2902 words
15 minutes
Lie Groups and Lie Algebras, Part II: The Lorentz Group

Prerequisites: Part I: Lie Groups and Lie Algebras, familiarity with index notation for vectors and matrices, and special relativity at the level of knowing what a Lorentz boost is.


In the previous post we built the general machinery of Lie groups, i.e. generators, structure constants, representations and the exponential map. Now we use all of it on the group behind all of relativistic physics, the Lorentz group.

In a sense, the Lorentz group is the answer to the question of which linear transformations preserve the structure of spacetime. Everything that follows in quantum field theory (spinors, the Dirac equation, gauge theories) is built on its representation theory. So it is worth going slowly here, even if some of it looks familiar from special relativity.


Definition: the orthogonal group O(n,m)O(n,m)#

Before going to spacetime, let me set up the general framework.

Take a space with coordinates (y1,…,ym,x1,…,xn)(y_1, \ldots, y_m, x_1, \ldots, x_n). The group of transformations that leaves the quadratic form

(y12+…+ym2)−(x12+…+xn2)(y_1^2 + \ldots + y_m^2) - (x_1^2 + \ldots + x_n^2)

invariant is called the orthogonal group O(n,m)O(n,m).

When all signs are the same (m=0m = 0), this is just the familiar rotation group O(n)O(n), which preserves x12+…+xn2x_1^2 + \ldots + x_n^2. The interesting physics happens when the signs are mixed, i.e. some coordinates enter with a plus and others with a minus. This indefinite signature is exactly what spacetime has.

Definition: the Lorentz group#

The Lorentz group is the group of linear coordinate transformations

xμ→x′μ=Λμν xνx^\mu \to x'^\mu = \Lambda^\mu{}_\nu \, x^\nu

that leave the following quantity invariant,

ημν xμxν=t2−x2−y2−z2\eta_{\mu\nu}\,x^\mu x^\nu = t^2 - x^2 - y^2 - z^2

which is the spacetime interval. We work in the mostly-minus convention η=diag(+1,−1,−1,−1)\eta = \text{diag}(+1, -1, -1, -1) throughout. Hence

Lorentz Group≡O(3,1)\boxed{\text{Lorentz Group} \equiv O(3,1)}

The "33" counts the spatial dimensions (which enter the quadratic form with a minus sign), and the "11" counts time (which enters with a plus). Whether one writes O(3,1)O(3,1) or O(1,3)O(1,3) is a convention, and some authors write it the other way. What matters is the signature, one plus and three minuses.


Lorentz invariance of the Minkowski metric#

Which condition does Λ\Lambda have to satisfy to be a Lorentz transformation? We require the spacetime interval to be invariant,

ημν x′μx′ν=ημν xμxν\eta_{\mu\nu}\,x'^\mu x'^\nu = \eta_{\mu\nu}\,x^\mu x^\nu

and substitute x′μ=Λμρ xρx'^\mu = \Lambda^\mu{}_\rho\, x^\rho,

ημν x′μx′ν=ημν(Λμρ xρ)(Λνσ xσ)=ηρσ xρxσ\eta_{\mu\nu}\,x'^\mu x'^\nu = \eta_{\mu\nu}(\Lambda^\mu{}_\rho\, x^\rho)(\Lambda^\nu{}_\sigma\, x^\sigma) = \eta_{\rho\sigma}\,x^\rho x^\sigma

where in the last step we relabelled the dummy indices so that the coefficients of xρxσx^\rho x^\sigma can be compared. Since this has to hold for any xμx^\mu, we can strip off the xx‘s,

ηρσ=ημν Λμρ Λνσ\boxed{\eta_{\rho\sigma} = \eta_{\mu\nu}\,\Lambda^\mu{}_\rho\,\Lambda^\nu{}_\sigma}

or, in matrix notation,

η=ΛTη Λ\eta = \Lambda^T \eta\, \Lambda

This is the defining equation of the Lorentz group. Compare it with the orthogonal group O(n)O(n), whose defining equation is 1=RTR\mathbf{1} = R^T R. The structure is the same, with the Minkowski metric η\eta in place of the Euclidean metric δ\delta.

A note on index gymnastics: The metric ημν\eta_{\mu\nu} is what raises and lowers indices, i.e. it converts contravariant (upper) indices into covariant (lower) ones and vice versa. In the equation above, Λμρ\Lambda^\mu{}_\rho can also change the indices on ημν\eta_{\mu\nu} itself, as the contraction on the right-hand side shows.


How the Lorentz group splits up#

The defining equation η=ΛTηΛ\eta = \Lambda^T \eta \Lambda constrains Λ\Lambda in a way that splits the Lorentz group into four disconnected components. Let us see how.

By determinant: proper vs. improper#

Taking the determinant of both sides of η=ΛTηΛ\eta = \Lambda^T \eta \Lambda,

det⁡(η)=det⁡(ΛT) det⁡(η) det⁡(Λ)\det(\eta) = \det(\Lambda^T)\,\det(\eta)\,\det(\Lambda)

and since det⁡(ΛT)=det⁡(Λ)\det(\Lambda^T) = \det(\Lambda) and det⁡(η)≠0\det(\eta) \neq 0, we can divide through,

1=(det⁡Λ)21 = (\det \Lambda)^2det⁡Λ=±1\boxed{\det \Lambda = \pm 1}

This gives two classes.

  • det⁡Λ=+1\det \Lambda = +1: Proper Lorentz transformations. These form the subgroup SO(3,1)SO(3,1), where the "SS" stands for special, i.e. unit determinant, as in SO(3)SO(3) for rotations.

  • det⁡Λ=−1\det \Lambda = -1: Improper Lorentz transformations. These include parity, time reversal and their combinations. They cannot be continuously connected to the identity (there is no smooth way from det⁡=+1\det = +1 to det⁡=−1\det = -1), so they cannot be reached by exponentiating generators. There is still a continuum of them (parity followed by any rotation is improper), but each one is a discrete operation such as PP composed with an ordinary proper transformation.

By the 00-component: orthochronous vs. non-orthochronous#

Next we look at the ρ=σ=0\rho = \sigma = 0 component of ηρσ=ημνΛμρΛνσ\eta_{\rho\sigma} = \eta_{\mu\nu}\Lambda^\mu{}_\rho\Lambda^\nu{}_\sigma,

1=(Λ00)2−∑i=13(Λi0)21 = (\Lambda^0{}_0)^2 - \sum_{i=1}^{3}(\Lambda^i{}_0)^2

which gives

(Λ00)2=1+∑i=13(Λi0)2≥1(\Lambda^0{}_0)^2 = 1 + \sum_{i=1}^{3}(\Lambda^i{}_0)^2 \geq 1

Since (Λ00)2≥1(\Lambda^0{}_0)^2 \geq 1, either Λ00≥1\Lambda^0{}_0 \geq 1 or Λ00≤−1\Lambda^0{}_0 \leq -1, with nothing in between. This is again a discrete split.

  • Λ00≥1\Lambda^0{}_0 \geq 1: Orthochronous, the transformation preserves the direction of time.
  • Λ00≤−1\Lambda^0{}_0 \leq -1: Non-orthochronous, the transformation reverses the direction of time.

The four components#

Combining the two binary choices, O(3,1)O(3,1) splits into four disconnected components,

Componentdet⁡Λ\det\LambdaΛ00\Lambda^0{}_0ContainsExample
L+↑\mathcal{L}^\uparrow_++1+1≥1\geq 1IdentityRotations, boosts
L+↓\mathcal{L}^\downarrow_++1+1≤−1\leq -1PTPTCombined parity + time reversal
L−↑\mathcal{L}^\uparrow_-−1-1≥1\geq 1PPParity (spatial inversion)
L−↓\mathcal{L}^\downarrow_-−1-1≤−1\leq -1TTTime reversal

Only L+↑\mathcal{L}^\uparrow_+, the proper orthochronous Lorentz group, is connected to the identity. This is the component that has a Lie algebra, and when physicists say “the Lorentz group” without further qualification they almost always mean it,

Orthochronous Proper Lorentz Transformations≡SO+(3,1)\boxed{\text{Orthochronous Proper Lorentz Transformations} \equiv SO^+(3,1)}

The other three components are obtained by applying the discrete transformations PP, TT or PTPT to elements of SO+(3,1)SO^+(3,1).

Non-orthochronous transformations#

When Λ00≤−1\Lambda^0{}_0 \leq -1, the transformation reverses the direction of time. Any non-orthochronous transformation can be written as an orthochronous one composed with a discrete inversion, and the relevant discrete operations are

  • time reversal TT: (t,x,y,z)→(−t,x,y,z)(t, x, y, z) \to (-t, x, y, z), which has det⁡Λ=−1\det\Lambda = -1 and Λ00=−1\Lambda^0{}_0 = -1,
  • the combination PTPT: (t,x,y,z)→(−t,−x,−y,−z)(t, x, y, z) \to (-t, -x, -y, -z), which has det⁡Λ=+1\det\Lambda = +1 and Λ00=−1\Lambda^0{}_0 = -1.

Improper Lorentz transformations#

Transformations with det⁡Λ=−1\det\Lambda = -1 are improper. Any improper transformation can be written as a proper one (det⁡=+1\det = +1) composed with a discrete transformation. Examples are

  • Parity: (t,x,y,z)→(t,−x,−y,−z)(t,x,y,z) \to (t, -x, -y, -z), which flips all spatial coordinates, det⁡=−1\det = -1, orthochronous,
  • Single-axis reflection: (t,x,y,z)→(t,−x,y,z)(t,x,y,z) \to (t, -x, y, z), which flips one spatial axis, det⁡=−1\det = -1, orthochronous,
  • Time reversal: (t,x,y,z)→(−t,x,y,z)(t,x,y,z) \to (-t, x, y, z), which flips time, det⁡=−1\det = -1, non-orthochronous.

So parity is improper but orthochronous (Λ00=+1\Lambda^0{}_0 = +1), while time reversal is both improper and non-orthochronous. These are genuinely different operations and live in different disconnected components of the Lorentz group.


Lorentz group representations#

Now we apply the representation theory from Part I. Recall that a set of objects ϕi\phi^i (with i=1,…,ni = 1, \ldots, n) transforms in an nn-dimensional representation RR of the Lorentz group if, under a Lorentz transformation,

ϕi→Λij ϕj=[exp⁡(−i2 ωμνJRμν)]ijϕj\phi^i \to \Lambda^i{}_j\,\phi^j = \left[\exp\left(-\frac{i}{2}\,\omega_{\mu\nu}J_R^{\mu\nu}\right)\right]^i{}_j \phi^j

Here Λ=exp⁡(−i2 ωμνJμν)\Lambda = \exp(-\tfrac{i}{2}\,\omega_{\mu\nu}J^{\mu\nu}) is a matrix representation of the abstract group element. The JRμνJ_R^{\mu\nu} are the Lorentz generators in the representation RR, which are n×nn \times n matrices. The parameters ωμν\omega_{\mu\nu} are antisymmetric (ωμν=−ωνμ\omega_{\mu\nu} = -\omega_{\nu\mu}), and an antisymmetric 4×44\times4 matrix has 4×32=6\frac{4\times3}{2} = 6 independent components, corresponding to three rotations and three boosts.

For infinitesimal transformations (ωμν\omega_{\mu\nu} small) we expand the exponential to first order,

δϕi=−i2 ωμν (JRμν)ij ϕj\delta\phi^i = -\frac{i}{2}\,\omega_{\mu\nu}\,(J_R^{\mu\nu})^i{}_j\,\phi^j

The pair (μ,ν)(\mu, \nu) labels which generator we mean (which rotation or boost), while (i,j)(i, j) are the matrix indices of that generator in the representation RR. The explicit form of (JRμν)ij(J_R^{\mu\nu})^i{}_j as an n×nn\times n matrix depends on the representation.

Let us go through the representations one at a time.


Scalar representation#

For a scalar ϕ\phi, the index ii takes only one value (i=1i = 1), so this is a 1-dimensional representation, and each generator (Jμν)ij(J^{\mu\nu})^i{}_j is a 1×11\times1 matrix, i.e. a single number for each pair (μ,ν)(\mu,\nu).

A scalar field is invariant under Lorentz transformations,

ϕ→Λϕ=1⋅ϕ=ϕ\phi \to \Lambda\phi = 1 \cdot \phi = \phi

(Invariant means the value does not change. In contrast, covariant means it transforms in a well-defined way under certain rules. All representations are covariant, but only the scalar is invariant.)

Since Λ=e0=1\Lambda = e^0 = 1 in this representation,

δϕ=0,Jμν=0\delta\phi = 0, \qquad J^{\mu\nu} = 0

A representation in which all generators vanish is a valid solution of the Lie algebra [Ta,Tb]=ifabcTc[T^a, T^b] = if^{ab}{}_c T^c (both sides are trivially zero), and it is called the trivial representation.

A typical Lorentz scalar in particle physics is the rest mass of a particle, on which all observers agree regardless of their frame.


Vector representation#

A contravariant 4-vector VμV^\mu transforms as

Vμ→Λμν VνV^\mu \to \Lambda^\mu{}_\nu\,V^\nu

and a covariant 4-vector VμV_\mu as

Vμ→Λμν VνV_\mu \to \Lambda_\mu{}^\nu\,V_\nu

with Λ\Lambda satisfying the Lorentz condition η=ΛTηΛ\eta = \Lambda^T\eta\Lambda. The spacetime coordinates xμx^\mu and the four-momentum pμ=(E,p)p^\mu = (E, \mathbf{p}) are the most important examples of contravariant 4-vectors.

This is a 4-dimensional representation, so each generator JμνJ^{\mu\nu} is a 4×44\times4 matrix, (Jμν)ρσ(J^{\mu\nu})^\rho{}_\sigma. Explicitly,

(Jμν)ρσ=i(ημρδσν−ηνρδσμ)\boxed{(J^{\mu\nu})^\rho{}_\sigma = i\left(\eta^{\mu\rho}\delta^\nu_\sigma - \eta^{\nu\rho}\delta^\mu_\sigma\right)}

This is antisymmetric in μν\mu\nu, as it must be since Jμν=−JνμJ^{\mu\nu} = -J^{\nu\mu}, and it is the unique generator consistent with the infinitesimal form of V′μ=ΛμνVνV'^\mu = \Lambda^\mu{}_\nu V^\nu.

To see this, take an infinitesimal Lorentz transformation Λμν=δνμ+ωμν\Lambda^\mu{}_\nu = \delta^\mu_\nu + \omega^\mu{}_\nu,

ΛμνVν=(δνμ+ωμν)Vν=Vμ+ωμνVν≡Vμ+δVμ\Lambda^\mu{}_\nu V^\nu = (\delta^\mu_\nu + \omega^\mu{}_\nu)V^\nu = V^\mu + \omega^\mu{}_\nu V^\nu \equiv V^\mu + \delta V^\mu

Comparing with the general formula δVρ=−i2ωμν(Jμν)ρσVσ\delta V^\rho = -\frac{i}{2}\omega_{\mu\nu}(J^{\mu\nu})^\rho{}_\sigma V^\sigma and putting in the explicit generator, one can check that the two agree, δVρ=ωρσVσ\delta V^\rho = \omega^\rho{}_\sigma V^\sigma.


Lorentz transformations of 4-vectors#

Now the explicit transformations. A general Lorentz transformation depends on six parameters,

x′μ=Λ(η,θ)μν xνx'^\mu = \Lambda(\boldsymbol{\eta}, \boldsymbol{\theta})^\mu{}_\nu\,x^\nu

where η=(ηx,ηy,ηz)\boldsymbol{\eta} = (\eta^x, \eta^y, \eta^z) are the three rapidity (boost) parameters and θ=(θx,θy,θz)\boldsymbol{\theta} = (\theta^x, \theta^y, \theta^z) are the three rotation angles.

Boosts#

A boost along the xx-axis can be written either with the velocity or with the rapidity. With the velocity βi=vi\beta^i = v^i (natural units, c=1c = 1) and γi=(1−(βi)2)−1/2\gamma^i = (1 - (\beta^i)^2)^{-1/2},

t→γx(t+βxx),x→γx(x+βxt)t \to \gamma^x(t + \beta^x x), \qquad x \to \gamma^x(x + \beta^x t)

Since −1<β<1-1 < \beta < 1, we can write βi=tanh⁡ηi\beta^i = \tanh\eta^i with −∞<ηi<∞-\infty < \eta^i < \infty the rapidity, and the same transformation becomes

t→(cosh⁡ηx) t+(sinh⁡ηx) x,x→(sinh⁡ηx) t+(cosh⁡ηx) xt \to (\cosh\eta^x)\,t + (\sinh\eta^x)\,x, \qquad x \to (\sinh\eta^x)\,t + (\cosh\eta^x)\,x

The rapidity form is nicer for a couple of reasons. Rapidities add when collinear boosts are composed (velocities do not), and the hyperbolic functions make the analogy with rotations obvious, since the cos⁡θ,sin⁡θ\cos\theta, \sin\theta of rotations become cosh⁡η,sinh⁡η\cosh\eta, \sinh\eta for boosts.

Boosts in the yy and zz directions work the same way, with the cosh⁡\cosh and sinh⁡\sinh in the appropriate row and column.

Boost matrices#

The explicit 4×44\times4 matrices for boosts along each axis are as follows.

Boost along xx (β\beta-form and η\eta-form):

Λ(βx)=(γxβxγx00βxγxγx0000100001),Λ(ηx)=(cosh⁡ηxsinh⁡ηx00sinh⁡ηxcosh⁡ηx0000100001)\Lambda(\beta^x) = \begin{pmatrix} \gamma^x & \beta^x\gamma^x & 0 & 0 \\ \beta^x\gamma^x & \gamma^x & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix}, \qquad \Lambda(\eta^x) = \begin{pmatrix} \cosh\eta^x & \sinh\eta^x & 0 & 0 \\ \sinh\eta^x & \cosh\eta^x & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix}

Boost along yy:

Λ(βy)=(γy0βyγy00100βyγy0γy00001),Λ(ηy)=(cosh⁡ηy0sinh⁡ηy00100sinh⁡ηy0cosh⁡ηy00001)\Lambda(\beta^y) = \begin{pmatrix} \gamma^y & 0 & \beta^y\gamma^y & 0 \\ 0 & 1 & 0 & 0 \\ \beta^y\gamma^y & 0 & \gamma^y & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix}, \qquad \Lambda(\eta^y) = \begin{pmatrix} \cosh\eta^y & 0 & \sinh\eta^y & 0 \\ 0 & 1 & 0 & 0 \\ \sinh\eta^y & 0 & \cosh\eta^y & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix}

Boost along zz:

Λ(βz)=(γz00βzγz01000010βzγz00γz),Λ(ηz)=(cosh⁡ηz00sinh⁡ηz01000010sinh⁡ηz00cosh⁡ηz)\Lambda(\beta^z) = \begin{pmatrix} \gamma^z & 0 & 0 & \beta^z\gamma^z \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ \beta^z\gamma^z & 0 & 0 & \gamma^z \end{pmatrix}, \qquad \Lambda(\eta^z) = \begin{pmatrix} \cosh\eta^z & 0 & 0 & \sinh\eta^z \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ \sinh\eta^z & 0 & 0 & \cosh\eta^z \end{pmatrix}

The pattern is that a boost mixes the time component (row/column 0) with the spatial component along the boost, and leaves the other two spatial components alone. The cosh⁡\cosh sits on the diagonal and the sinh⁡\sinh off the diagonal, in the same positions where cos⁡\cos and sin⁡\sin sit for rotations.

Rotation matrices#

Rotations do not touch the time component at all and act only on the spatial 3×33\times3 block.

Rotation about xx-axis:

Λ(θx)=(1000010000cos⁡θxsin⁡θx00−sin⁡θxcos⁡θx)\Lambda(\theta^x) = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & \cos\theta^x & \sin\theta^x \\ 0 & 0 & -\sin\theta^x & \cos\theta^x \end{pmatrix}

Rotation about yy-axis:

Λ(θy)=(10000cos⁡θy0−sin⁡θy00100sin⁡θy0cos⁡θy)\Lambda(\theta^y) = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & \cos\theta^y & 0 & -\sin\theta^y \\ 0 & 0 & 1 & 0 \\ 0 & \sin\theta^y & 0 & \cos\theta^y \end{pmatrix}

Rotation about zz-axis:

Λ(θz)=(10000cos⁡θzsin⁡θz00−sin⁡θzcos⁡θz00001)\Lambda(\theta^z) = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & \cos\theta^z & \sin\theta^z & 0 \\ 0 & -\sin\theta^z & \cos\theta^z & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix}

These are the familiar 3×33\times3 rotation matrices placed in the lower-right 3×33\times3 block of a 4×44\times4 matrix, with the time-time component equal to 1 and all time-space components equal to 0.


Rotation and boost generators#

We now extract the generators by differentiating the finite transformations and evaluating at the identity (all parameters zero).

Boost generators#

Ki=−i∂Λ(ηi)∂ηi∣ηi=0K^i = -i\frac{\partial\Lambda(\eta^i)}{\partial\eta^i}\bigg|_{\eta^i=0}

Using cosh⁡(0)=1\cosh(0) = 1 and sinh⁡(0)=0\sinh(0) = 0,

Kx=−i(0100100000000000),Ky=−i(0010000010000000),Kz=−i(0001000000001000)K^x = -i\begin{pmatrix} 0 & 1 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{pmatrix}, \quad K^y = -i\begin{pmatrix} 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{pmatrix}, \quad K^z = -i\begin{pmatrix} 0 & 0 & 0 & 1 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{pmatrix}

The KiK^i are symmetric matrices (the 4×44\times4 part in front of which the −i-i sits). This reflects the fact that boosts are not unitary transformations. The Lorentz group is non-compact, and boosts move you along a hyperbola instead of around a circle.

Rotation generators#

Ji=−i∂Λ(θi)∂θi∣θi=0J^i = -i\frac{\partial\Lambda(\theta^i)}{\partial\theta^i}\bigg|_{\theta^i=0}

Using cos⁡(0)=1\cos(0) = 1 and sin⁡(0)=0\sin(0) = 0,

Jx=i(00000000000−10010),Jy=i(0000000100000−100),Jz=i(000000−1001000000)J^x = i\begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & -1 \\ 0 & 0 & 1 & 0 \end{pmatrix}, \quad J^y = i\begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 \\ 0 & 0 & 0 & 0 \\ 0 & -1 & 0 & 0 \end{pmatrix}, \quad J^z = i\begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & -1 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{pmatrix}

The JiJ^i are antisymmetric (again the part after the ii), which reflects that rotations are unitary, since the rotation group SO(3)SO(3) is compact.

The Lorentz algebra#

The six generators satisfy the commutation relations

[Ji,Jj]=iϵijkJk[J^i, J^j] = i\epsilon^{ijk}J^k[Ji,Kj]=iϵijkKk[J^i, K^j] = i\epsilon^{ijk}K^k[Ki,Kj]=−iϵijkJk[K^i, K^j] = -i\epsilon^{ijk}J^k

and each one has a clear physical meaning.

  • [Ji,Jj]=iϵijkJk[J^i, J^j] = i\epsilon^{ijk}J^k: The rotation generators close among themselves and satisfy the su(2)\mathfrak{su}(2) algebra. This just says that the JiJ^i are angular momenta and that rotations form a subgroup.

  • [Ji,Kj]=iϵijkKk[J^i, K^j] = i\epsilon^{ijk}K^k: The boosts transform as a vector under rotations. If you rotate the coordinate system, the boost generators rotate with it, which is what one would expect. A boost “in the xx-direction” should become a boost “in the yy-direction” under a 90°90° rotation about zz.

  • [Ki,Kj]=−iϵijkJk[K^i, K^j] = -i\epsilon^{ijk}J^k: This is the most important one. The commutator of two boosts is a rotation, not another boost, so boosts do not form a subgroup. (The rotation obtained this way shows up physically as Thomas precession.) The minus sign, compared to [Ji,Jj][J^i, J^j], matters a lot. It is the reason the Lorentz group is non-compact, and the reason it has no finite-dimensional unitary representations apart from the trivial one. In Part III it will also be the reason left- and right-handed spinors are complex conjugates of each other.

If the third relation had a plus sign instead, [Ki,Kj]=+iϵijkJk[K^i, K^j] = +i\epsilon^{ijk}J^k, then (Ji,Ki)(J^i, K^i) would generate SO(4)SO(4), a compact group with perfectly well-behaved finite-dimensional unitary representations. That one minus sign changes everything.

Connecting to the tensor notation#

The six generators JiJ^i and KiK^i can be packaged into the antisymmetric tensor JμνJ^{\mu\nu} via

Ji=12ϵijkJjk,Ki=J0iJ^i = \frac{1}{2}\epsilon^{ijk}J^{jk}, \qquad K^i = J^{0i}
A sign caveat

If you plug the boxed vector-representation formula into this, you get J0iJ^{0i} equal to minus the KiK^i matrices we computed above, while the rotation generators agree. The mismatch comes from the sign convention for the rapidity. Differentiating Λ(η)\Lambda(\eta) as we did corresponds to writing Λ=e+iηK\Lambda = e^{+i\eta K}, whereas Λ=exp⁡(−i2ωμνJμν)\Lambda = \exp(-\tfrac{i}{2}\omega_{\mu\nu}J^{\mu\nu}) has the opposite sign in the exponent. This is harmless for the algebra, since all three commutation relations are unchanged under Ki→−KiK^i \to -K^i. From here on we use the standard identification Ki=J0iK^i = J^{0i} (as in Peskin & Schroeder).

The tensor form is useful because the Lorentz transformation can then be written compactly as Λ=exp⁡(−i2ωμνJμν)\Lambda = \exp(-\tfrac{i}{2}\omega_{\mu\nu}J^{\mu\nu}), where the six independent components of ωμν\omega_{\mu\nu} are the three rotation angles and the three boost rapidities.


Connection to what comes next#

We now have the complete Lorentz algebra, i.e. six generators, their explicit 4×44\times4 matrices and their commutation relations. The 4×44\times4 (vector) representation is only one possibility, though. The Lie algebra has infinitely many representations, and the physically most important ones are not the vector representation.

In the next post we complexify the Lorentz algebra by defining Ni±=12(Ji±iKi)N_i^\pm = \frac{1}{2}(J_i \pm iK_i), which splits it into su(2)⊕su(2)\mathfrak{su}(2) \oplus \mathfrak{su}(2). This gives the spinor representations, the (1/2,0)(1/2, 0) and (0,1/2)(0, 1/2) representations that describe left- and right-handed Weyl fermions. From there we construct Dirac and Majorana spinors, and eventually get to the Poincaré group and the Wigner classification of particles by mass and spin.

The minus sign in [Ki,Kj]=−iϵijkJk[K^i, K^j] = -i\epsilon^{ijk}J^k will be behind all of it.


Next post: Part III: Representations of the Lorentz Group


This post is based on my own self-study notes that I created in 2022 in order to get a deeper understanding of all of this.

Lie Groups and Lie Algebras, Part II: The Lorentz Group
https://rohankulkarni.me/posts/notes/lorentz-group/
Author
Rohan Kulkarni
Published at
2026-04-04
License
CC BY-NC-SA 4.0
Share Post
Comments