Recall that
Sym:Rn×n→Sn by
Sym:A↦2A+A⊤. Let
∥⋅∥ denote the usual operator norm. For
positive semidefinite matrices, this corresponds to the maximum
eigenvalue.
Question 1. What values can
λmin(Sym(∏i=1m∥Ai∥Ai))
take if Ai are positive semidefinite?
Note that by convexity and submultiplicativity, we have
∥∥∥∥∥∥Sym(i∏∥Ai∥Ai)∥∥∥∥∥∥≤∥∥∥∥∥∥i∏∥Ai∥Ai∥∥∥∥∥∥≤1,
so that −I⪯Sym(i∏∥Ai∥Ai)⪯I.
Clearly, the latter bound is tight (e.g., take A1=⋯=Am=I).
In this post, we will consider the first bound more carefully and show:
Proposition 1. Suppose n≥2. Let
A1,…,Am∈S+n. Then, Sym(i∏∥Ai∥Ai)⪰−cos(m+1π)m+1I.
For example, the above proposition implies that for any
A,B,C⪰0, we have −81I⪯∥A∥op∥B∥opSym(AB)⪯I,−41I⪯∥A∥op∥B∥op∥C∥opSym(ABC)⪯I.
Proof. By homogeneity, it suffices to consider the case where
∥A1∥=⋯=∥Am∥≤1.
Let A1,…,Am∈S+n minimize Opt:=A1,…,Ammin{λmin(Sym(i∏Ai)):0⪯Ai⪯I}.
Without loss of generality, each Ai is a rank-one matrix. Indeed,
suppose rank(Ak)≥2. Let v0∈Sn−1
correspond to a minimum eigenvalue of
Sym(∏iAi). Let
w=(∏i>kAi)v0 and set
A~k=∥Akw∥2(Akw)(Akw)⊤.
Note that A~kw=w⊤(Ak)2ww⊤AkwAkw.
Then λmin(Sym(i<k∏Ai⋅A~k⋅i>k∏Ai))≤v0⊤(i<k∏Ai⋅A~k⋅i>k∏Ai)v0=w⊤(Ak)2ww⊤Akwλmin(Sym(i∏Ai))≤Opt. The last inequality
follows from the assumption that 0⪯Ak⪯I.
We may thus write each Ai=vivi⊤ for some
vi∈Sn−1. Equivalently, v0,v1,…,vm minimizes
v0,v1,…,vm∈Sn−1min⟨v0,v1⟩⟨v1,v2⟩…⟨vm,v0⟩.
By negating vi if necessary, we may assume that
⟨v0,v1⟩,…,⟨vm−1,vm⟩>0
and ⟨vm,v0⟩<0. By optimality and our
assumptions on the signs, we have that for each i=1,…,m−1,
vi=∥vi−1+vi+1∥vi−1+vi+1.
We deduce that
v1,…,vm−1∈span{v0,vm} so
that without loss of generality
v0,v1,…,vm∈S1. In particular, we may
parameterize vi=(cos(θ0+iη),sin(θ0+iη))
for some θ0 and η. Then, Opt=η∈Rmincos(η)ncos(nη).
This expression is minimized at η=n+1π where
cos(nη)=−cos(η). ◻