Principal Axes and Principal Moments of Inertia

Here is a claim that sounds too good to be true. Take any rigid object — a brick, a wrench, a potato, an asteroid shaped like nothing at all — and there exist exactly three mutually perpendicular directions through its center of mass about which it will spin perfectly smoothly, angular momentum locked parallel to angular velocity, no wobble, no bearing forces, no tumble. Not "objects with nice symmetry." Any object. The potato has no symmetry whatsoever; the three directions are hiding in it anyway. The guarantee does not come from geometry — it comes from linear algebra. The moment of inertia tensor that maps spin to angular momentum, $\mathbf{L}=\mathbf{I}\boldsymbol{\omega}$ , is a real symmetric matrix, and the spectral theorem promises every such matrix a complete set of real eigenvalues and orthogonal eigenvectors. Those eigenvectors are the principal axes; the eigenvalues are the principal moments of inertia $I_1, I_2, I_3$ . Finding them is the single most consequential diagonalization in classical mechanics: in the principal frame the products of inertia vanish, angular momentum and kinetic energy fall apart into three independent pieces, and the equations of rigid-body motion become tractable at all. This page does the eigenvalue problem honestly: the characteristic cubic and its frame-independent invariants, the full proof that the eigenvalues are real and the axes orthogonal, what freedom degeneracy leaves you, and the symmetry shortcuts that in practice hand you the axes without solving anything. The pattern of the three eigenvalues then sorts every rigid body in the universe into exactly three families — spherical, symmetric, and asymmetric tops — a classification you will meet again from molecular spectra to tumbling spacecraft.

Key takeaways

The inertia tensor $\mathbf{I}$ is real and symmetric, so the spectral theorem guarantees three real eigenvalues (the principal moments $I_1,I_2,I_3$ ) and an orthonormal triad of eigenvectors (the principal axes) — for every rigid body, symmetric or not.
Spin about a principal axis and $\mathbf{L}=I_a\boldsymbol{\omega}$ with an honest scalar: no wobble, no bearing torque. About any other axis, $\mathbf{L}$ tilts away from $\boldsymbol{\omega}$ and precesses.
In the principal frame the tensor is diagonal: $\mathbf{L}=(I_1\omega_1,I_2\omega_2,I_3\omega_3)$ and $T=\tfrac12(I_1\omega_1^2+I_2\omega_2^2+I_3\omega_3^2)$ — three decoupled directions.
The characteristic cubic's coefficients $\mathrm{tr}\,\mathbf{I}$ , $\tfrac12[(\mathrm{tr}\,\mathbf{I})^2-\mathrm{tr}(\mathbf{I}^2)]$ , $\det\mathbf{I}$ are rotational invariants — free sanity checks on any diagonalization.
Symmetry shortcuts do the work in practice: an $n\ge3$ -fold symmetry axis is principal and forces the perpendicular plane degenerate; a mirror plane makes its normal principal.
The eigenvalue pattern classifies all bodies: spherical top ( $I_1{=}I_2{=}I_3$ ), symmetric top ( $I_1{=}I_2{\ne}I_3$ ), asymmetric top (all distinct) — with the triangle inequalities $I_i+I_j\ge I_k$ as reality constraints.
Key idea: the inertia tensor is real symmetric, hence orthogonally diagonalizable — its eigenvectors define the principal axes and its eigenvalues are the principal moments, and in that frame the rotational dynamics decouples.

The core insight

The inertia tensor is real symmetric, hence orthogonally diagonalizable: its eigenvectors define the principal axes and its eigenvalues are the principal moments. In that frame $\mathbf{L}$ and $\boldsymbol\omega$ are related component-by-component, decoupling the rotational dynamics.

Where this lives today

Molecular spectroscopy runs on this article's taxonomy. A molecule's rotational energy levels are set by its principal moments, so the eigenvalue pattern — spherical, symmetric, or asymmetric top — dictates the entire structure of its microwave spectrum. Methane, a spherical top, has no permanent dipole along any principal axis and shows no pure rotational spectrum at all; ammonia's symmetric-top ladder is so clean it powered the first maser; water, an asymmetric top, produces the tangled lines that are both the bane of infrared astronomy and — as megamasers — the beacon by which radio astronomers weigh supermassive black holes. When a 2026 radio survey identifies a molecule in an interstellar cloud, the identification is, at bottom, a fit of observed lines to the eigenvalues of an inertia tensor. The engineering version is attitude dynamics. Spacecraft flight software carries the vehicle's principal moments as configuration constants, and mission rules are written in eigenvalue language: spin-stabilize about the maximum-moment axis, never trust the intermediate one. The active frontier is estimating the tensor of things nobody built: on-orbit servicing and debris-removal missions must identify the principal axes of a tumbling, uncooperative satellite from lightcurves and lidar before daring an approach — this article's eigenvalue problem solved in reverse. And every robotics simulator and CAD package quietly reports principal moments and axes for each designed part — the diagonalization Euler did blind, now running thousands of times per second in physics engines.

The physical picture

Spin a rigid body about an arbitrary axis through its center of mass and you feel the problem immediately: the body fights you. Its angular momentum $\mathbf{L}=\mathbf{I}\boldsymbol{\omega}$ points somewhere else than $\boldsymbol{\omega}$ — the tensor article computed a $53.1°$ tilt for its dumbbell — and as the body turns, that misaligned $\mathbf{L}$ sweeps a cone. A rotating $\mathbf{L}$ demands torque, so the bearings get hammered once per revolution. Now rotate your chosen axis slowly through the body and watch the wobble: for three special, mutually perpendicular directions it dies out exactly. Those are the principal axes — the directions $\boldsymbol{\omega}$ can point without $\mathbf{I}$ rotating it, which is precisely the eigenvector condition $\mathbf{I}\hat e = I\hat e$ . What is remarkable is the census. However lopsided the body, there are always at least three such directions, always arrangeable as an orthogonal triad — and the number of ways nature can distribute them is small. If all three eigenvalues differ, the triad is unique: three lonely axes frozen in the body. If two eigenvalues coincide, an entire plane of principal directions opens up. If all three coincide, every axis in space is principal. That eigenvalue pattern — all different, two equal, all equal — is the complete taxonomy of rigid bodies:

Figure

The three families of rigid bodies — the eigenvalue pattern sets the freedom

Note what the figure does not say: it never mentions the body's shape, only its eigenvalues. A uniform cube is a spherical top — every axis through its center spins as cleanly as a sphere's (Derivation C proves it). Symmetry is sufficient to locate principal axes, and we will prove exactly how much it buys; but it is not necessary. The lumpy asteroid has its three axes too — you just have to diagonalize to find them, which is the eigenvalue problem this page now sets up.

The eigenvalue problem

Asking "about which directions does the body spin without wobble?" is asking for the vectors that $\mathbf{I}$ stretches without rotating. That is an eigenvector demand:

\mathbf{I}\,\hat e_a = I_a\,\hat e_a, \qquad a=1,2,3 \qquad (1)

where: $\hat e_a$ — the $a$ -th principal axis, a unit eigenvector fixed in the body; $I_a$ — the corresponding principal moment of inertia (the eigenvalue). Feed a spin along $\hat e_a$ into $\mathbf{L}=\mathbf{I}\boldsymbol{\omega}$ and out comes $\mathbf{L}=I_a\boldsymbol{\omega}$ : parallel, scalar, freshman-simple. Everything special about principal axes is this one line. Nontrivial solutions of $(\mathbf{I}-\lambda\mathbb{1})\hat e=0$ exist only where the matrix is singular, which turns the search into a polynomial root problem:

\det(\mathbf{I}-\lambda\,\mathbb{1})=0 \;\;\Longleftrightarrow\;\; -\lambda^3+(\mathrm{tr}\,\mathbf{I})\,\lambda^2-\tfrac12\!\left[(\mathrm{tr}\,\mathbf{I})^2-\mathrm{tr}(\mathbf{I}^2)\right]\lambda+\det\mathbf{I}=0 \qquad (2)

where: $\lambda$ — the eigenvalue variable, whose three roots are $I_1,I_2,I_3$ ; $\mathbb{1}$ — the $3\times3$ identity. The three coefficients are the rotational invariants of the tensor: $\mathrm{tr}\,\mathbf{I}=I_1+I_2+I_3$ , the middle invariant equals $I_1I_2+I_2I_3+I_3I_1$ , and $\det\mathbf{I}=I_1I_2I_3$ . Because a similarity transformation cannot change them, they are the same in every axis frame — free error detectors: after any diagonalization, check that the trace of the original matrix equals the sum of your eigenvalues; if not, the answer is wrong. But why should the roots of a cubic be real numbers, and why should the eigenvectors of a lopsided mass distribution come out perpendicular? The tensor article invoked the spectral theorem on faith. This is the page where the proof belongs.

Derivation AEigenvalues are real and principal axes are orthogonal

We are proving the two claims that make the principal-axis frame possible: every eigenvalue of the inertia tensor is a real number, and eigenvectors belonging to distinct eigenvalues are exactly perpendicular — with the degenerate case handled honestly at the end, since that is where the freedom of symmetric and spherical tops lives. The strategy is the classic Hermitian-matrix argument, using only one structural fact, $I_{ij}=I_{ji}$ , which follows from the definition $I_{ij}=\sum m(r^2\delta_{ij}-x_ix_j)$ by inspection. The move to watch: we temporarily allow the eigenvector to be complex, because reality of the eigenvalues is exactly what must be earned, not assumed — and the symmetry of $\mathbf{I}$ is what forces the imaginary parts to die.

Step 1. Let $\mathbf{I}\hat e_a = I_a \hat e_a$ with $\hat e_a$ possibly complex, and contract with the conjugate transpose $\hat e_a^{\dagger}$ :

\hat e_a^{\dagger}\,\mathbf{I}\,\hat e_a = I_a\, \hat e_a^{\dagger}\hat e_a

Why this combination: $\hat e_a^{\dagger}\hat e_a=|\hat e_a|^2$ is real and positive, so whatever we learn about the left side transfers directly to $I_a$ .

Step 2. Conjugate the left side and use $\mathbf{I}^{\dagger}=\mathbf{I}$ (real entries, symmetric):

\left(\hat e_a^{\dagger}\,\mathbf{I}\,\hat e_a\right)^{*} = \hat e_a^{\dagger}\,\mathbf{I}^{\dagger}\,\hat e_a = \hat e_a^{\dagger}\,\mathbf{I}\,\hat e_a

Why this finishes reality: the quadratic form equals its own complex conjugate, hence is real; dividing by the real, positive $|\hat e_a|^2$ makes $I_a$ real. Physically this had to happen — $I_a$ is a moment of inertia, measurable in $\mathrm{kg\,m^2}$ — but the mathematics needed the symmetry to deliver it.

Step 3. Now take two eigenpairs with $I_a \ne I_b$ and contract each eigenvalue equation with the other eigenvector:

\hat e_b^{\mathsf T}\,\mathbf{I}\,\hat e_a = I_a\,\hat e_b^{\mathsf T}\hat e_a, \qquad \hat e_a^{\mathsf T}\,\mathbf{I}\,\hat e_b = I_b\,\hat e_a^{\mathsf T}\hat e_b

Step 4. Symmetry makes the two left sides equal — $\hat e_b^{\mathsf T}\mathbf{I}\hat e_a=(\mathbf{I}\hat e_b)^{\mathsf T}\hat e_a$ — so subtracting gives

(I_a - I_b)\; \hat e_a\cdot\hat e_b = 0

Why the conclusion is forced: the first factor is nonzero by assumption, so the dot product must vanish. Orthogonality is not chosen for convenience; distinct eigenvalues compel it. The wobble-free directions of any rigid body arrive pre-packaged as a perpendicular set.

Step 5. Degenerate case. If $I_a = I_b$ , Step 4 says nothing, and indeed any linear combination $\alpha\hat e_a + \beta\hat e_b$ is again an eigenvector with the same eigenvalue: the eigenspace is a full plane (or all of space if triply degenerate). Within that subspace orthogonality is not forced but available — pick any orthonormal basis and the triad completes. This is precisely the freedom of the symmetric top (choose any perpendicular pair in the plane) and the spherical top (choose any orthonormal triad whatsoever). Verified: this is the spectral theorem for real symmetric matrices, specialized from the standard Hermitian proof; matches Goldstein §5.4 and Morin ch. 9. Cross-check: the cube-corner tensor in Worked example 2 has eigenvalues $\{0.018, 0.099, 0.099\}\ \mathrm{kg\,m^2}$ — real, with the degenerate pair spanning a plane perpendicular to the nondegenerate axis, exactly as Steps 4-5 predict.

With the triad in hand, "diagonalizing" is nothing more than adopting it as your coordinate frame. Stack the principal axes as the rows of a rotation matrix and the tensor's components transform as

\mathbf{I}' = R\,\mathbf{I}\,R^{\mathsf T} = \mathrm{diag}(I_1, I_2, I_3) \qquad (3)

where: $R$ — the orthogonal matrix whose rows are $\hat e_1^{\mathsf T}, \hat e_2^{\mathsf T}, \hat e_3^{\mathsf T}$ , so $R^{\mathsf T}=R^{-1}$ . Nothing about the body changed — the same mass sits in the same places — but described along its own eigen-directions the tensor keeps only its diagonal. The products of inertia vanish in this frame by construction: a product of inertia is an off-diagonal matrix element between two axes, and eigenvectors have none between them. The geometry of equation (3) is worth seeing rather than believing. The tensor's quadratic form traces out ellipses (in cross-section; ellipsoids in full 3D — the inertia ellipsoid is next in the track). In a generic frame the ellipse sits tilted, and the cross term $I_{xy}$ measures the tilt; rotating your axes onto the ellipse's own symmetry directions kills the cross term. Diagonalization is that rotation:

Figure

Diagonalization is a rotation of your axes onto the eigen-directions

What the principal frame buys

Write every rotational quantity along the eigen-triad and the matrix machinery collapses. Angular momentum, which in a generic frame mixes all three spin components into each of its own, becomes component-by-component:

\mathbf{L} = I_1\omega_1\,\hat e_1 + I_2\omega_2\,\hat e_2 + I_3\omega_3\,\hat e_3 \qquad (4)

where: $\omega_a=\boldsymbol{\omega}\cdot\hat e_a$ — the spin components along the principal axes. Each component of $\mathbf{L}$ answers only to its own component of $\boldsymbol{\omega}$ , scaled by its own moment. The misalignment of $\mathbf{L}$ and $\boldsymbol{\omega}$ is now transparent: unless the three $I_a$ are equal, the three components stretch by different factors, bending $\mathbf{L}$ away from $\boldsymbol{\omega}$ — except when $\boldsymbol{\omega}$ lies along a single axis and only one term survives. The full anatomy of that misalignment is the subject of angular momentum of a rigid body. The kinetic energy, a nine-term double sum $\tfrac12\sum I_{ij}\omega_i\omega_j$ in a generic frame, keeps only its squares:

T = \tfrac12\left(I_1\omega_1^2 + I_2\omega_2^2 + I_3\omega_3^2\right) \qquad (5)

where: the three terms are the energies of three independent "rotational directions" — the rigid-body analogue of writing a coupled oscillator problem in normal coordinates. This diagonal form is the starting point for the Lagrangian of every spinning top and for Euler's equations of motion, neither of which is tractable with cross terms alive. One constraint rounds out the algebra. Each principal moment is built from mass times squared distances, so none can be negative — and pairs of them bound the third:

I_i + I_j \;\ge\; I_k \qquad (i,j,k\ \text{distinct}) \qquad (6)

where: equality holds exactly for a planar body (a lamina), for which the perpendicular-axis relation $I_3=I_1+I_2$ saturates the bound. The inequality is a reality check in the literal sense: a matrix with eigenvalues $(1,1,3)$ , in any units, cannot be the inertia tensor of anything — no distribution of mass in three-dimensional space produces it. Each $I_i=\int\rho\,(x_j^2+x_k^2)\,\mathrm{d}V$ counts two coordinate directions, so summing two moments counts every coordinate at least as often as the third moment does.

you want	use	notes
the principal moments	roots of the cubic (2)	check: sum = trace, product = det
the axes themselves	eigenvectors of $\mathbf{I}$ ; symmetry shortcuts first	Derivation B below
$\mathbf{L}$ , $T$ in a generic frame	$\mathbf{L}=\mathbf{I}\boldsymbol{\omega}$ , $T=\tfrac12\boldsymbol{\omega}^{\mathsf T}\mathbf{I}\boldsymbol{\omega}$	tensor article's equations
$\mathbf{L}$ , $T$ in the principal frame	(4) and (5)	the frame Euler's equations live in
scalar $I$ about a known axis $\hat{\mathbf n}$	$I_a=\sum_k I_k n_k^2$	no diagonalization needed if the $I_k$ are known
sanity of any claimed $I_1,I_2,I_3$	positivity + triangle inequality (6)	lamina saturates; $(1,1,3)$ impossible

Symmetry shortcuts: axes for free

Solving the cubic is the method of last resort. Real bodies — machined rotors, molecules, textbook shapes — carry symmetries, and every symmetry pins down principal structure before any integral is computed. The workhorse rule: mass mirror-symmetric across a plane makes the plane's normal a principal axis (the products-of-inertia article proves this by pairing mass elements). The deeper rule concerns rotational symmetry, and it comes with a bonus that explains the entire symmetric-top family:

Go deeper · Derivation B · 2 minA symmetry axis is a principal axis and forces degeneracy

Derivation BA symmetry axis is a principal axis and forces degeneracy

The claim: if a body's mass distribution is unchanged by rotation through $2\pi/n$ about some axis with $n\ge3$ , then that axis is principal, and the two perpendicular principal moments are equal — the body is automatically a symmetric top. (For $n=2$ , mirror-type symmetry, the axis is still principal but the perpendicular moments may differ.) The strategy is an invariance argument, the kind that generalizes far beyond mechanics: the tensor is built from the mass distribution, so any rotation that fixes the distribution must fix the tensor's components — and we simply ask which matrices survive that demand. Watch how the conclusion strengthens as $n$ passes from 2 to 3: the constraint equations acquire more independent directions to act in, and pin down more of the matrix.

Step 1. Take the symmetry axis as $\hat z$ and let $R_z(\phi)$ be the rotation through the symmetry angle $\phi=2\pi/n$ . Invariance of the mass distribution under this rotation means the tensor's components are the same before and after:

\mathbf{I} = R_z(\phi)\,\mathbf{I}\,R_z(\phi)^{\mathsf T}

Why this is the right statement: the components of $\mathbf{I}$ are integrals over $\rho(\mathbf{r})$ ; if the rotated body is indistinguishable from the original, every such integral — computed in the same fixed axes — returns the same value.

Step 2. Impose the condition on the $(x,z)$ and $(y,z)$ components. Under $R_z(\phi)$ the pair $(I_{xz}, I_{yz})$ transforms like a 2D vector rotating through $\phi$ :

\begin{pmatrix} I_{xz} \\ I_{yz} \end{pmatrix} \;\longmapsto\; \begin{pmatrix} \cos\phi & -\sin\phi \\ \sin\phi & \cos\phi \end{pmatrix} \begin{pmatrix} I_{xz} \\ I_{yz} \end{pmatrix} = \begin{pmatrix} I_{xz} \\ I_{yz} \end{pmatrix}

Why this kills both products: a 2D rotation through any $\phi\ne 0,\pi,2\pi,\dots$ has no nonzero fixed vector — its eigenvalues are complex. So invariance forces $I_{xz}=I_{yz}=0$ , and $\hat z$ is an eigenvector of $\mathbf{I}$ : a principal axis. Note this already works for $n=2$ ( $\phi=\pi$ maps the pair to its negative, same conclusion).

Step 3. Impose the condition on the upper-left $2\times2$ block. The in-plane components $(I_{xx}-I_{yy},\ 2I_{xy})$ transform as a 2D vector rotating through $2\phi$ :

\begin{pmatrix} I_{xx}-I_{yy} \\ 2 I_{xy} \end{pmatrix} \;\longmapsto\; R(2\phi) \begin{pmatrix} I_{xx}-I_{yy} \\ 2 I_{xy} \end{pmatrix} = \begin{pmatrix} I_{xx}-I_{yy} \\ 2 I_{xy} \end{pmatrix}

Why the doubled angle matters: for $n=2$ , $2\phi=2\pi$ and the condition is empty — a two-fold symmetric body (a brick face) keeps $I_{xx}\ne I_{yy}$ and stays asymmetric. But for $n\ge3$ , $2\phi$ is not a multiple of $2\pi$ , the fixed vector must vanish, and $I_{xx}=I_{yy}$ with $I_{xy}=0$ : the perpendicular plane is a degenerate eigenplane.

Step 4. Read off the taxonomy. Any body with one $n\ge3$ axis is a symmetric top (triangular prism, hexagonal nut, cylinder as $n\to\infty$ ). Any body with $n\ge3$ axes in two different directions — a cube has four 3-fold body-diagonal axes — must be degenerate in two independent planes at once, which forces all three eigenvalues equal: a spherical top, no sphere required. Verified: consistent with Landau-Lifshitz §32's classification of tops; the cube conclusion is confirmed numerically in Derivation C (all three moments $Ma^2/6$ ) and the hexagonal/triangular cases match direct integration.

The recipe that falls out, in the order a practitioner actually uses it: first hunt for an $n\ge3$ axis (symmetric or spherical top — done, no integrals); then hunt for mirror planes (each normal is principal; three mutually perpendicular ones fix the whole triad); only then — for the genuinely lumpy body — assemble the six independent components and hand the cubic to a computer.

Go deeper · Derivation C · 2 minPrincipal moments of a uniform rectangular block and the cube surprise

Derivation CPrincipal moments of a uniform rectangular block and the cube surprise

We are computing the full principal triad of the most-used asymmetric top — the uniform block of mass $M$ and edges $a,b,c$ — and extracting from it the least intuitive fact in this article: the uniform cube is a spherical top. The strategy: symmetry hands us the axes for free (three mirror planes), so the only work is three diagonal integrals; the payoff is a formula you will reuse constantly, and the $a=b=c$ limit delivers the surprise. Watch that each moment sees only the two transverse edges — the mass's distance from an axis does not involve the coordinate along it.

Step 1. Fix axes through the center, along the edges. Each coordinate plane is a mirror plane of the block, so all three axis directions are normals of mirror planes: the axes are principal by symmetry, and all products of inertia vanish — e.g. $I_{xy}=-\int\rho\,xy\,\mathrm{d}V=0$ because the integrand is odd in $x$ over the symmetric interval $[-a/2,a/2]$ .

Step 2. Compute one diagonal component. With $\rho=M/(abc)$ ,

I_1 = I_{xx} = \int \rho\,(y^2+z^2)\,\mathrm{d}V = \frac{M}{b}\int_{-b/2}^{b/2} y^2\,\mathrm{d}y + \frac{M}{c}\int_{-c/2}^{c/2} z^2\,\mathrm{d}z = \frac{M}{12}\left(b^2+c^2\right)

Why the split works: $y^2+z^2$ separates, and each factor integrates over its own edge, giving $\mathrm{edge}^2/12$ each — the same $1/12$ as a thin rod's central moment, which is no coincidence: a block is a bundle of rods.

Step 3. Cycle the labels: $I_2=\tfrac{M}{12}(a^2+c^2)$ , $I_3=\tfrac{M}{12}(a^2+b^2)$ . Ordering follows the edges: the axis along the longest edge has the smallest moment (its transverse edges are the short ones). For $a>b>c$ : $I_1<I_2<I_3$ — a bona fide asymmetric top with a unique triad and an unstable middle axis.

Step 4. The cube surprise. Set $a=b=c$ :

I_1=I_2=I_3=\frac{M a^2}{6}

Why this makes the cube spherical: with all three eigenvalues equal, $\mathbf{I}=\tfrac{Ma^2}{6}\,\mathbb{1}$ commutes with every rotation, so every axis through the center — edge, face diagonal, body diagonal — has the identical moment $Ma^2/6$ and is principal. Dynamically a uniform cube is indistinguishable from a sphere of equal moment. This is Step 4 of Derivation B made concrete: four independent 3-fold axes force total degeneracy. *Verified: numerically, $M=1$ , $a=b=c=1$ : direct integration gives $I=0.1667=1/6$ about a face axis and $\hat{\mathbf n}^{\mathsf T}\mathbf{I}\hat{\mathbf n}=1/6$ about the body diagonal $\hat{\mathbf n}=(1,1,1)/\sqrt3$ alike; matches Morin ch. 9 and Taylor ch. 10 tabulated values.*

Limiting cases

Case	Result	Meaning
Spherical top ( $I_1=I_2=I_3$ )	$\mathbf{I}=I\,\mathbb{1}$	Every axis principal; $\mathbf{L}=I\boldsymbol{\omega}$ always. Sphere, uniform cube, regular tetrahedron, methane.
Symmetric top ( $I_1=I_2\ne I_3$ )	degenerate eigenplane $\perp\hat e_3$	Symmetry axis plus any perpendicular pair; free precession lives here. Cylinder, frisbee, benzene.
Asymmetric top ( $I_1<I_2<I_3$ )	unique triad	Three lonely axes; spin about $\hat e_2$ is unstable — the tennis-racket effect. Brick, phone, water molecule.
Planar lamina	$I_3=I_1+I_2$	Triangle inequality (6) saturated — the perpendicular-axis theorem's fingerprint in the eigenvalues.
Thin rod	$I_1\to0$ , $I_2=I_3$	Mass on the axis resists nothing; a degenerate "prolate" extreme of the symmetric top.
$n=2$ symmetry only	axis principal, no degeneracy	A brick face's 2-fold axis is principal but $I_{xx}\ne I_{yy}$ — why $n\ge3$ is the degeneracy threshold.

Common trapPrincipal axes belong to the body and a point, not to the body alone

"The principal axes of a brick point along its edges" is true only about its center of mass. The tensor is origin-dependent — shift the reference point by $\mathbf{a}$ and the Steiner correction $M(a^2\delta_{ij}-a_ia_j)$ adds new off-diagonal structure that generally rotates the eigenvectors. About a generic corner of the same brick, the principal axes point along entirely different directions (Worked example 2 exhibits this for a cube — and shows the one exception: shifts along an existing principal axis leave it principal). A second overreach to resist: $\mathbf{L}\parallel\boldsymbol{\omega}$ holds only when $\boldsymbol{\omega}$ points along one principal axis — in the principal frame but with a generic $\boldsymbol{\omega}$ , the three components stretch differently and $\mathbf{L}$ still tilts away.

Worked examples

1 · Your phone is an asymmetric top (confidence builder)

A phone of mass $M=0.21\ \mathrm{kg}$ measures $a\times b\times c = 0.165\times0.076\times0.008\ \mathrm{m}$ (length $x$ , width $y$ , thickness $z$ ). Find the principal axes and moments, classify the body, verify the triangle inequality, and compute the kinetic energy of a 2-revolutions-per-second flip about the width axis and about the face normal.

Step 1. Axes and moments by symmetry. The three central coordinate planes are mirror planes, so the length, width, and thickness directions are the principal axes — Derivation C applies verbatim with $M/12 = 0.0175\ \mathrm{kg}$ :

I_x = \tfrac{M}{12}(b^2+c^2) = 0.0175\,(0.076^2+0.008^2) = 1.02\times10^{-4}\ \mathrm{kg\,m^2}

I_y = \tfrac{M}{12}(a^2+c^2) = 4.78\times10^{-4}\ \mathrm{kg\,m^2}, \qquad I_z = \tfrac{M}{12}(a^2+b^2) = 5.78\times10^{-4}\ \mathrm{kg\,m^2}

Step 2. Classify. Three distinct values in ratio $1 : 4.67 : 5.65$ — an asymmetric top. The middle moment belongs to the width axis $\hat y$ : that is the intermediate axis, and the end-over-end flip about it is the one that tumbles (the tennis-racket instability — try it, with the phone case on).

Step 3. Triangle inequality audit: $I_x + I_y = 5.80\times10^{-4} \ge I_z = 5.78\times10^{-4}$ ✓ — satisfied with only $0.4\%$ slack. Why so tight: at $8\ \mathrm{mm}$ thick the phone is nearly a lamina, and a lamina saturates the bound exactly ( $I_z=I_x+I_y$ , the perpendicular-axis theorem). The $c^2$ terms are the entire slack.

Step 4. Energies at $\omega = 2\ \mathrm{rev/s} = 4\pi\ \mathrm{rad/s}$ . Face-normal spin (pizza-toss style): $T = \tfrac12 I_z\omega^2 = \tfrac12(5.78\times10^{-4})(4\pi)^2 = \boxed{45.6\ \mathrm{mJ}}$ . Width-axis flip: $T = \tfrac12 I_y \omega^2 = \boxed{37.7\ \mathrm{mJ}}$ . Same spin rate, different price — the tensor is the exchange rate between spin and energy, direction by direction.

2 · A cube spun about its corner diagonal — degeneracy in action (exam level)

A uniform cube, $M=1.2\ \mathrm{kg}$ , edge $a=0.3\ \mathrm{m}$ , pivots about one corner. Find the inertia tensor about the corner, its principal moments and axes, and the moment about the body diagonal through that corner. ( $Ma^2 = 0.108\ \mathrm{kg\,m^2}$ .)

Step 1. Start from the center, where the cube is a spherical top (Derivation C): $\mathbf{I}^{\mathrm{cm}} = \tfrac{Ma^2}{6}\,\mathbb{1} = 0.018\,\mathbb{1}$ . Shift to the corner with the tensor Steiner theorem, displacement $\mathbf{d}=\tfrac{a}{2}(1,1,1)$ , $d^2 = \tfrac34 a^2$ :

I_{ij} = \tfrac{Ma^2}{6}\,\delta_{ij} + M\!\left(d^2\delta_{ij} - d_i d_j\right) = \begin{pmatrix} 0.072 & -0.027 & -0.027 \\ -0.027 & 0.072 & -0.027 \\ -0.027 & -0.027 & 0.072 \end{pmatrix}\ \mathrm{kg\,m^2}

with diagonal $\tfrac{Ma^2}{6}+\tfrac{Ma^2}{2} = \tfrac{2Ma^2}{3} = 0.072$ and off-diagonal $-\tfrac{Ma^2}{4} = -0.027$ . The corner frame is not principal: the edge axes now carry products of inertia.

Step 2. Diagonalize by symmetry, not by cubic. The body diagonal through the corner is a 3-fold symmetry axis of the cube, so by Derivation B it is principal and the perpendicular plane is degenerate. Guess $\hat{\mathbf n} = (1,1,1)/\sqrt3$ and check: each row of $\mathbf{I}$ dotted into $(1,1,1)$ gives $0.072 - 2(0.027) = 0.018$ , so $\mathbf{I}\hat{\mathbf n} = 0.018\,\hat{\mathbf n}$ — an eigenvector, eigenvalue $\boxed{I_{\parallel} = 0.018\ \mathrm{kg\,m^2} = \tfrac{Ma^2}{6}}$ .

Step 3. The remaining pair from the invariants — no cubic needed. $\mathrm{tr}\,\mathbf{I} = 3(0.072) = 0.216 = I_{\parallel} + 2I_{\perp}$ (degenerate pair), so $I_{\perp} = \tfrac{0.216-0.018}{2} = \boxed{0.099\ \mathrm{kg\,m^2} = \tfrac{11\,Ma^2}{12}}$ . Determinant audit: $I_{\parallel}I_{\perp}^2 = 0.018\times0.099^2 = 1.764\times10^{-4}$ , matching $\det\mathbf{I}$ computed directly ✓. Principal axes: the diagonal $\hat{\mathbf n}$ , plus any orthogonal pair in the plane perpendicular to it — Step 5 of Derivation A's freedom, realized in wood and glue.

Step 4. The sting in the tail: the diagonal moment about the corner equals the moment about the center, $Ma^2/6$ — the Steiner shift changed nothing along that axis. Why: the body diagonal through the corner passes through the center of mass, so the displacement $\mathbf{d}$ is parallel to the axis and adds no perpendicular distance: the Steiner term $M(d^2\delta_{ij}-d_id_j)$ has $\hat{\mathbf n}$ as a zero-eigenvalue eigenvector. Spinning a cube about its body diagonal costs the same whether you grip it at the center or at the corner.

Figure

A cube pivoted at a corner — the 3-fold diagonal axis and its degenerate plane

Origins · 1 minLeonhard Euler and how this was discovered

Leonhard Euler

1707–1783 · St Petersburg and Berlin

Principal axes entered physics as a puzzle Euler inherited. In 1755 his Hungarian correspondent Johann Andreas Segner discovered that every rigid body admits axes of "permanent rotation" and counted three of them — but could not build a dynamics on the observation. Euler could. His 1765 treatise Theoria motus corporum solidorum seu rigidorum defined the moments of inertia, proved that the three axes exist for every body and are mutually perpendicular — effectively diagonalizing a symmetric operator ninety years before matrices were invented — and then wrote the equations of rotational motion in exactly the frame those axes define, where they take their clean, workable form. By this point Euler was almost totally blind, dictating calculations of legendary accuracy to his assistants in St Petersburg. The eigenvalue problem this article solves is, historically, the first eigenvalue problem in physics — solved before the mathematics it belongs to had a name, by a man who could no longer see the spinning bodies he understood better than anyone alive.

History

1755

— Segner identifies the three axes of permanent rotation possessed by any rigid body, the first sighting of principal axes.

1765

— Euler's Theoria motus corporum solidorum proves the orthogonal triad exists for every body, defines the principal moments, and builds rigid-body dynamics in the principal frame.

1829

— Cauchy proves the spectral theorem for symmetric quadratic forms, giving Euler's axes their permanent algebraic foundation — the same mathematics that later organizes stress tensors and quantum observables.

1926

— Dennison works out the quantum rotational energy levels of symmetric and asymmetric tops, turning the eigenvalue pattern of the inertia tensor into measurable molecular spectra.

1985

— Dzhanibekov's wing-nut flip aboard Salyut 7 makes the asymmetric top's unstable middle axis a viral demonstration, decades before the word existed.

Going further

References. Morin, Introduction to Classical Mechanics, ch. 9 · Goldstein, Poole & Safko, Classical Mechanics 3rd ed., ch. 5 · Landau & Lifshitz, Mechanics Vol. 1, §32 · Taylor, Classical Mechanics, ch. 10

Provenance. Definitions and the spectral-theorem argument follow Goldstein, Poole & Safko §5.4 and Morin ch. 9. Box and cube principal moments ( $I=\tfrac{M}{12}(b^2+c^2)$ , cube $Ma^2/6$ ) are standard textbook values verifiable by direct integration. Triangle inequalities and top classification from Landau-Lifshitz §32. A verifier should recompute the block integrals, confirm $\mathrm{tr}\,\mathbf{I}$ and $\det\mathbf{I}$ as invariants, and check the orthogonality/reality proof. Worked-example numbers on this page were recomputed numerically for this version: phone $I=(1.02, 4.78, 5.78)\times10^{-4}\ \mathrm{kg\,m^2}$ , $T=45.6/37.7\ \mathrm{mJ}$ at $4\pi\ \mathrm{rad/s}$ ; cube corner $I_\parallel=Ma^2/6=0.018$ , $I_\perp=11Ma^2/12=0.099\ \mathrm{kg\,m^2}$ , trace $0.216$ , det $1.764\times10^{-4}$ .

Builds on: Moment of Inertia Tensor · Products of Inertia and the Inertia Matrix · Parallel Axis (Steiner) Theorem Leads to: Inertia Ellipsoid · Classification of Rigid Bodies: Spherical, Symmetric, and Asymmetric Tops · Euler's Equations of Motion · Angular Momentum of a Rigid Body

Do physicists actually solve the characteristic cubic by hand to find principal moments?

Almost never, and the article's own examples show the escape routes. First resort: symmetry — an $n\ge3$ axis or a stack of mirror planes fixes the axes outright, and then the moments are just three diagonal integrals (the phone). Second resort: partial symmetry plus invariants — in the cube-corner example, symmetry supplied one eigenvector, and the trace and determinant then determined the other two eigenvalues by algebra, no cubic in sight. Third resort: a computer — for a genuinely lumpy body you hand the $3\times3$ matrix to an eigensolver, which is numerically trivial. The cubic formula itself is the one tool nobody reaches for; its invariants — trace, determinant — are used constantly as free checks.

For a symmetric top, which perpendicular pair of axes am I supposed to choose?

Whichever you like — and that freedom is not sloppiness, it is the physics. Inside a degenerate eigenplane every direction responds to spin with the same moment $I_1=I_2$ , so no experiment can distinguish one in-plane pair from another; the choice is as arbitrary as choosing $x$ and $y$ axes in empty space. Convention picks something convenient: for a cylinder, any diameter at $t=0$ ; for a molecule, axes adapted to the nuclei. The freedom is exact, protected by the symmetry (Derivation B): free precession of a symmetric top is the same about every in-plane axis, which is why the motion is so clean.

Is the principal frame of an asymmetric top completely unique?

Unique as a set of three lines, yes — that is what distinct eigenvalues buy (Derivation A, Step 4 forces orthogonality and one-dimensional eigenspaces). But three residual freedoms remain, all trivial: you may flip the sign of any axis ( $\hat e_a\to-\hat e_a$ is the same eigenvector), you may reorder the labels 1, 2, 3, and physicists conventionally fix both by demanding $I_1\le I_2\le I_3$ and a right-handed triad. Watch handedness: after sign flips you can accidentally hold a left-handed frame, where Euler's equations pick up sign errors — numerical eigensolvers promise neither ordering nor handedness, so attitude-dynamics code always post-processes into the convention.

Does every principal axis have to be a symmetry axis of the body?

No — that direction of the logic fails completely, and the potato is the proof. Symmetry implies principal (Derivation B), but principal does not imply symmetry: the spectral theorem guarantees three orthogonal principal axes for any mass distribution, including one with no symmetry element whatsoever. A random asteroid has its full triad; you find it by diagonalizing, not by staring at the shape. Principal axes are an algebraic birthright of every real symmetric tensor; symmetry is merely the happy circumstance that lets you locate them without computing. Absence of symmetry costs you only convenience, never existence.

Are the principal axes through my chosen pivot the same as through the center of mass?

Generally no — each reference point owns its own triad. Shifting the origin by $\mathbf{a}$ adds the Steiner term $M(a^2\delta_{ij}-a_ia_j)$ , which carries its own eigenstructure (its zero eigenvalue points along $\mathbf{a}$ ), and the sum of two symmetric matrices does not share eigenvectors with either unless they commute. The cube-corner example shows both faces of this: the body diagonal survived as principal because the shift was along it, while the edge axes — principal at the center — acquired products of inertia at the corner and rotated into a new degenerate pair. Practical rule: quote principal axes with their reference point; unqualified, they mean the center-of-mass triad — the one CAD software reports.

What happens to the principal axes if the body is not quite rigid — a spinning gymnast, or a satellite with moving parts?

The tensor becomes time-dependent, and with it the eigenvalues and eigenvectors: a gymnast tucking mid-somersault is deliberately editing her principal moments (slashing the one about her somersault axis to spin faster), and a satellite deploying panels migrates its principal axes in flight. At each instant the body still has a principal triad (the spectral theorem asks only for a symmetric matrix, not a permanent one), so the classification applies frame by frame; what fails is the convenience of body-fixed constants, and Euler's equations acquire extra terms from $\dot{\mathbf{I}}$ . The catastrophic version is Explorer 1's lesson — flexible parts dissipate energy and drag a "rigid" body's spin toward the maximum-moment axis — covered in the tensor article's FAQ.

Why does the intermediate principal axis get singled out as unstable when the eigenvalue problem treats all three axes identically?

The eigenproblem is indeed symmetric among the axes — instability enters only when you add dynamics on top of the algebra. Perturb a spin about principal axis $\hat e_a$ and Euler's equations couple the two small transverse spin components with a feedback coefficient proportional to $(I_a-I_b)(I_a-I_c)/I_bI_c$ . About the largest or smallest moment the two differences share a sign and the perturbation just circulates — stable wobble. About the middle moment they have opposite signs, the feedback is anti-restoring, and the perturbation grows exponentially: the tennis-racket flip. The asymmetry is dynamical, not algebraic — it needs the ordering $I_1<I_2<I_3$ , so it does not exist for symmetric or spherical tops. The full derivation belongs to Euler's equations.

Principal Axes and Principal Moments of Inertia

Where this lives today

The physical picture

The eigenvalue problem

What the principal frame buys

Symmetry shortcuts: axes for free

Limiting cases

Worked examples

1 · Your phone is an asymmetric top (confidence builder)

2 · A cube spun about its corner diagonal — degeneracy in action (exam level)

Going further

Before this

Where this leads