Home / Learn / Exam prep / GATE Physics / Lesson 1.2
All lessons

Module 1: Mathematical physics · Lesson 1.2 · about 55 min

Hermitian and unitary matrices, diagonalisation

Hermitian, anti-Hermitian, unitary and normal matrices and where their eigenvalues live, unitary diagonalisation and the spectral theorem, commuting matrices, functions of matrices from Pauli exponentials to Jordan blocks, Cayley-Hamilton, and the trace and determinant tricks that save minutes in GATE.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 1 · about 55 minutes. By the end you will be able to say at a glance where the eigenvalues of a special matrix must lie, diagonalise Hermitian and normal matrices with a unitary change of basis, decide when two matrices can be diagonalised together, compute functions of matrices (exponentials, square roots, powers) without long multiplication, and use the Cayley-Hamilton theorem and trace and determinant tricks under exam time pressure. The basics (characteristic equation, 2 × 2 eigenvalues, inverse, rank) are in the IIT JAM lesson on matrices and eigenvalues; here we go one level deeper. The language of the previous lesson on vector spaces and Dirac notation is used throughout.

Every Hermitian or unitary matrix is a diagonal matrix seen in a rotated basis. Find that basis, and powers, exponentials, square roots and inverses all act on the eigenvalues one at a time.

1. The family and where its eigenvalues live

For a complex square matrix, A† is the conjugate transpose. Four kinds matter in physics:

Kind Definition Eigenvalues
Hermitian H†=H real
anti-Hermitian A†=−A purely imaginary (or 0)
unitary U†U=I modulus 1 (on the unit circle)
normal AA†=A†A anywhere

All three special kinds are normal, and normality is exactly the condition for being diagonalised by a unitary matrix. The eigenvalue facts take one line each:

  • Hermitian: λ⟨v|v⟩=⟨v|Hv⟩=⟨Hv|v⟩=λ∗⟨v|v⟩, so λ is real.
  • Anti-Hermitian: iA is Hermitian, so λ=−i×(real) is imaginary. Real antisymmetric matrices are the real case: eigenvalues 0,±iω, determinant ≥0, and trA2=−∑Aij2≤0.
  • Unitary: ‖Uv‖=|λ|‖v‖=‖v‖, so |λ|=1. A real orthogonal 2×2 with determinant −1 is a reflection with eigenvalues +1 and −1.

Careful with complex matrices: symmetric is not Hermitian. A=(1ii1) is symmetric, but A†≠A. In fact AA†=2I, so A/2 is unitary, with eigenvalues (1±i)/2 on the unit circle.

Hermitian H
an observable: its eigenvalues are the possible results of a measurement
Unitary U
a change of basis or a time evolution that keeps all probabilities and overlaps
U†HU
the same observable written in another orthonormal basis: same eigenvalues, different matrix
[A, B] = 0
A and B can be measured together: they have a common eigenbasis
Physical meaning
Diagonalising H means finding the states with definite values (energies, spins, normal-mode frequencies); every other state is a superposition of these.

2. Unitary diagonalisation and the spectral theorem

If A is normal, its normalised eigenvectors are orthonormal and can be put as the columns of a unitary U:

A=UDU†,D=diag(λ1,…,λn),A=∑kλkPk  with Pk=|k⟩⟨k|

The second form, the spectral decomposition, is the most useful: the projectors Pk satisfy PjPk=δjkPk and ∑Pk=I, so any function acts eigenvalue by eigenvalue:

f(A)=∑kf(λk)Pk

Worked example. σx has eigenvectors (1,1)/2 and (1,−1)/2, so the Hadamard matrix U=12(111−1) gives U†σxU=σz. Now take H=2P1+5P2 with P1=12(1111) and P2=12(1−1−11). The positive square root is H=2P1+5P2, whose top-right entry is 12(2−5)≈−0.411. Taking square roots of the entries of H would be nonsense.

Quadratic forms and principal axes. A real symmetric A is diagonalised by a rotation, so the curve x→TAx→=1 becomes λ1X2+λ2Y2=1 in the eigenvector axes: an ellipse if both eigenvalues are positive, a hyperbola if they have opposite signs, empty if both are negative. The off-diagonal entry tilts the axes and decides the shape: x2+4xy+y2=1 has eigenvalues 3 and −1, a hyperbola.

The curve xᵀAx = 1 for a symmetric 2 × 2 matrix, with its two perpendicular eigenvector axes. Slide the off-diagonal entry b: the axes turn, and when det A changes sign the ellipse breaks into a hyperbola.

The same idea solves small oscillations with unequal masses: Mx→¨=−Kx→ with trial solution a→cos⁡ωt gives the generalised eigenvalue problem

det⁡(K−ω2M)=0

For M=diag(1,2) and K=(3−1−12) this is 2ω4−8ω2+5=0, so ω2=2±6/2. Diagonalising K alone would ignore the masses. For a refresher on normal modes with equal masses, see the JAM lesson on superposition and coupled oscillators.

3. Commuting matrices

Theorem. Two diagonalisable matrices can be diagonalised by the same similarity transformation exactly when they commute. For Hermitian matrices the common transformation is unitary.

One direction is immediate: diagonal matrices commute. For the other, if AB=BA and Av→=av→, then A(Bv→)=a(Bv→), so B maps each eigenspace of A into itself.

  • If every eigenvalue of A is non-degenerate, each eigenspace is a line, so every eigenvector of A is an eigenvector of B: B is diagonal in A’s eigenbasis (in fact a polynomial in A).
  • If A is degenerate, B only has to be block-diagonal: inside a degenerate eigenspace you must diagonalise B separately. That is how quantum mechanics labels states with several quantum numbers (energy, then L2, then Lz).

Quick test. A=(2112) has distinct eigenvalues with eigenvectors (1,±1), so the matrices that commute with it are exactly (abba); (3−1−13) qualifies and (1002) does not.

4. Functions of matrices

Define f(A) by the power series of f. Four methods cover every GATE question:

  1. Diagonalise: f(A)=Uf(D)U†, eigenvalue by eigenvalue. Eigenvalues of f(A) are f(λk), so trf(A)=∑f(λk) and det⁡f(A)=∏f(λk). For A=(0ln⁡2ln⁡20), treA=eln⁡2+e−ln⁡2=2.5.

  2. A matrix that squares to itself or to the identity. If (nˆ⋅σ→)2=I (any unit vector nˆ), the series splits into even and odd parts:

    eiθnˆ⋅σ→=Icos⁡θ+i(nˆ⋅σ→)sin⁡θ

    So exp⁡(iπσx/2)=iσx and treiθnˆ⋅σ→=2cos⁡θ. If A2=cA (as for any rank-one matrix with trace c), then An=cn−1A and eA=I+ec−1cA.

  3. Jordan blocks (not diagonalisable). For A=λI+N with N2=0, Taylor expansion stops after one term: f(A)=f(λ)I+f′(λ)N. So (1101)=(11/201) and etA=eλt(I+tN).

  4. Cayley-Hamilton reduces any power series to a polynomial of degree n−1 (next section).

Two identities to know by heart: det⁡eA=etrA for every square matrix, and eiH is unitary whenever H is Hermitian (which is why time evolution e−iHt/ℏ conserves probability). But eAeB≠eA+B unless A and B commute. When [A,B] commutes with both, eAeB=eA+B+12[A,B] (the simplest Baker-Campbell-Hausdorff formula, used for xˆ and pˆ).

5. Cayley-Hamilton at GATE level

Every square matrix satisfies its own characteristic equation p(λ)=det⁡(λI−A)=0. Uses beyond the 2×2 reduction you saw in JAM:

  • Inverses without row reduction. If p(λ)=λ3−6λ2+11λ−6, then A3−6A2+11A−6I=0, so A−1=16(A2−6A+11I).
  • High powers. Divide λN by p(λ); only the remainder (degree below n) survives. For 2×2 with distinct eigenvalues, AN=αA+βI where λiN=αλi+β for both eigenvalues. The Fibonacci matrix (1110) gives An=(Fn+1FnFnFn−1), with Fn=(φn−ψn)/5 from its eigenvalues (1±5)/2.
  • Adjugate. adjA=(det⁡A)A−1 has eigenvalues det⁡A/λk; for 3×3, tradjA=λ1λ2+λ2λ3+λ3λ1.

6. Trace and determinant tricks

Worked example (rank one). Find det⁡(I+A) for A=(111222333).

  1. Every row is a multiple of (1,1,1): A=u→v→T with u→=(1,2,3)T, v→=(1,1,1)T.
  2. Au→=u→(v→Tu→)=6u→, and anything orthogonal to v→ is sent to zero: eigenvalues 6,0,0.
  3. I+A has eigenvalues 7,1,1, so det⁡(I+A)=7=1+v→Tu→.

More shortcuts that GATE rewards:

  • Spectrum fixed by an equation. H2=I forces eigenvalues ±1; the trace then counts them. A 4×4 Hermitian H with H2=I and trace 2 has eigenvalues 1,1,1,−1 and det⁡H=−1. A real 4×4 with A2=−I has eigenvalues i,i,−i,−i (complex roots of a real polynomial pair up) and det⁡A=1.
  • Equality in the triangle inequality. A 3×3 unitary with trace 3 has three unit-modulus eigenvalues adding to 3, so all equal 1; being normal it is diagonalisable, hence U=I.
  • Circulants and permutations. A matrix whose rows are cyclic shifts has eigenvectors built from roots of unity; the row sum is always an eigenvalue (eigenvector (1,1,…,1)). The cyclic permutation of three objects has eigenvalues 1,ω,ω2.
  • Kronecker products. A⊗B has eigenvalues aibj: tr(A⊗B)=trAtrB and det⁡(A⊗B)=(det⁡A)m(det⁡B)n for A of size n and B of size m.
  • Positive definite Hermitian matrices: all eigenvalues positive, equivalently v†Hv>0 for all v≠0, equivalently H=B†B with B invertible. Positive diagonal entries are necessary but not sufficient.
  • Nilpotent: if N2=0 then the image of N lies in its kernel, so rankN≤n/2.

“If all the eigenvalues of a matrix equal 1, the matrix is the identity.” Only if the matrix is diagonalisable. The shear (1101) has eigenvalue 1 twice but only one eigenvector, and it is certainly not I. For normal matrices (Hermitian, unitary, anti-Hermitian) the conclusion is safe, because they are always diagonalisable: U=VIV†=I. Before reading off a matrix from its eigenvalues, ask whether it is normal.

Check your diagonalisations in fibub’s matrix calculator, and watch a matrix act on vectors with the matrix transform tool.

  • Hermitian: real eigenvalues. Anti-Hermitian: imaginary. Unitary: modulus 1. All three are normal, and normal matrices are exactly the unitarily diagonalisable ones. Complex symmetric is not Hermitian.
  • Spectral decomposition A=∑λkPk gives f(A)=∑f(λk)Pk; trf(A)=∑f(λk), det⁡f(A)=∏f(λk), det⁡eA=etrA.
  • eiθnˆ⋅σ→=Icos⁡θ+inˆ⋅σ→sin⁡θ; for a Jordan block f(λI+N)=f(λ)I+f′(λ)N; eAeB=eA+B only if they commute.
  • Commuting diagonalisable matrices share an eigenbasis; with degeneracy, diagonalise the second matrix inside each eigenspace.
  • Small oscillations: det⁡(K−ω2M)=0. Quadratic forms: the signs of the eigenvalues decide ellipse or hyperbola.
  • Cayley-Hamilton gives inverses and high powers; tradjA is the sum of pairwise eigenvalue products; det⁡(I+u→v→T)=1+v→Tu→.

Practise: exam questions on this lesson

From warm-up to top-rank level, with full solutions. For the BSc-level basics, the JAM matrices questions are a good warm-up. More on the GATE practice page for this topic.

Before this: lesson 1.1, linear vector spaces and Dirac notation. Next: lesson 1.3, analytic functions and Cauchy’s theorem. Back to the GATE Physics course home.

Going on to CSIR NET? Group theory for physicists shows how unitary matrices like these form groups such as SU(2).

Check yourself

1Where do the eigenvalues of a unitary matrix lie in the complex plane?
2When can two Hermitian matrices be diagonalised by the same unitary matrix?
3For A = [[1, 1], [0, 1]], what is the top-right entry of the square root of A that has eigenvalues +1?

Sources

  • G. Strang, Introduction to Linear Algebra, 5th ed. (2016), chapters 6 and 7
  • R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. (2013), chapters 1, 2 and 4
  • G. B. Arfken, H. J. Weber and F. E. Harris, Mathematical Methods for Physicists, 7th ed. (2013), chapter 6
  • H. Goldstein, C. Poole and J. Safko, Classical Mechanics, 3rd ed. (2002), chapter 6 (small oscillations)
  • GATE Physics (PH) syllabus, Mathematical Physics section, published by the organising institute each year

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Not affiliated with the IITs, IISc or the GATE organising institute. The general aptitude part of GATE is not covered here. Check the official GATE website for the current pattern. All questions are original.