Rotation matrices

The formula

R_z(θ) = [ cos θ   −sin θ    0 ]
         [ sin θ    cos θ    0 ]
         [   0        0      1 ]

The elementary rotation about z. At θ = 30°, cos θ = 0.8660 and sin θ = 0.5, so the x-axis unit vector (1, 0, 0) lands at (0.8660, 0.5, 0) — swung 30° toward y, length unchanged. R_x and R_y follow the same pattern on their own axes.

Rᵀ · R = I,    det(R) = +1

The membership card for SO(3), the group of all proper rotations. Columns are unit-length and mutually perpendicular, so the inverse is the transpose — undoing a rotation costs nothing but an index swap. A determinant of −1 means you built a reflection, not a rotation, and your frame just went left-handed.

C[i][j] = cos( angle between new axis i and old axis j )

Why flight software calls it the DCM. Each of the nine entries is literally the cosine of the angle between one axis of the body frame and one axis of the reference frame. Column 1 of a body-to-nav DCM is the body x-axis written out in nav coordinates — equivalently, row 1 of the nav-to-body DCM is the nav x-axis written out in body coordinates.

R_total = R₂ · R₁        (R₁ acts first)

Composition is matrix multiplication, read right to left — and it does not commute. Two rotations in the other order are a different attitude, full stop.

C_nav←body = R_z(ψ) · R_y(θ) · R_x(φ);   C_body←nav = its transpose = R_x(φ)ᵀ · R_y(θ)ᵀ · R_z(ψ)ᵀ
θ = −asin(C[2][0]),  ψ = atan2(C[1][0], C[0][0]),  φ = atan2(C[2][1], C[2][2])

The aerospace yaw-pitch-roll (ZYX) construction and its extraction — the angle formulas apply to C_nav←body, the body-to-nav DCM. Three angles in, nine cosines out, and the trip back through asin and atan2 recovers them — check: ψ, θ, φ of 30°, 20°, 10° builds a body-to-nav matrix with C[2][0] = −0.342, and −asin(−0.342) returns exactly 20°.

θ_rot = acos( (tr(R) − 1) / 2 )

The total rotation angle hiding in any R, read off the diagonal. A 90° rotation about z has trace 0 + 0 + 1 = 1, and acos(0) gives back 90°. This is Euler's theorem in one line: every rotation matrix is a single turn about a single axis, whether you built it that way or not.

Where you meet it

Sensor alignment on the test stand. The triaxial accelerometer never mounts perfectly along the article's axes. The alignment matrix in the data-reduction script — measured on a rate table or from a boresight survey — is a rotation matrix, and every channel of flight-quality vibration data goes through it before anyone sees a plot. When the lateral channel shows suspicious coupling, the first question is whether that matrix got transposed.
Attitude in flight software. The navigation filter carries vehicle attitude as a DCM (or a quaternion it converts to one), transforming accelerometer and gyro measurements from body frame to nav frame every cycle. Guidance, the Kalman filter, and the telemetry display each want a different frame, and rotation matrices are the plumbing between all of them.
Robot cells and CMM work. A six-axis arm is a chain of rotation matrices multiplied left to right from base to tool flange. Teach a fixture offset with the wrong frame selected and the part gets machined in a coordinate system that exists only in the pendant.
The review board argument. Someone presents an attitude error plot, someone else asks whether the DCM is body-to-nav or nav-to-body, and twenty minutes disappear. Both conventions are in the room's heritage code. The matrix that settles it is the transpose of the one that doesn't.

How it works

Nine numbers, but only three degrees of freedom — orthonormality imposes six constraints, which is exactly why the DCM is redundant storage for an attitude and why corrupted ones are detectable. If the rows of a matrix claiming to be a DCM don't each have unit norm, it isn't one; something upstream truncated, transposed, or overwrote it.

The mistake people actually make is active versus passive. The same nine numbers can mean "rotate this vector within a fixed frame" or "re-express this fixed vector in a rotated frame." The two interpretations are transposes of each other, and both are legitimate conventions used by legitimate references. Every organization has a story about two subsystems that each tested clean and produced a sign error when integrated, because one team's R was the other team's Rᵀ. The fix is boring and permanent: name every frame in the variable (C_body_from_nav), and test at an attitude where the transpose gives a different answer — everything passes at zero roll, zero pitch.

Order matters, physically. Take the vector (0, 1, 0). Rotate 90° about z, then 90° about x: it ends at (−1, 0, 0). Same two rotations, opposite order: (0, 0, 1). Not close, not a sign flip — pointing along a different axis entirely. Any test sequence, tooling operation, or maneuver plan that composes rotations has to fix the order in writing.

Euler angles are a readout, not a home. Three angles feel natural and fit on a telemetry page, but there are twelve valid sequences and every one of them has a singularity. For aerospace ZYX, it's pitch = ±90°: the asin saturates, yaw and roll collapse into one indistinguishable angle, and the kinematic rate equations divide by cos θ. That's gimbal lock — the software version of the mechanical problem — and it's why attitude gets stored as a DCM or quaternion and only displayed as Euler angles. Vertical launch vehicles and tumbling test articles will find the singularity for you.

Numerical hygiene. Multiply thousands of small incremental rotations together — which is exactly what a strapdown attitude propagator does — and floating-point error slowly walks the product away from orthogonality; the columns stop being perpendicular and the determinant drifts off 1. The drift per multiply is tiny, but propagators run at hundreds of hertz for hours. Standard practice is periodic re-orthogonalization: Gram–Schmidt the columns, or round-trip through a normalized quaternion. Related trap: never interpolate or average rotation matrices element-by-element. The midpoint of two rotation matrices is not a rotation matrix; interpolate the angle, or slerp the quaternion.

Know its limits. A rotation matrix encodes orientation only. Rigid-body motion needs a translation too (the 4×4 homogeneous transform bolts one on), and anything that scales, shears, or reflects is outside SO(3) by definition — if your "rotation" changes a vector's length, it isn't your convention that's wrong, it's your matrix.

History

The core theorem is older than matrices. In 1775 Leonhard Euler, working at the St. Petersburg Academy, proved by spherical geometry that any displacement of a rigid body about a fixed point is a single rotation about some axis — published in 1776 as "Formulae generales pro translatione quacunque corporum rigidorum" in the Novi Commentarii academiae scientiarum Petropolitanae ^[1]^[2]. The same program of rigid-body work gave us the three orientation angles that still carry his name, along with direction cosines as the bookkeeping for them ^[1]^[3]. Euler had the full content of the rotation matrix a half-century before anyone wrote mathematics in rows and columns.

The next move came from an unlikely quarter. Olinde Rodrigues (1795–1851) had left academic mathematics to become a Paris banker, financing French railways, and in 1840 published a paper deriving, by purely geometric means, the law for composing two finite rotations into one — axis and angle in, axis and angle out ^[4]^[5]. His composition rule is, entry for entry, the multiplication law for unit quaternions, worked out three years before Hamilton had his flash of insight on Broome Bridge in 1843 ^[4]. The paper appeared in the Journal de Mathématiques Pures et Appliquées and then sat largely unnoticed for decades ^[5]^[6].

The twentieth century turned the theorem into flight hardware, and flight hardware exposed the parametrization's weak spot. The Apollo command module carried its attitude reference on a three-gimbal inertial platform — engineers had weighed a fourth gimbal and rejected it for size, simplicity, and reliability, accepting that the platform would tumble its reference if the middle gimbal neared 90° ^[7]. The computer flashed a warning at 70° and froze the IMU at 85° ^[7]. During Apollo 11's lunar orbit operations, with Houston calling out how close he was drifting to the forbidden zone, Mike Collins radioed the line every guidance engineer since has quoted: "How about sending me a fourth gimbal for Christmas" ^[7]^[8]. He was asking, in effect, for a nonsingular attitude representation — which is precisely what storing the DCM or the quaternion, rather than three Euler angles, gives the software versions of that platform today.

Related tools

/tools/convert-angle — degrees, radians, and mils on the way into every cos θ and back out of every atan2
/tools/moment-of-inertia-shapes — inertia properties transform between frames by the same R, applied as R · I · Rᵀ
/tools/orbital-velocity-period — the orbit frame your body-to-LVLH rotation matrix is defined against
/tools/hohmann-transfer — transfer burns count only if attitude control has the DCM pointing the engine the right way
/tools/antenna-gain-beamwidth — pointing a narrow beam is an az/el frame transformation done with these matrices

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →