Identify R2 with C via (x1,x2)↔x=x1+ix2. For A=(acbd)∈M2(R) put z=2(a+d)+i(c−b),w=2(a−d)+i(b+c). 1. Show that Ax=zx+wxˉ for all x∈C. 2. Show that detA=∣z∣2−∣w∣2. 3. Show that the singular values of A are σ1=∣z∣+∣w∣ and σ2=∣z∣−∣w∣. 4. Show that ∥A∥F2=2(∣z∣2+∣w∣2).
Hint
For 1, compute z+w and z−w. For 3, compute ∣zx+wxˉ∣2 for ∣x∣=1, x=eiφ: the term 2Re(zwˉe2iφ) describes an interval when φ varies.
Show solutionHide solution
Solution
Idea. Every real-linear map of C splits uniquely into a complex-linear part x↦zx (a rotation-dilation) and a conjugate-linear part x↦wxˉ (a reflection-dilation). The coordinates (z,w) make the geometry of M2(R) explicit.
Part 1. First, z+w=22a+i(2c)=a+ic,z−w=22d+i(−2b)=d−ib. With x=x1+ix2 we have zx+wxˉ=(z+w)x1+i(z−w)x2=(a+ic)x1+(id+b)x2, whose real and imaginary parts are ax1+bx2 and cx1+dx2. That is exactly Ax read as a complex number.
(Conversely a=Re(z+w), c=Im(z+w), d=Re(z−w), b=−Im(z−w), so A↦(z,w) is a linear bijection M2(R)→C2.)
Part 2. Expanding, ∣z∣2−∣w∣2=4(a+d)2+(c−b)2−(a−d)2−(b+c)2=44ad+(c−b)2−(b+c)2=44ad−4bc=ad−bc.
Part 3. For x=eiφ we have ∣zx+wxˉ∣2=∣z∣2+∣w∣2+2Re(zx⋅wxˉ) and wxˉ=wˉx, so ∣Ax∣2=∣z∣2+∣w∣2+2Re(zwˉe2iφ). As φ runs through [0,2π), the number zwˉe2iφ runs around the circle of radius ∣z∣∣w∣, and its real part takes all values in [−∣z∣∣w∣,∣z∣∣w∣]. Hence ∣Ax∣2 ranges over [∣z∣2+∣w∣2−2∣z∣∣w∣,∣z∣2+∣w∣2+2∣z∣∣w∣]=[(∣z∣−∣w∣)2,(∣z∣+∣w∣)2]. Now ∣Ax∣2=xTATAx on the unit circle ranges exactly between the smallest and the largest eigenvalue of ATA, that is between σ22 and σ12. Therefore σ1=∣z∣+∣w∣ and σ2=∣z∣−∣w∣.
Part 4.∥A∥F2=σ12+σ22=(∣z∣+∣w∣)2+(∣z∣−∣w∣)2=2(∣z∣2+∣w∣2). (Directly: ∣z∣2+∣w∣2=4(a+d)2+(a−d)2+(c−b)2+(b+c)2=2a2+b2+c2+d2.)
Check (numerical). For 200 random 2×2 matrices, all four identities hold to machine precision. For the rotation R(θ) one finds z=eiθ, w=0, and for diag(1,−1) one finds z=0, w=1.
Remark.det is a quadratic form of signature (2,2) on M2(R)≅R4: ∣z∣2 contributes two positive squares and −∣w∣2 two negative ones. The formula ∥A∥=∣z∣+∣w∣ for the operator norm shows that its unit ball is the ℓ1-sum of two Euclidean discs.
Lesson. In M2(R), choose coordinates adapted to the action of SO(2)×SO(2) (rotations on both sides): z and w are the natural ones.