSW
[11] D. M ANSFIELD , N.J. W ILDBERGER, Plimpton 322 Sydney 2052 Australia
is Babylonian Exact Sexagesimal Trigonometry, His-
toria Math. 44(4) (2017), 395–419.
54