Why observables have real eigenvalues, and what e^{-iHt} really means.
Why is e−iHt unitary? What do ρ and logρ mean? One theorem answers all of these: normal operators are exactly those diagonalized by an orthonormal basis, and after that a function of an operator is just a function of numbers.
What you’ll learn
State and prove the finite-dimensional spectral theorem via the Schur decomposition.
Define f(A) for normal A by functional calculus and use its algebraic properties.
Derive e−iθn^⋅σ/2=cos2θI−isin2θn^⋅σ.
Prove e−iHt−e−iH′t∞≤t∥H−H′∥∞ and state the first-order Trotter bound.
Explain operator monotonicity and why A≥B≥0 does not imply A2≥B2.
Normal operators and the spectral theorem
DerivationProof via Schur triangularization
A=UTU†,Uunitary,Tupper triangular
Schur. Every A∈Cd×d has an eigenvector (its characteristic polynomial has a root in C). Normalize it, extend to an orthonormal basis, and induct on the orthogonal complement:
(T†T)11=∣t11∣2,(TT†)11=j=1∑d∣t1j∣2
Unitary conjugation preserves normality, so T is normal. Compare the (1,1) entries of T†T (column 1 of T) and TT† (row 1):
T=diag(λ1,…,λd)⇒A=i∑λi∣ui⟩⟨ui∣
Equality forces t1j=0 for j>1. Delete the first row and column (still normal and triangular) and repeat: T is diagonal, and the columns ui of U are orthonormal eigenvectors.
AA†=UDD∗U†=UD∗DU†=A†A
Conversely, if A=UDU† with D diagonal, then both products equal U∣D∣2U†. Grouping equal eigenvalues gives the projector form.
Check your understanding
Which matrix is normal but neither Hermitian nor unitary?
Y
(1011)
(0010)
diag(1,2i)
Show the answer
D.diag(1,2i)
Every diagonal matrix is normal. diag(1,2i) has a non-real eigenvalue (so it is not Hermitian) of modulus 2 (so it is not unitary). Normal means unitarily diagonalizable, with arbitrary complex eigenvalues.
Functional calculus
The map f↦f(A) is an algebra homomorphism: (fg)(A)=f(A)g(A), (f+g)(A)=f(A)+g(A) and fˉ(A)=f(A)†. All the f(A) are diagonal in one basis, so they commute; specf(A)=f(specA); and ∥f(A)∥∞=maxk∣f(λk)∣. This one definition supplies every operator function in quantum information:
ρ and ρ−1/2 (fidelity, the pretty-good measurement), and logρ (von Neumann entropy−Trρlogρ, with 0log0=0);
∣A∣=A†A for any A (trace norm, polar decomposition A=V∣A∣);
e−iHt (Hamiltonian evolution) and e−βH/Tre−βH (thermal states).
DerivationPauli exponentials from the spectral theorem
σn^2=∣n^∣2I=I
Let n^∈R3 be a unit vector and σn^=nxX+nyY+nzZ. The anticommutation relations {σi,σj}=2δijI give
P±=21(I±σn^),σn^=P+−P−,P++P−=I
So σn^ is Hermitian with eigenvalues ±1 (traceless, so one of each), with spectral projectors
e−iθσn^/2=e−iθ/2P++eiθ/2P−
Apply f(λ)=e−iθλ/2:
e−iθσn^/2=cos2θI−isin2θσn^
Collect terms. This is Rn^(θ), the rotation by θ about n^ on the Bloch sphere. At θ=2π it equals −I.
What e^{-iHt} really means
For time-independent H=∑kEkPk the Schrödinger equationidtd∣ψ⟩=H∣ψ⟩ (with ℏ=1) is solved in one line:
∣ψ(t)⟩=U(t)∣ψ(0)⟩,U(t)=e−iHt=k∑e−iEktPk.
Each energy component only rotates its phase. Populations ∥Pkψ∥2 never change; relative phases advance at the Bohr frequencies Ek−El. Since e−iEkt=1, U(t) is unitary, with U(s)U(t)=U(s+t) and U(t)†=U(−t).
Conversely, every continuous one-parameter unitary group on Cd is e−iHt for a unique Hermitian H (the finite-dimensional case of Stone's theorem, treated in Beyond Finite Dimensions). Hermiticity of the generator is equivalent to conservation of probability.
DerivationDuhamel's formula and a perturbation bound
F(s)=e−iH(t−s)e−iH′s,F(0)=e−iHt,F(t)=e−iH′t
How much does a Hamiltonian error V=H′−H matter? Interpolate between the two evolutions:
F′(s)=e−iH(t−s)(iH−iH′)e−iH′s=−ie−iH(t−s)Ve−iH′s
Differentiate; each exponential commutes with its own generator:
e−iH′t−e−iHt=−i∫0te−iH(t−s)Ve−iH′sds
Integrate from 0 to t (Duhamel's formula):
e−iHt−e−iH′t∞≤t∥H−H′∥∞
Take operator norms; unitaries have norm 1. Errors in the generator grow at most linearly in time.
Predict before you look
A Hamiltonian is miscalibrated by V with ∥V∥∞=10−3 (ℏ=1). After evolving for t=200, what is the largest operator-norm error e−iHt−e−iH′t∞ the bound allows?
Show the answer
0.2
t∥V∥∞=200×10−3=0.2. The bound is nearly tight when V commutes with H: the error is then maxk1−e−ivkt=2∣sin(vkt/2)∣≈0.1997 for an eigenvalue vk=10−3.
Order, monotonicity and trace inequalities
Hermitian matrices are partially ordered: A≥B means A−B is positive semidefinite. Functional calculus interacts with this order in surprising ways. A function f is operator monotone if A≥B implies f(A)≥f(B) for Hermitian matrices of every size with spectra in its domain.
Think first Find the catch: does A≥B≥0 imply A2≥B2? Test A=(2111), B=(1000).
A−B=(1111)≥0 (eigenvalues 0,2) and B≥0. But A2−B2=(4332) has determinant 8−9=−1<0, hence a negative eigenvalue. Squaring is monotone on numbers, not on matrices: non-commutativity breaks it.
Trace functions behave better. The Golden–Thompson inequalityTreA+B≤Tr(eAeB) holds for all Hermitian A,B, while its naive three-matrix extension fails. Lieb's concavity theorem (1973), joint concavity of (A,B)↦Tr(K†ApKB1−p) on positive matrices for 0≤p≤1, led to Lieb and Ruskai's proof of strong subadditivity of von Neumann entropy, the backbone of quantum Shannon theory.
Key takeaways
Normal ⟺ unitarily diagonalizable: A=∑kλkPk with orthogonal spectral projectors. Hermitian, unitary and positive operators are the normal ones with real, unimodular and non-negative spectra.
Functional calculus f(A)=∑kf(λk)Pk defines e−iHt, ρ, logρ and ∣A∣; it is an algebra homomorphism with specf(A)=f(specA).
e−iθn^⋅σ/2=cos2θI−isin2θn^⋅σ, and a 2π rotation is −I.
Duhamel: e−iHt−e−iH′t≤t∥H−H′∥; first-order Trotter error is at most 2rt2∥[A,B]∥.
Matrix order is subtle: tp is operator monotone exactly for 0≤p≤1, while trace inequalities (Golden–Thompson, Lieb) power entropy theory.
Check yourself
In the interactive lesson these are scored with your confidence, and what you miss comes back for review.
Question 1
What is e−iπX/2 as a matrix?
X
−iX
iX
−I
Show the answer
B.−iX
cos2πI−isin2πX=−iX. As a gate it equals X up to the global phase −i, but as a matrix it is −iX, and the difference matters once the operation is controlled.
Question 2
Which operator is RX(2π)=e−iπX?
I
X
−iX
−I
Show the answer
D.−I
cosπI−isinπX=−I. A full turn returns every Bloch vector to itself but multiplies the state by −1, the hallmark of spin-½ (SU(2) double-covers SO(3)).
Question 3
Let A≥B≥0. Which conclusion is always valid?
A2≥B2
eA≥eB
A≥B
A−1≥B−1 (when both are invertible)
Show the answer
C.A≥B
t↦t is operator monotone (Löwner–Heinz with p=21). Squaring and exponentiating are increasing on numbers but not monotone on matrices.
Question 4
Which condition guaranteeseA+B=eAeB for square matrices A,B?
A and B are both Hermitian
[A,B]=0
Tr(AB)=Tr(BA)
A and B are both unitary
Show the answer
B.[A,B]=0
If A and B commute, the binomial theorem applies to (A+B)n and the series factor. Otherwise the BCH series adds 21[A,B]+… to the exponent.
Further reading
R. Bhatia, Matrix Analysis, Graduate Texts in Mathematics 169 (Springer, 1997).
R. A. Horn & C. R. Johnson, Matrix Analysis (Cambridge, 2nd ed., 2013).
E. H. Lieb & M. B. Ruskai, "Proof of the strong subadditivity of quantum-mechanical entropy," J. Math. Phys. (1973).
A. M. Childs, Y. Su, M. C. Tran, N. Wiebe & S. Zhu, "Theory of Trotter error with commutator scaling," Phys. Rev. X (2021).
A. Gilyén, Y. Su, G. H. Low & N. Wiebe, "Quantum singular value transformation and beyond: exponential improvements for quantum matrix arithmetics," STOC (2019).