Introduction
Spin polarization of electrons in intense laser fields is usually treated as a radiative process: the spin flips when a photon is emitted, and the flip rate follows the structure of the field [1 , 2 , 3 , 4 , 5 , 6 ] . That channel closes as the quantum parameter χ \chi χ becomes small, its probability per pulse falling as α α 0 2 N \alpha\alpha_0^2 N α α 0 2 N for N N N cycles at normalized amplitude a 0 a_0 a 0 . The spin keeps moving below it. Inside a pulse with a 0 ∼ 10 a_0 \sim10 a 0 ∼ 10 the Thomas–Bargmann–Michel–Telegdi (BMT) equation [7 ] turns the rest-frame polarization through radians, and nearly all of that rotation is undone while the pulse passes.
What survives is a holonomy. For a plane-wave background, A μ = A μ ( η ) A^\mu= A^\mu(\eta) A μ = A μ ( η ) with η = k ⋅ x \eta= k \cdot x η = k ⋅ x the light-front phase, and for a pulse of finite duration, the rotation surviving after the field has gone is Θ n e t = − 1 2 a e 2 A \Theta_{\rm net} = -\frac{1}{2}a_e^2\mathcal{A} Θ net = − 2 1 a e 2 A , where a e = ( g − 2 ) / 2 a_e = (g-2)/2 a e = ( g − 2 ) /2 is the anomalous magnetic moment and A \mathcal{A} A is twice the signed area that the vector potential encloses in the polarization plane. Every property of the pulse acts through the curve that a ⊥ ( η ) a_\perp(\eta) a ⊥ ( η ) traces in that plane, and at leading order through its area alone. The electron energy, the carrier frequency, the carrier-envelope phase and the rate at which the curve is traced drop out. Chirp and envelope shape do not: they deform the curve, and A \mathcal{A} A follows.
The area carries a physical name. In the Coulomb gauge the spin angular momentum density of the field is ϵ 0 E × A \epsilon_0 \mathbf{E} \times\mathbf{A} ϵ 0 E × A , and for a plane wave E ∝ − a ˙ ⊥ \mathbf{E} \propto-\dot{\mathbf{a}}_\perp E ∝ − a ˙ ⊥ while A ∝ a ⊥ \mathbf{A} \propto\mathbf{a}_\perp A ∝ a ⊥ , so that A \mathcal{A} A is the spin angular momentum the pulse carries per unit area. The net spin rotation of the electron therefore measures the helicity of the light, weighted by the square of the anomaly. Section III makes that statement quantitative.
Two of the three ingredients of this result are in print, in two literatures that have never cited one another, and Sec. IV sets out the division in full. The exact spin problem for an electron with an anomalous moment in a plane wave was solved by the Tomsк school in the 1960s [8 , 9 , 10 , 11 ] and rediscovered in a gravitational context three decades later [12 ] : the residual dynamics is a rotation whose coupling is the anomaly alone, the first order in the anomaly is algebraic in a ⊥ ( η ) a_\perp(\eta) a ⊥ ( η ) , and exactly two profiles integrate in closed form. Separately, the net reorientation of a precessing magnetic dipole after a burst of radiation has passed is known to be of second order in the coupling and to equal the helicity of the burst [13 ] , a statement phrased as a holonomy in the memory literature [14 ] . Neither line supplies the general second order for a pulse of arbitrary shape, and neither treats a charged relativistic particle, where the reduction of the coupling from g / 2 g/2 g /2 to a e a_e a e is the whole of the physics. Substituting an electron into the dipole formula overestimates the answer by six orders of magnitude.
The rotation axis is the propagation direction in the electron rest frame, so in a head-on or copropagating geometry the helicity is conserved through O ( a e 6 ) O(a_e^6) O ( a e 6 ) and the measurable quantity is the rotation of a transverse polarization relative to the momentum. At oblique incidence that axis leaves the momentum and the helicity statement lapses. Linear polarization gives no net rotation at any g g g , exactly and at every order, because the curve degenerates to a segment.
Sections II and III carry the derivation. Section IV is the delimitation against earlier work. Section V reports the numerical verification, Sec. VI the departure from the law in a focused beam, and Sec. VII the signal-to-background budget, which is unfavorable and is given in full. Technical material, the validation tables and the extended algebra are in the Supplemental Material, reproduced in full in the Appendix of this preprint.
Spin transport in a plane-wave pulse
Background, orbit and assumptions
The background is an exact plane wave in the metric ( + , − , − , − ) (+,-,-,-) ( + , − , − , − ) : the dimensionless potential a μ ( η ) = e A μ ( η ) / m a^\mu(\eta)=eA^\mu(\eta)/m a μ ( η ) = e A μ ( η ) / m depends on the light-front phase alone, with k 2 = 0 k^2=0 k 2 = 0 and k ⋅ a = 0 k\cdot a=0 k ⋅ a = 0 . Its transverse part is a ⊥ \mathbf{a}_\perp a ⊥ , with components ( a 1 , a 2 ) = ( a x , a y ) (a_1,a_2)=(a_x,a_y) ( a 1 , a 2 ) = ( a x , a y ) along a right-handed polarization basis e ^ 1 × e ^ 2 = n ^ \hat{\mathbf e}_1\times\hat{\mathbf e}_2=\hat{\mathbf n} e ^ 1 × e ^ 2 = n ^ . The amplitude parameter is a 0 a_0 a 0 . It is the peak of ∣ a ⊥ ∣ |\mathbf{a}_\perp| ∣ a ⊥ ∣ for linear polarization, while for the elliptical family of ellipticity δ \delta δ scanned below that peak is a 0 / 1 + δ 2 a_0/\sqrt{1+\delta^2} a 0 / 1 + δ 2 , or a 0 / 2 a_0/\sqrt{2} a 0 / 2 in the circular case. Finiteness of the pulse enters through one condition,
a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) = 0 , (1) \mathbf{a}_\perp(+\infty)=\mathbf{a}_\perp(-\infty)=0, \tag*{(1)} a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) = 0 , (1)
which is the statement that the transverse electric field has no dc component. The phase average of a ⊥ \mathbf{a}_\perp a ⊥ itself is unconstrained. A pulse carrying ∫ a ⊥ d η ≠ 0 \int\mathbf{a}_\perp\,d\eta\ne0 ∫ a ⊥ d η = 0 obeys everything below.
We treat the spin as a classical four-vector S μ S^\mu S μ with S ⋅ u = 0 S\cdot u=0 S ⋅ u = 0 and S 2 = − 1 S^2=-1 S 2 = − 1 , evolving by the BMT equation [7 , 15 , 16 ] ,
d S μ d τ = e m [ g 2 F μ ν S ν + ( g 2 − 1 ) u μ ( S α F α ν u ν ) ] , (2) \frac{dS^\mu}{d\tau}=\frac{e}{m}\left[\frac{g}{2}F^{\mu\nu}S_\nu+\left(\frac{g}{2}-1\right)u^\mu(S_\alpha F^{\alpha\nu}u_\nu)\right], \tag*{(2)} d τ d S μ = m e [ 2 g F μν S ν + ( 2 g − 1 ) u μ ( S α F α ν u ν ) ] , (2)
with no Stern–Gerlach force, no radiation reaction and no electric dipole moment. The first omission is the mildest: relative to the Lorentz force the Stern–Gerlach force carries a factor ℏ ω / m c 2 ≃ 3 × 10 − 6 \hbar\omega/mc^2\simeq3\times10^{-6} ℏ ω / m c 2 ≃ 3 × 1 0 − 6 for Ti:sapphire light in the rest frame of a slow electron, and its role in classical electron dynamics has been isolated numerically [17 , 18 ] . The other two are genuine restrictions. Section VII weighs them.
Within the plane-wave sector Eq. (2) is not an approximation. Chakrabarti solved the Dirac–Pauli equation exactly for fields of the form F = F ( k ⋅ x ) F=F(k\cdot x) F = F ( k ⋅ x ) with k 2 = 0 k^2=0 k 2 = 0 and found that the polarization four-vector obeys the BMT equation exactly, for an arbitrary anomalous magnetic moment and an arbitrary electric dipole moment [9 ] . Barducci and Giachetti later built the eigenfunctions and the Green function of the same problem [19 ] . Outside the locally constant field approximation the electron–wave interaction is nonlocal and an anomalous magnetic moment cannot be introduced in the usual way [20 ] . We hold a e a_e a e at its vacuum value throughout. That costs a relative O ( χ ) O(\chi) O ( χ ) . It is harmless at the working points of Sec. VII.
The orbit is the exact Volkov solution [21 ] . Because k μ F μ ν = 0 k_\mu F^{\mu\nu}=0 k μ F μν = 0 , the light-front projection κ = k ⋅ u \kappa=k\cdot u κ = k ⋅ u is a constant of the motion, d / d τ = κ d / d η d/d\tau=\kappa\,d/d\eta d / d τ = κ d / d η , and the four-velocity returns to its initial value, u ( ± ∞ ) = u 0 u(\pm\infty)=u_0 u ( ± ∞ ) = u 0 ; no integration is needed for the orbit, which removes one source of numerical error from everything below.
A single antisymmetric generator
Write ( p ∧ q ) μ ν = p μ q ν − q μ p ν (p\wedge q)^{\mu\nu}=p^\mu q^\nu-q^\mu p^\nu ( p ∧ q ) μν = p μ q ν − q μ p ν for the wedge of two four-vectors, which acts on S S S as p ( q ⋅ S ) − q ( p ⋅ S ) p(q\cdot S)-q(p\cdot S) p ( q ⋅ S ) − q ( p ⋅ S ) ; the symbol never denotes a three-dimensional wedge. Every sign statement below is read off that convention. With f = ( e / m ) F = k ∧ a ′ f=(e/m)F=k\wedge a' f = ( e / m ) F = k ∧ a ′ and w μ = f μ ν u ν w^\mu=f^{\mu\nu}u_\nu w μ = f μν u ν , and using u ⋅ S = 0 u\cdot S=0 u ⋅ S = 0 to bring the right-hand side of Eq. (2) to antisymmetric form, one obtains a single generator,
d S μ d η = 1 κ Ω μ ν S ν , Ω = ( 1 + a e ) k ∧ a ′ + a e ( u ∧ w ) . (3) \frac{dS^\mu}{d\eta}=\frac{1}{\kappa}\Omega^\mu{}_\nu S^\nu,\qquad\Omega=(1+a_e)\,k\wedge a'+a_e\,(u\wedge w). \tag*{(3)} d η d S μ = κ 1 Ω μ ν S ν , Ω = ( 1 + a e ) k ∧ a ′ + a e ( u ∧ w ) . (3)
The completion to antisymmetric form is unique, because two antisymmetric matrices that act identically on the three-space orthogonal to u u u differ by an antisymmetric matrix with a three-dimensional kernel, and the rank of an antisymmetric matrix is even.
Expanded in the basis { k , k ˉ , e 1 , e 2 } \{k,\bar{k},e_1,e_2\} { k , k ˉ , e 1 , e 2 } with k ⋅ k ˉ = 2 k\cdot\bar{k}=2 k ⋅ k ˉ = 2 , the anomalous piece u ∧ w u\wedge w u ∧ w has components along the boost k ˉ ∧ k \bar{k}\wedge k k ˉ ∧ k and along k ˉ ∧ e i \bar{k}\wedge e_i k ˉ ∧ e i and also inside the little algebra of k k k . Hence at g ≠ 2 g\ne2 g = 2 the generator leaves that little algebra. The pieces together generate s l ( 2 , R ) \mathfrak{sl}(2,\mathbb R) sl ( 2 , R ) already for linear polarization, and the η \eta η ordering in Eq. (3) is not degenerate.
One simplification comes first. A null rotation N N N from the little group I S O ( 2 ) k ISO(2)_k I S O ( 2 ) k leaves k k k fixed and sends e i ↦ e i + c i k e_i\mapsto e_i+c_i k e i ↦ e i + c i k , so that N f N − 1 = k ∧ ( a ′ + c k ) = f NfN^{-1}=k\wedge(a'+c k)=f N f N − 1 = k ∧ ( a ′ + c k ) = f : the background is invariant, which is stronger than gauge equivalence. Since I S O ( 2 ) k ISO(2)_k I S O ( 2 ) k acts transitively on the mass shell at fixed κ \kappa κ , any u 0 u_0 u 0 can be carried to ( k + k ˉ ) / 2 (k+\bar{k})/2 ( k + k ˉ ) /2 , the net map of the original geometry is the conjugate N − 1 Λ n e t N N^{-1}\Lambda_{\mathrm{net}}N N − 1 Λ net N , and a rotation angle is a conjugation invariant. It is therefore enough to work head-on, normalized to κ = 1 \kappa=1 κ = 1 , and Secs. III and V do so.
Reduction to a connection, and the holonomy
The g = 2 g=2 g = 2 propagator
At a e = 0 a_e=0 a e = 0 the generator is Ω = k ∧ a ′ = a 1 ′ N 1 + a 2 ′ N 2 \Omega=k\wedge a'=a'_1N_1+a'_2N_2 Ω = k ∧ a ′ = a 1 ′ N 1 + a 2 ′ N 2 with N i = k ∧ e i N_i=k\wedge e_i N i = k ∧ e i . Direct multiplication gives ( N i N j ) μ ν = δ i j k μ k ν (N_iN_j)^\mu{}_\nu=\delta_{ij}k^\mu k_\nu ( N i N j ) μ ν = δ ij k μ k ν , which is symmetric in ( i j ) (ij) ( ij ) , so [ N 1 , N 2 ] = 0 [N_1,N_2]=0 [ N 1 , N 2 ] = 0 and N i 3 = 0 N_i^3=0 N i 3 = 0 ; the generator therefore commutes with itself at different phases, the ordering collapses, and
Λ 0 ( η ) = exp [ κ − 1 ( a 1 N 1 + a 2 N 2 ) ] , Λ 0 ( + ∞ ) = I , (4) \Lambda_0(\eta)=\exp\left[\kappa^{-1}(a_1N_1+a_2N_2)\right],\qquad\Lambda_0(+\infty)=\mathbb{I}, \tag*{(4)} Λ 0 ( η ) = exp [ κ − 1 ( a 1 N 1 + a 2 N 2 ) ] , Λ 0 ( + ∞ ) = I , (4)
by Eq. (1) . Hence no property of u 0 u_0 u 0 enters: u u u does not appear in Ω \Omega Ω at a e = 0 a_e=0 a e = 0 at all. The mechanism is not ours. It is the one Kupersztych identified: the operator that transports an electron and its spin through a plane wave is of Lorentz type and acts on the wave as a gauge transformation [22 ] . In modern language the generator sits in the Abelian translation subgroup of the little group of a null vector [23 ] , and Eq. (1) closes the loop.
Interaction picture and the exact reduction
Since ( 1 + a e ) f + a e u ∧ w = f + a e ( f + u ∧ w ) (1+a_e)f+a_eu\wedge w=f+a_e(f+u\wedge w) ( 1 + a e ) f + a e u ∧ w = f + a e ( f + u ∧ w ) , factoring the evolution as U = Λ 0 V U=\Lambda_0V U = Λ 0 V gives an interaction-picture equation in which the anomaly is the only coupling,
d V d η = a e M ( η ) V , M = Ad Λ 0 − 1 ( f + u ∧ w ) . (5) \frac{dV}{d\eta}=a_e\,M(\eta)V,\qquad M=\operatorname{Ad}_{\Lambda_0^{-1}}(f+u\wedge w). \tag*{(5)} d η d V = a e M ( η ) V , M = Ad Λ 0 − 1 ( f + u ∧ w ) . (5)
Write A = k ∧ a ⊥ = a 1 N 1 + a 2 N 2 A=k\wedge a_\perp=a_1N_1+a_2N_2 A = k ∧ a ⊥ = a 1 N 1 + a 2 N 2 , so that Λ 0 = e A \Lambda_0=e^A Λ 0 = e A and M = e − ad A X M=e^{-\operatorname{ad}_A}X M = e − ad A X with X = f + u ∧ w X=f+u\wedge w X = f + u ∧ w . The series terminates after three terms: on the whole of s o ( 1 , 3 ) \mathfrak{so}(1,3) so ( 1 , 3 ) the operator A A A raises the grading k ∧ e i → { k ˉ ∧ k , e 1 ∧ e 2 } → k ∧ e i → 0 k\wedge e_i\to\{ \bar{k}\wedge k,e_1\wedge e_2\}\to k\wedge e_i\to0 k ∧ e i → { k ˉ ∧ k , e 1 ∧ e 2 } → k ∧ e i → 0 , so ad A 3 = 0 \operatorname{ad}_A^3=0 ad A 3 = 0 and M = X − [ A , X ] + 1 2 [ A , [ A , X ] ] M=X-[A,X]+\frac{1}{2}[A,[A,X]] M = X − [ A , X ] + 2 1 [ A , [ A , X ]] .
Two cancellations follow, and they are of different kinds. Everything along the boost k ˉ ∧ k \bar{k}\wedge k k ˉ ∧ k and along e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 is removed by the first commutator, with no input beyond the algebra, while the residual null rotations along k ∧ e i k\wedge e_i k ∧ e i are removed by the two algebraic identities
U a 1 − W a 2 = ∣ a ⊥ ∣ 2 a 1 ′ , U a 2 + W a 1 = ∣ a ⊥ ∣ 2 a 2 ′ , (6) Ua_1-Wa_2=|a_\perp|^2a'_1,\qquad Ua_2+Wa_1=|a_\perp|^2a'_2, \tag*{(6)} U a 1 − W a 2 = ∣ a ⊥ ∣ 2 a 1 ′ , U a 2 + W a 1 = ∣ a ⊥ ∣ 2 a 2 ′ , (6)
with U = a ⊥ ⋅ a ⊥ ′ U=a_\perp\cdot a'_\perp U = a ⊥ ⋅ a ⊥ ′ and W = a 1 a 2 ′ − a 2 a 1 ′ W=a_1a'_2-a_2a'_1 W = a 1 a 2 ′ − a 2 a 1 ′ , which follow from those definitions by inspection. What survives is a pure rotation of the rest frame about an axis lying in the polarization plane,
M ( η ) d η = n ^ ∧ d a ⊥ , d ζ d η = a e ( n ^ × a ⊥ ′ ) × ζ , (7) M(\eta)\,d\eta=\hat{n}\wedge da_\perp,\qquad\frac{d\zeta}{d\eta}=a_e\left(\hat{n}\times a'_\perp\right)\times\zeta, \tag*{(7)} M ( η ) d η = n ^ ∧ d a ⊥ , d η d ζ = a e ( n ^ × a ⊥ ′ ) × ζ , (7)
with ζ \zeta ζ the rest-frame spin and n ^ = 1 2 ( k / κ − κ k ˉ ) \hat{n}=\frac{1}{2}(k/\kappa-\kappa\bar{k}) n ^ = 2 1 ( k / κ − κ k ˉ ) the unit spacelike vector along the wave in the rest frame of u 0 u_0 u 0 . In the convention fixed at Eq. (3) the sign in Eq. (7) is positive, the curvature is [ n ^ ∧ e 1 , n ^ ∧ e 2 ] = + e 1 ∧ e 2 [\hat{n}\wedge e_1,\hat{n}\wedge e_2]=+e_1\wedge e_2 [ n ^ ∧ e 1 , n ^ ∧ e 2 ] = + e 1 ∧ e 2 , and e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 generates a right-handed rotation about n ^ \hat{n} n ^ . Table S1 of the Supplemental Material carries the three terms of e − ad A X e^{-\operatorname{ad}_A}X e − ad A X out at one numerical point and shows where each cancellation happens, and we also checked Eq. (7) against the components of Eq. (3) on a phase grid, where the largest discrepancy is 2.7 × 10 − 15 2.7\times10^{-15} 2.7 × 1 0 − 15 .
Equation (7) is a connection one-form on the polarization plane, and its coefficients are constants. Two consequences are immediate. The phase enters only through the path, so the ordered exponential of a e M a_eM a e M depends on the curve C : η ↦ ( a x , a y ) \mathcal{C}:\eta\mapsto(a_x,a_y) C : η ↦ ( a x , a y ) alone and not on the rate at which it is traced. And the integral of a one-form with constant coefficients along a path depends on the endpoints alone, so
a e ∫ M d η = a e n ^ ∧ Δ a ⊥ = 0 (8) a_e\int M\,d\eta=a_e\hat{n}\wedge\Delta a_\perp=0 \tag*{(8)} a e ∫ M d η = a e n ^ ∧ Δ a ⊥ = 0 (8)
by Eq. (1) , exactly, at any a 0 a_0 a 0 and for any curve. Hence the first order in the anomaly cancels for every polarization. A weaker fact is sometimes offered in its place and does not do the work. That Λ n e t \Lambda_{\mathrm{net}} Λ net lies in the stabilizer S O ( 3 ) u 0 SO(3)_{u_0} S O ( 3 ) u 0 follows from u ( ± ∞ ) = u 0 u(\pm\infty)=u_0 u ( ± ∞ ) = u 0 and the conservation of S ⋅ u S\cdot u S ⋅ u , but it confines the first-order coefficient to a three-dimensional space and constrains it no further. Section VI makes the distinction concrete.
For linear polarization the axis n ^ × a ⊥ ′ \hat{n}\times a'_\perp n ^ × a ⊥ ′ is fixed, the rotations at different phases commute, and V ( + ∞ ) = exp [ a e Δ a 1 n ^ ∧ e 1 ] = I V(+\infty)=\exp[a_e\Delta a_1\hat{n}\wedge e_1]=\mathbb{I} V ( + ∞ ) = exp [ a e Δ a 1 n ^ ∧ e 1 ] = I exactly at any g g g and any a 0 a_0 a 0 . The conjugation of Sec. II B carries that conclusion to any geometry, since N − 1 I N = I N^{-1}\mathbb{I}N=\mathbb{I} N − 1 I N = I . This is perturbative in nothing, and it agrees with the closed-form spin dynamics obtained for intense linearly polarized fields [24 ] .
Non-Abelian Stokes evaluation
The base is the polarization plane R 2 \mathbb{R}^2 R 2 , the bundle is the trivial principal bundle R 2 × S O ( 3 ) u 0 \mathbb{R}^2\times SO(3)_{u_0} R 2 × S O ( 3 ) u 0 , and M u 0 = 0 M_{u_0}=0 M u 0 = 0 . The corresponding Lie algebra because M u 0 = 0 M_{u_0}=0 M u 0 = 0 . The connection one-form is ω = a e ( G 1 d a 1 + G 2 d a 2 ) \omega=a_e(G_1\,da_1+G_2\,da_2) ω = a e ( G 1 d a 1 + G 2 d a 2 ) with G i = n ^ ∧ e i G_i=\hat{n}\wedge e_i G i = n ^ ∧ e i , and its curvature is constant over the plane,
F = d ω + ω ∧ ω = a e 2 ( e 1 ∧ e 2 ) d a 1 ∧ d a 2 . (9) F=d\omega+\omega\wedge\omega=a_e^2(e_1\wedge e_2)\,da_1\wedge da_2. \tag*{(9)} F = d ω + ω ∧ ω = a e 2 ( e 1 ∧ e 2 ) d a 1 ∧ d a 2 . (9)
The path C \mathcal{C} C closes for exactly one reason, Eq. (1) , and that is the only place where finiteness of the pulse is used. Thus the net map is the holonomy Λ n e t = P exp ∮ C ω \Lambda_{\mathrm{net}}=\mathcal{P}\exp\oint_{\mathcal{C}}\omega Λ net = P exp ∮ C ω . No adiabatic parameter appears and no eigenspace is being followed, so this is ordinary non-Abelian parallel transport, and the resemblance to the geometric phases of Refs. [25 , 26 ] is an analogy.
The surface-ordered form of the non-Abelian Stokes theorem [27 , 28 ] turns the ordered exponential around C \mathcal{C} C into a surface-ordered exponential of the curvature carried back to a base point along a mesh of paths. Constancy of F F F in the coordinates of the base does not make that ordering trivial, because what is ordered is not F F F but its adjoint transport along the mesh. The theorem delivers the leading term and nothing more: a rotation about n ^ \hat{n} n ^ through a e 2 a_e^2 a e 2 times the oriented area enclosed by C. The second Magnus term [29 , 30 ] gives the same result in three lines and fixes the sign, and the two together give
Θ n e t = − 1 2 a e 2 A [ 1 + O ( a e 2 a 0 2 ) ] , A = ∫ ( a x a y ′ − a y a x ′ ) d η (10) \Theta_{\rm net}=-\frac{1}{2}a_{\rm e}^{2}\mathcal{A}\left[1+O(a_{\rm e}^{2}a_{0}^{2})\right],\qquad\mathcal{A}=\int(a_xa_y'-a_ya_x')\,d\eta \tag*{(10)} Θ net = − 2 1 a e 2 A [ 1 + O ( a e 2 a 0 2 ) ] , A = ∫ ( a x a y ′ − a y a x ′ ) d η (10)
in the right-handed triad e ^ 1 × e ^ 2 = n ^ \hat{\mathbf{e}}_1 \times\hat{\mathbf{e}}_2=\hat{\mathbf{n}} e ^ 1 × e ^ 2 = n ^ , and Sec. S2 of the Supplemental Material carries both evaluations out in full.
Two definitions have to be pinned down before Eq. (10) can be read literally. First, A \mathcal{A} A is twice the signed area enclosed by C C C , counted with multiplicity, so a curve that winds N N N times about the origin contributes N N N times over. Second, we take Θ n e t \Theta_{\rm net} Θ net to be the angle lifted continuously in a e a_{\rm e} a e from Θ n e t = 0 \Theta_{\rm net}=0 Θ net = 0 at a e = 0 a_{\rm e}=0 a e = 0 . The angle read off an S O ( 3 ) SO(3) S O ( 3 ) matrix is its reduction modulo 2 π 2\pi 2 π , and the two coincide while ∣ Θ n e t ∣ < π |\Theta_{\rm net}|<\pi ∣ Θ net ∣ < π .
The area is gauge invariant. Within the plane-wave sector the freedom left is a μ → a μ + c μ a^\mu\to a^\mu+c^\mu a μ → a μ + c μ with c c c constant and k ⋅ c = 0 k\cdot c=0 k ⋅ c = 0 , the gauge function being c ⋅ x c\cdot x c ⋅ x . The field tensor f = k ∧ a ′ f=k\wedge a' f = k ∧ a ′ does not see it. The transverse part of c c c translates C C C bodily in the polarization plane, and the signed area of a closed curve does not change under a translation. Equation (1) removes the freedom altogether in any case, since it picks out the one representative that vanishes at both ends of the pulse.
The electric pulse area of the unipolar-pulse literature [31 , 32 ] is a different functional: ∫ E d t \int E\,dt ∫ E d t is linear in the field and vanishes for every pulse obeying Eq. (1), while A \mathcal{A} A is quadratic and stays finite there. For linear polarization C C C degenerates to a segment, a segment encloses no area, and A = 0 \mathcal{A}=0 A = 0 . The vanishing there is exact at every order, by the fixed-axis argument of Sec. III B and not by Eq. (10). The same holds for any curve traced out and back, such as a y ∝ a x 2 a_y\propto a_x^2 a y ∝ a x 2 .
The area as the spin angular momentum of the pulse
In the Coulomb gauge the spin part of the electromagnetic angular momentum has density ϵ 0 E × A \epsilon_0\mathbf{E}\times\mathbf{A} ϵ 0 E × A . For a plane wave written in terms of the dimensionless potential, A = ( m / e ) a ⊥ \mathbf{A}=(m/e)\mathbf{a}_{\perp} A = ( m / e ) a ⊥ and E = − ∂ t A = − ω ( m / e ) a ⊥ ′ \mathbf{E}=-\partial_t\mathbf{A}=-\omega(m/e)\mathbf{a}_{\perp}' E = − ∂ t A = − ω ( m / e ) a ⊥ ′ , so that
ϵ 0 ∫ ( E × A ) d t = ϵ 0 ( m e ) 2 A n ^ . (11) \epsilon_0\int(\mathbf{E}\times\mathbf{A})\,dt=\epsilon_0\left(\frac{m}{e}\right)^2\mathcal{A}\hat{\mathbf{n}}. \tag*{(11)} ϵ 0 ∫ ( E × A ) d t = ϵ 0 ( e m ) 2 A n ^ . (11)
Thus A \mathcal{A} A is, up to a positive constant, the spin angular momentum the pulse carries per unit cross-sectional area, and Eq. (10) reads: the net rotation of the electron spin is the helicity of the pulse times the square of the anomaly. The identification is not ours. It is the reading given to the same functional in the orientation memory of a magnetic dipole [13 ] , where the area is the difference between the numbers of right- and left-handed photons in the burst. Equation (11) also separates the present effect from the Pauli coupling σ ⋅ ( E × A ) \boldsymbol{\sigma}\cdot(\mathbf{E}\times\mathbf{A}) σ ⋅ ( E × A ) , which rotates the spin already at g = 2 g=2 g = 2 in two counterpropagating elliptically polarized beams [33 ] and is quantum in origin.
Order of the remainder, and a resummation
The order of the remainder in Eq. (10) is a theorem, and the proof takes two lines. Grade s o ( 3 ) u \mathfrak{so}(3)_{u} so ( 3 ) u by calling G 1 G_1 G 1 and G 2 G_2 G 2 odd and ϵ 1 ∧ ϵ 2 \epsilon_1\wedge\epsilon_2 ϵ 1 ∧ ϵ 2 even. The bracket respects that Z 2 \mathbb{Z}_2 Z 2 grading, since [ o d d , o d d ] = e v e n [\mathrm{odd},\mathrm{odd}]=\mathrm{even} [ odd , odd ] = even and [ e v e n , o d d ] = o d d [\mathrm{even},\mathrm{odd}]=\mathrm{odd} [ even , odd ] = odd . Every term of the Magnus series is an integral of n n n -fold nested brackets of M M M , and M M M is odd, so the n n n th term has parity n n n . Three consequences follow: the component of the exponent along n ^ \hat{\mathbf{n}} n ^ carries only even powers of a e a_{\rm e} a e , which is the O ( a e 4 ) O(a_{\rm e}^{4}) O ( a e 4 ) correction to Eq. (10); the leading correction to the axis is the third-order in-plane term, so that at fixed curve the axis is tilted out of n ^ \hat{\mathbf{n}} n ^ by O ( a e ) O(a_{\rm e}) O ( a e ) ; and the change in the projection of the spin on n ^ \hat{\mathbf{n}} n ^ is quadratic in that tilt, hence O ( a e 6 ) O(a_{\rm e}^{6}) O ( a e 6 ) . The helicity statement of Sec. I rests on the last of the three. That one is measured. With the spin started along n ^ \hat{\mathbf{n}} n ^ and the anomaly inflated so that a e a 0 a_{\rm e}a_0 a e a 0 runs from 0.03 to 0.5, the change in the longitudinal projection follows Δ S ∥ = − 1.80 × 10 − 3 ( a e a 0 ) 6 \Delta S_{\parallel}=-1.80\times10^{-3}(a_{\rm e}a_0)^6 Δ S ∥ = − 1.80 × 1 0 − 3 ( a e a 0 ) 6 , with local log-log exponents from 6.00 to 5.98 at the small-annoy end. The coefficient repeats to 0.5% at a 0 = 1 a_0=1 a 0 = 1 , 3 and 10. The dependence is on the product a e a 0 a_{\rm e}a_0 a e a 0 alone. Over the same range the angle scales as a e 2.000 a_{\rm e}^{2.000} a e 2.000 and the tilt ψ \psi ψ of the axis out of n ^ \hat{\mathbf{n}} n ^ as a e 0.996 a_{\rm e}^{0.996} a e 0.996 , indicating Δ S ∥ = − ( 1 − cos Θ n e t ) sin 2 ψ \Delta S_{\parallel}=-(1-\cos\Theta_{\rm net})\sin^2\psi Δ S ∥ = − ( 1 − cos Θ net ) sin 2 ψ holds to 5 × 10 − 3 5\times10^{-3} 5 × 1 0 − 3 , so the sixth power is the product of the two factors the grading predicts. Two limits attach to it. For a spin that does not start along n ^ \hat{\mathbf{n}} n ^ the projection changes at O ( a e 3 ) O(a_{\rm e}^{3}) O ( a e 3 ) instead, measured exponent 3.00, so helicity is conserved to sixth order as a property of the helicity state. The coefficient itself was measured at N = 1 N=1 N = 1 , γ = 1 \gamma=1 γ = 1 and circular polarization. It does not carry to the working point a 0 = 75 a_0=75 a 0 = 75 , N = 32 N=32 N = 32 of Sec. VII.
One curve does better than the order counting. On a plateau where C C C is a circle of radius a 0 a_0 a 0 traced at unit rate, n ^ × a ⊥ ′ \hat{\mathbf{n}}\times\mathbf{a}_{\perp}' n ^ × a ⊥ ′ has fixed magnitude a 0 a_0 a 0 and rotates about n ^ \hat{\mathbf{n}} n ^ at unit rate, and Eq. (7) becomes the magnetic-resonance problem for a rotating field of strength a e a 0 a_{\rm e}a_0 a e a 0 on exact resonance. In the corotating frame the rotation vector is constant, of length 1 + a e 2 a 0 2 \sqrt{1+a_{\rm e}^{2}a_0^{2}} 1 + a e 2 a 0 2 and tilted from n ^ \hat{\mathbf{n}} n ^ by arctan ( a e a 0 ) \arctan(a_{\rm e}a_0) arctan ( a e a 0 ) , and returning to the laboratory frame subtracts one turn per unit phase,
Θ ≃ ( 1 + a e 2 a 0 2 − 1 ) A a 0 2 . (12) \Theta\simeq\left(\sqrt{1+a_{\rm e}^{2}a_0^{2}}-1\right)\frac{\mathcal{A}}{a_0^2}. \tag*{(12)} Θ ≃ ( 1 + a e 2 a 0 2 − 1 ) a 0 2 A . (12)
Three conditions attach to Eq. (12): the idealized circle violates Eq. (1) and is therefore not an admissible pulse; composing two rotations about different axes returns a single angle only for an integer number of turns; and for a pulse with fronts the phase interval has to be read as A / a 0 2 \mathcal{A}/a_0^2 A / a 0 2 , which is the form written above, since reading it as the plateau length instead costs 19% at a 0 = 1 a_0=1 a 0 = 1 , a e = 0.05 a_{\rm e}=0.05 a e = 0.05 . The residual error is real. At $a_{\rm e}a_0=0.3 Eq. (12) is still off by 0.1%, so it improves the asymptotics without becoming exact. The word exact does not belong to it. The expansion parameter is a e a 0 a_{\mathrm{e}}a_0 a e a 0 , not a e a_{\mathrm{e}} a e . With the physical anomaly a e a 0 = 1.16 × 10 − 3 a 0 a_{\mathrm{e}}a_0 = 1.16 \times10^{-3}a_0 a e a 0 = 1.16 × 1 0 − 3 a 0 , and Eq. (10) stays accurate to 10 − 2 10^{-2} 1 0 − 2 up to a 0 ∼ 100 a_0 \sim100 a 0 ∼ 100 for a circular curve. For a strongly elliptical or chirped curve the remainder depends on the shape of C \mathcal{C} C and not on its area alone, and we have not bounded it there.
Relation to earlier work
Three separate lines of work reach parts of Eq. (10) , and the boundaries between them and the present paper are sharp enough to be drawn one by one.
The exact plane-wave spin problem
The Dirac–Pauli equation for a particle with an anomalous moment in a plane wave of arbitrary profile was solved exactly in the 1960s. Ternov, Bagrov and Klimenko treated a circularly polarized wave and read off the shift of the spin precession frequency caused by the anomaly [8 ] . Chakrabarti obtained the polarization directly in integrated form for a class of fields F = F ( k ⋅ x ) F = F(k \cdot x) F = F ( k ⋅ x ) , with an anomalous magnetic moment and an anomalous electric dipole moment both arbitrary [9 ] . The material is collected in the monograph of Bagrov and Gitman, Sec. 5.3 of Ref. [11 ] and Chap. 8 of the earlier edition [10 ] , and four of its statements bear directly on this work.
First, after the Volkov factor is separated the residual spinor equation is driven by the anomalous moment alone, with the wave entering through a ⊥ ′ ( η ) a'_{\perp}(\eta) a ⊥ ′ ( η ) , which is the content of Eq. (5) above, in quantum form and three decades earlier. Second, the first order in the anomaly is written there in closed form, and the operator is an algebraic function of a ⊥ ( η ) a_{\perp}(\eta) a ⊥ ( η ) and not an integral of it, so that our Eq. (8) is one line away from that expression, although the asymptotic step and the statement about a pulse are not taken there. That step is short, and it was not taken in that line. Third, linear polarization of arbitrary profile is solved exactly, the evolution operator being a function of the instantaneous a ⊥ ( η ) a_{\perp}(\eta) a ⊥ ( η ) . For a pulse that operator is the identity, which is the exact zero quoted in Sec. III B. Fourth, a potential modulus with an arbitrary phase function is solved exactly, and the answer depends on the total phase swept and not on the phase function, which is Eq. (12) together with its independence of chirp. The same reduction was found independently three decades later, for electromagnetic and gravitational pp-wave backgrounds, by Balakin, Kurbanova and Zimdahl [12 ] . They state explicitly that nothing rotates at g = 2 g = 2 g = 2 .
One thing that line does not contain is the general second order. Bagrov and Gitman say plainly that two cases are known for which the residual equation is solved exactly, and those are the two just listed; neither an asymptotic limit for a pulse of finite duration nor an observable net rotation is formulated anywhere in that literature. The language throughout is one of wave functions, quantum numbers and almost-periodicity. Equation (10) is the second order for an elliptical curve of arbitrary profile, and the two exactly solvable cases are its degenerate limits: for linear polarization the curve collapses and A = 0 \mathcal{A} = 0 A = 0 , while for a circular plateau the area is known in closed form and Eq. (10) is the leading term of Eq. (12) .
One consequence of that reading is worth recording. In the Dirac–Pauli problem the anomalous magnetic and electric moments enter through a single modulus and a constant mixing angle, and a constant rotation in the transverse plane leaves the signed area unchanged. Equation (10) therefore holds for a particle carrying both, with a e a_{\mathrm{e}} a e replaced by the modulus of the combined anomalous moment.
The vanishing at g = 2 g = 2 g = 2
That a plane wave leaves the asymptotic spin of a g = 2 g = 2 g = 2 electron untouched is old and well recorded. Kuperzstych derived the underlying gauge mechanism in 1976 [22 ] . Aleksandrov et al. found that the spin change in a linearly polarized pulse is proportional to the electric pulse area ∫ E d t \int E\,\mathrm{d}t ∫ E d t [31 ] , which is zero for any pulse with a finite envelope. For arbitrary polarization the same statement follows in one line from the Volkov solution [21 ] . Ilderton, King and Tang proved it in the quantum theory and wrote the first order in the anomaly explicitly: the amplitude for reversing the spin state is ∣ c − ∣ = 1 2 ∣ ∫ d η μ b ( a ⊥ ′ − i a ⊥ ′ ) ∣ |c_{-}| = \frac{1}{2}\left|\int\mathrm{d}\eta\,\mu_b\left(a'_{\perp} - \mathrm{i}a'_{\perp}\right)\right| ∣ c − ∣ = 2 1 ∫ d η μ b ( a ⊥ ′ − i a ⊥ ′ ) , with μ b \mu_b μ b their phase-dependent anomaly and the light-front integral running over the whole pulse [34 ] . The result is stated verbatim in the review literature [5 , 6 ] . Section III A adds nothing to it and cites it as the starting point.
The delimitation against Ref. [34 ] needs care, because their vanishing condition is parity: the amplitude is zero when the potential is an even function of the phase. Unipolarity enters only in the weak-field limit. For a pulse of zero electric area that is odd in the phase, produced by a shifted carrier-envelope phase, a chirp or an asymmetric envelope, their first order survives and can exceed the holonomy. By how much we have not evaluated. A constant μ b \mu_b μ b returns μ b Δ a ⊥ = 0 \mu_b \Delta a_{\perp} = 0 μ b Δ a ⊥ = 0 by Eq. (1) , so what is left of their integral is the variation of μ b \mu_b μ b across the pulse, and that variation is itself of order χ \chi χ . Their expression is first order in a dressed anomaly μ b ( χ ) \mu_b(\chi) μ b ( χ ) . Ours is second order in the vacuum anomaly, and it does not vanish when theirs does. The two are complementary. The frame common to both, that non-radiative effects dominate at small χ \chi χ , is theirs.
Area, helicity and memory
The functional A \mathcal{A} A and its identification with helicity are also in print, in a literature with no overlap with strong-field physics. Oblak and Seraj computed the net reorientation of a magnetic dipole after a burst of radiation has passed. They found a second-order term equal to the signed area swept by the transverse potential, and they identify that area as the optical helicity of the burst [13 ] . Seraj and Neogi phrase memory observables of this kind as holonomies [14 ] . Up to the coupling constant, their nonlinear term and Eq. (10) are the same expression, and we take both the area law and the helicity reading from them, so that Sec. III D is written accordingly.
The physics is different, and the difference is the point. Their test body is a neutral, static, nonrelativistic dipole obeying M ˙ i = k F i j M j \dot{M}^{i} = \mathrm{k}\,F^{ij}M_{j} M ˙ i = k F ij M j , and the coupling k \mathrm{k} k is the full gyromagnetic ratio, a number of order unity: there is no charge, no orbit, no Thomas precession and no g g g factor. Substituting an electron into that formula with k = g / 2 \mathrm{k} = g/2 k = g /2 overestimates the rotation by a factor ( g / 2 ) 2 / a e 2 = 7.4 × 10 5 (g/2)^2/a_{\mathrm{e}}^2 = 7.4 \times10^5 ( g /2 ) 2 / a e 2 = 7.4 × 1 0 5 . For a charged relativistic electron on an exact Volkov orbit the terms proportional to g / 2 g/2 g /2 cancel identically against the boost and Thomas parts of the generator, as Eq. (5) shows, and what remains couples through a e a_{\mathrm{e}} a e alone. Hence their result cannot be applied to an electron before that reduction is proved, and ours does not follow from theirs. The reverse implication also fails. Their formula carries a term linear in the field, the radial magnetic field at null infinity, which is identically zero for an exact plane wave.
The suppression is not a defect. A g g g -independent background of the same geometric form is absent by Eq. (4) , so in an ideal plane wave the entire net rotation is anomalous. The factor 1.3 × 10 − 6 1.3 \times10^{-6} 1.3 × 1 0 − 6 is the price of a channel with nothing to subtract. Section VI shows how a finite waist puts the background back.
Schemes that do rotate the spin
Two published mechanisms produce a net rotation of the same observable and must not be confused with this one. Wei et al. rotate the transverse polarization of a relativistic beam through an arbitrary angle with a copropagating chirped or subcycle pulse [35 ] . Their pulses are temporally asymmetric, which is a ⊥ ( + ∞ ) ≠ a ⊥ ( − ∞ ) a_{\perp}(+\infty) \ne a_{\perp}(-\infty) a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) and breaks Eq. (1) at order a e a_{\mathrm{e}} a e . The dominant coupling in their analytic estimate is the kinematic 1 / γ 1/\gamma 1/ γ and not the anomaly. Their result is chirp dependent while ours is not, and there is no contradiction. What depends on the chirp there is the accumulated precession phase, whereas the invariant here is the net rotation relative to the momentum, that is, the difference between the precession phase and the momentum phase. Tikhomirov showed that a phase velocity v p h ≠ c v_{\mathrm{ph}} \ne c v ph = c produces a spin precession exceeding radiative effects by orders of magnitude [36 , 37 ] ; that effect vanishes at v p h = c v_{\mathrm{ph}} = c v ph = c , so it confirms the plane-wave theorem from the other side and identifies a medium as a way of breaking it.
What is new here
Four items remain, and they are the claims of this paper. (i) The net rotation for a pulse of finite duration and arbitrary elliptical profile is − 1 2 a e 2 A -\frac{1}{2}a_{\mathrm{e}}^2\mathcal{A} − 2 1 a e 2 A , Eq. (10) , which is the general second order that the two known exact cases bound. (ii) The geometric reading: holonomy of a connection with constant coefficients on the polarization plane, a signed area evaluated by the non-Abelian Stokes theorem, and a Z 2 \mathbb{Z}_2 Z 2 grading that fixes the remainder at O ( a e 4 ) O(a_{\mathrm{e}}^4) O ( a e 4 ) , the axis tilt at O ( a e ) O(a_{\mathrm{e}}) O ( a e ) and the helicity change at O ( a e 6 ) O(a_{\mathrm{e}}^6) O ( a e 6 ) . (iii) The transfer of the area mechanism from a neutral dipole to a charged relativistic electron, where the correct coupling is the anomaly and not g / 2 g/2 g /2 . That is the 1.3 × 10 − 6 1.3 \times10^{-6} 1.3 × 1 0 − 6 suppression, and it is also the absence of a g g g -independent background. (iv) Invariance under reparameterization of the phase, together with the quantitative bound on finite focusing of Sec. VI. Everything else in Secs. II and III we take from the works named above.
Numerical verification
We integrated Eq. (2) along the analytic Volkov orbit with the DOP853 scheme at r t o l = 10 − 13 \mathrm{rtol} = 10^{-13} rtol = 1 0 − 13 and a t o l = 10 − 16 \mathrm{atol} = 10^{-16} atol = 1 0 − 16 , transporting an orthonormal rest-frame triad and reading the net rotation off the resulting matrix. Reading the angle from a transported frame is what makes the result independent of the initial polarization, and a second implementation in covariant bivector form, run at a looser tolerance, returns the same zeros on a floor four to seven orders coarser. At the end of the pulse ∣ S ⋅ u ∣ |S \cdot u| ∣ S ⋅ u ∣ and ∣ S ⋅ S + 1 ∣ |S \cdot S + 1| ∣ S ⋅ S + 1∣ stay below 3 × 10 − 12 3 \times10^{-12} 3 × 1 0 − 12 at γ = 1 \gamma= 1 γ = 1 in the first scan and below 6 × 10 − 11 6 \times10^{-11} 6 × 1 0 − 11 in the fit of Figure 1 . Both floors grow with γ \gamma γ , approaching γ 2 \gamma^2 γ 2 above γ ≃ 10 2 \gamma\simeq10^2 γ ≃ 1 0 2 , while the measured Θ n e t \Theta_{\mathrm{net}} Θ net at g = 2 g = 2 g = 2 does not move with γ \gamma γ at all.
Figure 1. Net spin rotation in a plane-wave pulse against the signed area of the vector potential. Each symbol is one solution of Eq. (2) along the exact Volkov orbit, and the solid line is Eq. (10) with no fitted parameter. The scan covers Gaussian and cos 2 \cos^2 cos 2 envelopes, N = 0.5 N = 0.5 N = 0.5 –16 cycles, chirped and polarization-gated pulses, ellipticities δ = 0.05 \delta= 0.05 δ = 0.05 –1, carrier-envelope phases over [ 0 , 2 π ) [0, 2\pi) [ 0 , 2 π ) and Lorentz factors γ = 1 \gamma= 1 γ = 1 –10 4 10^4 1 0 4 , with the anomaly set to a e = 10 − 2 a_e = 10^{-2} a e = 1 0 − 2 so that the rotation stands above the roundoff floor. The g = 2 g = 2 g = 2 null family of the text is a separate set of runs covering γ ≤ 5 × 10 3 \gamma\le5 \times10^3 γ ≤ 5 × 1 0 3 . Color is the ellipticity ∣ δ ∣ |\delta| ∣ δ ∣ of the pulse, and for the polarization-gated family that of the residual circular component, ∣ ( 1 − q ) / ( 1 + q ) ∣ |(1-q)/(1+q)| ∣ ( 1 − q ) / ( 1 + q ) ∣ with q q q the amplitude ratio of the two counter-rotating pulses; it adds nothing beyond the scan parameters. Open symbols are the two families for which A = 0 \mathcal{A}=0 A = 0 : linearly polarized pulses, and the zero-area curve a y ∝ a x 2 a_y \propto a_x^2 a y ∝ a x 2 traced out and back. All 42 points lie between 4.1 × 10 − 17 4.1 \times10^{-17} 4.1 × 1 0 − 17 and 9.95 × 10 − 12 9.95 \times10^{-12} 9.95 × 1 0 − 12 rad for anomalies up to a e = 0.6 a_e = 0.6 a e = 0.6 , the two maxima being 1.29 × 10 − 12 1.29 \times10^{-12} 1.29 × 1 0 − 12 (linear) and 9.95 × 10 − 12 9.95 \times10^{-12} 9.95 × 1 0 − 12 (zero area). Inset: the ratio Θ n e t / Θ l a w \Theta_{\mathrm{net}}/\Theta_{\mathrm{law}} Θ net / Θ law , Θ l a w = − 1 2 a e 2 A \Theta_{\mathrm{law}}=-\frac{1}{2}a_e^2\mathcal{A} Θ law = − 2 1 a e 2 A , against a e a 0 a_ea_0 a e a 0 at a 0 = 1 a_0=1 a 0 = 1 for a circularly polarized flat-top pulse whose plateau is a circle of radius a 0 a_0 a 0 , together with the resummation of Eq. (12).
Figure 1 collects the scan. Over 180 configurations at g = 2 g = 2 g = 2 , covering two envelopes, γ \gamma γ from 1 to 5 × 10 3 5 \times10^3 5 × 1 0 3 , a 0 a_0 a 0 from 0.1 to 10, and three carrier-envelope phases, the largest net rotation is below 3 × 10 − 13 3 \times10^{-13} 3 × 1 0 − 13 rad. At a 0 = 10 a_0 = 10 a 0 = 10 that is the roundoff floor. Section S4.3 of the Supplementary Material shows the tolerance scan that establishes it. For linear polarization the same floor persists once the anomaly is switched on and inflated by up to four orders of magnitude above its physical value, and at γ ≤ 10 2 \gamma\le10^2 γ ≤ 1 0 2 the residue stays at or below 4.1 × 10 − 11 4.1 \times10^{-11} 4.1 × 1 0 − 11 rad.
Away from linear polarization, Eq. (10) reproduces the measured angle to three digits across the scan of Figure 1 and to five in the ellipticity scan. The ratio Θ n e t / a e 2 \Theta_{\mathrm{net}}/a_{\mathrm{e}}^2 Θ net / a e 2 is constant to 0.8 % 0.8\% 0.8% across a 26-fold range in a e a_{\mathrm{e}} a e , the scaling Θ n e t ∝ δ / ( 1 + δ 2 ) \Theta_{\mathrm{net}} \propto\delta/(1+\delta^2) Θ net ∝ δ / ( 1 + δ 2 ) in the ellipticity holds to five digits, and the rotation grows linearly with pulse length. Envelope shape acts through the area and through nothing else, and the effect is measurable: at a 0 = 10 a_0=10 a 0 = 10 and N = 1 N=1 N = 1 the Gaussian gives A = 334.412 \mathcal{A}=334.412 A = 334.412 against 323.603 323.603 323.603 for cos 2 \cos^2 cos 2 , a difference of 3.2 % 3.2\% 3.2% that propagates straight into Θ n e t \Theta_{\mathrm{net}} Θ net . Reparameterizing the phase at fixed curve leaves the net rotation unchanged to eight or nine digits, even at the inflated value a e = 0.25 a_e=0.25 a e = 0.25 . That invariance holds by construction, so it tests the code and not the theorem.
The absence of any γ \gamma γ dependence is kinematic. A null rotation removes the transverse initial momentum while leaving k k k , κ \kappa κ and the field tensor untouched, and κ \kappa κ itself survives only as the normalization of k k k , that is, as a reparameterization of the phase. Neither operation changes a ⊥ ( η ) a_\perp(\eta) a ⊥ ( η ) . Hence the only inputs left are the curve C \mathcal{C} C and the number a e a_e a e . Values at γ = 1 \gamma=1 γ = 1 and γ = 10 4 \gamma=10^4 γ = 1 0 4 agree to six digits. The storage-ring enhancement a e γ a_e\gamma a e γ [38 , 39 ] compares precession rates about a static field at one instant of laboratory time. Here the quantity it would multiply is exactly zero.
At the physical anomaly the angles are small. A circularly polarized Gaussian pulse gives Θ n e t = 2.25 × 10 − 6 a 0 2 N \Theta_{\mathrm{net}}=2.25\times10^{-6}a_0^2N Θ net = 2.25 × 1 0 − 6 a 0 2 N rad, against 2.18 × 10 − 6 a 0 2 N 2.18\times10^{-6}a_0^2N 2.18 × 1 0 − 6 a 0 2 N rad for the cos 2 \cos^2 cos 2 envelope. Section VII turns those numbers into a working point.
Finite focusing
Equation (10) rests on A = A ( η ) \mathcal{A}=\mathcal{A}(\eta) A = A ( η ) . Every real beam has a waist, and in a focused pulse the spin moves already at g = 2 g=2 g = 2 . The failure of the Lawson–Woodward theorem in a focused beam [40 , 41 ] leaves a residual momentum transfer. The hierarchy of spin contributions in the diffraction parameter ϵ = 1 / ( k w 0 ) \epsilon=1/(kw_0) ϵ = 1/ ( k w 0 ) was set out qualitatively in Ref. [34 ] , and a helicity change of up to 10 % 10\% 10% during ponderomotive scattering was reported in Ref. [42 ] . Neither work gives the exponent.
We modeled a focused Gaussian pulse with the Lax–Louisell–McKnight expansion carried to third order in ϵ \epsilon ϵ , closing the Coulomb gauge at the retained order [43 , 44 ] , and integrated the Lorentz and BMT equations in laboratory time. The Maxwell residuals of that model and the systematic spread over its one free constant are tabulated in Sec. S5 of the Supplemental Material. The observable is the spin rotation relative to the momentum: the exit reference triad is carried from the entrance triad by the pure boost u 0 → u f u_0\to u_f u 0 → u f , so the kinematic transport that accompanies ponderomotive deflection is subtracted. Five triads with anomalies a e ∈ { 0 , ± h , ± 2 h } a_e\in\{0,\pm h,\pm2h\} a e ∈ { 0 , ± h , ± 2 h } , h = 0.05 h=0.05 h = 0.05 , then share one orbit, and a five-point finite difference gives the rotation vector as r = r 0 + a e r 1 + a e 2 r 2 \mathbf{r}=\mathbf{r}_0+a_e\mathbf{r}_1+a_e^2\mathbf{r}_2 r = r 0 + a e r 1 + a e 2 r 2 .
Every violation enters at order ϵ 2 \epsilon^2 ϵ 2 for ϵ ≲ 0.15 \epsilon\lesssim0.15 ϵ ≲ 0.15 , as Figure 2 shows. The second-order coefficient keeps its plane-wave value with a relative error 1.05 ϵ 2 1.05\epsilon^2 1.05 ϵ 2 at γ = 10 \gamma=10 γ = 10 , rising to 3.4 ϵ 2 3.4\epsilon^2 3.4 ϵ 2 at γ = 1 \gamma=1 γ = 1 , and the local log-log slope of that error approaches 2.00 2.00 2.00 as ϵ → 0 \epsilon\to0 ϵ → 0 . Two terms that are identically zero in a plane wave grow alongside it, ∣ r 0 ∣ |\mathbf{r}_0| ∣ r 0 ∣ and ∣ r 1 ∣ |\mathbf{r}_1| ∣ r 1 ∣ , so both the g g g -independent rotation and the first order in the anomaly are restored; their exponents fitted over the whole plotted range are 2.87 and 2.89, with local slopes that fall to 2.02 at the small-ϵ \epsilon ϵ end, so that the quadratic law holds only as an asymptote and not over the whole scan. A direct run at the physical anomaly gives
Figure 2. Departure from the holonomy for a focused Gaussian pulse against the diffraction parameter ε = 1 / ( k w 0 ) \varepsilon= 1/(k w_0) ε = 1/ ( k w 0 ) , for a 0 = 1 a_0 = 1 a 0 = 1 , N = 1 N = 1 N = 1 , circular polarization, head-on collision at γ = 10 \gamma= 10 γ = 10 . Plotted are the relative deviation of the second-order coefficient r 2 r_2 r 2 from − 1 2 A n ^ -\frac{1}{2} A \hat{\mathbf{n}} − 2 1 A n ^ , and the magnitudes of the zeroth- and first-order coefficients r 0 r_0 r 0 and r 1 r_1 r 1 , which vanish identically in a plane wave. All three follow ε 2 \varepsilon^2 ε 2 for ε ≲ 0.15 \varepsilon\lesssim0.15 ε ≲ 0.15 . The dashed guide has slope 2, and the local log-log slopes measured over ε = 0.025 \varepsilon= 0.025 ε = 0.025 –0.15 0.15 0.15 run from 1.92 to 2.00 for the deviation of r 2 r_2 r 2 and from 2.71 to 2.02 for r 0 r_0 r 0 . The upper axis gives the waist in wavelengths, w 0 / λ = 1 / ( 2 π ε ) w_0/\lambda= 1/(2\pi\varepsilon) w 0 / λ = 1/ ( 2 π ε ) . Open circles repeat the deviation of r 2 r_2 r 2 for linear polarization, where it stays at or below the roundoff floor 7.5 × 10 − 11 7.5 \times10^{-11} 7.5 × 1 0 − 11 for every ε \varepsilon ε .
Θ a n o m ( ϵ ) 1 2 a e 2 A − 1 ≃ 10.7 a 0 ϵ 2 ( γ = 10 ) , (13) \frac{\Theta_{\mathrm{anom}}(\epsilon)}{\frac{1}{2}a_e^2\mathcal{A}}-1 \simeq\frac{10.7}{a_0}\epsilon^2 \qquad(\gamma=10), \tag*{(13)} 2 1 a e 2 A Θ anom ( ϵ ) − 1 ≃ a 0 10.7 ϵ 2 ( γ = 10 ) , (13)
the coefficient being measured as 10.7, 10.8 and 13.5 at a 0 = 1 a_0=1 a 0 = 1 and ϵ = 0.025 \epsilon=0.025 ϵ = 0.025 , 0.050 and 0.150. Therefore Eq. (10) holds to 1% for w 0 > 5.2 λ w_0>5.2\lambda w 0 > 5.2 λ at a 0 ≃ 1 a_0\simeq1 a 0 ≃ 1 and γ = 10 \gamma=10 γ = 10 . That number carries the systematic uncertainty of the field model: a factor of three on the coefficient, holding steady as ε \varepsilon ε falls, so at the far end of the spread the requirement becomes w 0 > 9 λ w_0 > 9\lambda w 0 > 9 λ . The exponent survives that spread. At γ = 1 \gamma= 1 γ = 1 the same accuracy needs w 0 > 99 λ w_0 > 99\lambda w 0 > 99 λ , and 10% accuracy still needs w 0 > 31 λ w_0 > 31\lambda w 0 > 31 λ ; the waist requirement relaxes as the intensity rises, as 1 / a 0 1/a_0 1/ a 0 up to a 0 ≃ 2 a_0 \simeq2 a 0 ≃ 2 , and we have not tested it above a 0 = 24 a_0 = 24 a 0 = 24 .
However, the dominant focusing term carries no information about g g g . For γ ≥ 100 \gamma\ge100 γ ≥ 100 we measure ∣ r 0 ∣ = ∣ Δ u ⊥ ∣ |r_0| = |\Delta u_\perp| ∣ r 0 ∣ = ∣Δ u ⊥ ∣ to four digits with the axis orthogonal to the momentum, which identifies r 0 r_0 r 0 as the Thomas–Wigner rotation [15 , 23 ] that accompanies the ponderomotive kick [45 ] and is fixed by the orbit alone. With it go two properties of the plane-wave result: the net rotation is nonzero already at g = 2 g = 2 g = 2 , and helicity is no longer conserved, because the rotation axis has turned perpendicular to the momentum. At ε = 0.15 \varepsilon= 0.15 ε = 0.15 and a 0 = 24 a_0 = 24 a 0 = 24 the anomalous part amounts to a 5 × 10 − 5 5 \times10^{-5} 5 × 1 0 − 5 fraction of the total signal. At a 0 = 1 a_0 = 1 a 0 = 1 the same ratio is 1.25 × 10 − 3 1.25 \times10^{-3} 1.25 × 1 0 − 3 , twenty-five times larger. One property does survive: for linear polarization r 2 r_2 r 2 stays at the roundoff floor for every ε \varepsilon ε up to 0.30, although nothing in Eq. (10) guarantees it there. We have no argument for why. All of this is one electron on one orbit. We did not average over a beam, and the deposited scans give ∣ r 0 ∣ |r_0| ∣ r 0 ∣ varying by a factor of 20 with impact parameter inside a single waist.
Signal, background and the classical window
The estimates below use published scalings, and no radiative process was integrated here, classically or quantum mechanically. The quantum nonlinearity parameter is χ ≃ 6.07 × 10 − 6 γ a 0 \chi\simeq6.07 \times10^{-6} \gamma a_0 χ ≃ 6.07 × 1 0 − 6 γ a 0 at ℏ ω = 1.55 e V \hbar\omega= 1.55\,\mathrm{eV} ℏ ω = 1.55 eV . The three limits on a classical treatment are the radiated fraction R C ≃ 3 2 α a 0 χ N R_C \simeq\frac{3}{2}\alpha a_0 \chi N R C ≃ 2 3 α a 0 χ N , the photon number N γ ∼ α a 0 N N_\gamma\sim\alpha a_0 N N γ ∼ α a 0 N , and the radiative spin-flip probability P s f ∼ α a 0 χ 2 N P_{\mathrm{sf}} \sim\alpha a_0 \chi^2 N P sf ∼ α a 0 χ 2 N [2 , 3 , 46 ] . All three tighten with γ \gamma γ while Θ n e t \Theta_{\mathrm{net}} Θ net does not move with it, so the optimum runs against intuition: keep the electron slow and pay in intensity and duration.
Take the best classical point of the plane-wave grid, γ = 1 \gamma= 1 γ = 1 , a 0 = 75 a_0 = 75 a 0 = 75 and N = 32 N = 32 N = 32 , an intensity of 1.2 × 10 22 W c m − 2 1.2 \times10^{22}\,\mathrm{W\,cm}^{-2} 1.2 × 1 0 22 W cm − 2 at 800 nm, where Θ n e t ≃ 0.4 \Theta_{\mathrm{net}} \simeq0.4 Θ net ≃ 0.4 rad and the transverse depolarization is sin 2 ( Θ n e t / 2 ) ≃ 4 × 10 − 2 \sin^2(\Theta_{\mathrm{net}}/2) \simeq4 \times10^{-2} sin 2 ( Θ net /2 ) ≃ 4 × 1 0 − 2 , with χ ≃ 2 × 10 − 4 \chi\simeq2 \times10^{-4} χ ≃ 2 × 1 0 − 4 , R C ≃ 3 × 10 − 3 R_C \simeq3 \times10^{-3} R C ≃ 3 × 1 0 − 3 and P s f ∼ 10 − 6 P_{\mathrm{sf}} \sim10^{-6} P sf ∼ 1 0 − 6 . That probability sits four to five orders below the depolarization it would have to compete with, and it acts differently: a radiative flip removes one spin from the polarized fraction and leaves the angle of the remainder untouched. Radiation reaction is the more dangerous of the two, because it shifts the angle itself. Its weight can be estimated as a relative distortion 2 R C 2R_C 2 R C of the curve traced in the rest frame against a holonomy of relative size a e a 0 a_e a_0 a e a 0 , giving Θ R R / Θ n e t ∼ 2 R C / ( a e a 0 ) \Theta_{\mathrm{RR}}/\Theta_{\mathrm{net}} \sim2R_C/(a_e a_0) Θ RR / Θ net ∼ 2 R C / ( a e a 0 ) . At the point just quoted that estimator is 6 × 10 − 2 6 \times10^{-2} 6 × 1 0 − 2 , and at γ = 10 \gamma= 10 γ = 10 , a 0 = 75 a_0 = 75 a 0 = 75 , N = 8 N = 8 N = 8 it is 0.30. A correction of six to thirty percent sits badly with a law quoted to a percent, and only an integration of the Landau–Lifshitz equation would settle the matter. We did not perform one. At γ = 10 3 \gamma= 10^3 γ = 1 0 3 the same estimator reaches 4, and P s f P_{\mathrm{sf}} P sf exceeds the geometric signal outright: at a 0 = 10 a_0 = 10 a 0 = 10 and N = 8 N = 8 N = 8 the two are 2 × 10 − 3 2 \times10^{-3} 2 × 1 0 − 3 and 1.8 × 10 − 3 1.8 \times10^{-3} 1.8 × 1 0 − 3 rad. The correction has overtaken the quantity it corrects.
The focusing background is the harder problem, and the numbers of Sec. VI settle it against us. The criterion is ∣ r 0 ∣ < a e 2 ∣ r 2 ∣ |r_0| < a_e^2|r_2| ∣ r 0 ∣ < a e 2 ∣ r 2 ∣ . Relative accuracy of r 2 r_2 r 2 is a different question. At the softest point of the ε \varepsilon ε scan, ε = 0.025 \varepsilon= 0.025 ε = 0.025 or w 0 = 6.4 λ w_0 = 6.4\lambda w 0 = 6.4 λ , at a 0 = 1 a_0 = 1 a 0 = 1 , N = 1 N = 1 N = 1 and γ = 10 \gamma= 10 γ = 10 , the measured ∣ r 0 ∣ = 1.44 × 10 − 5 |r_0| = 1.44 \times10^{-5} ∣ r 0 ∣ = 1.44 × 1 0 − 5 stands against an anomalous signal a e 2 ∣ r 2 ∣ = 2.25 × 10 − 6 a_e^2|r_2| = 2.25 \times10^{-6} a e 2 ∣ r 2 ∣ = 2.25 × 1 0 − 6 , so the g g g -independent background exceeds the whole signal by a factor of 6.4. Hence pushing the background below the signal requires ε 2 < a e 2 ∣ r 2 ∣ / c 0 \varepsilon^2 < a_e^2|r_2|/c_0 ε 2 < a e 2 ∣ r 2 ∣/ c 0 with the measured c 0 c_0 c 0 , that is w 0 > 16 λ w_0 > 16\lambda w 0 > 16 λ at γ = 10 \gamma= 10 γ = 10 and w 0 > 13.6 λ w_0 > 13.6\lambda w 0 > 13.6 λ at γ = 10 3 \gamma= 10^3 γ = 1 0 3 .
However, two axes save that comparison at high energy and destroy it at low. For γ ≥ 100 \gamma\ge100 γ ≥ 100 the background axis is orthogonal to the momentum, ∣ r 0 ∥ ∣ / ∣ r 0 ∣ ≤ 6 × 10 − 5 |r_{0\parallel}|/|r_0| \le6 \times10^{-5} ∣ r 0 ∥ ∣/∣ r 0 ∣ ≤ 6 × 1 0 − 5 at γ = 10 3 \gamma= 10^3 γ = 1 0 3 , while the holonomy lies along the momentum, so a measurement of the longitudinal component separates the two. The separation is two orders of magnitude weaker at γ = 10 2 \gamma= 10^2 γ = 1 0 2 , where the ratio is 5.7 × 10 − 3 5.7 \times10^{-3} 5.7 × 1 0 − 3 , and it fails below γ ≃ 3 \gamma\simeq3 γ ≃ 3 : at γ = 1 \gamma= 1 γ = 1 , ε = 0.15 \varepsilon= 0.15 ε = 0.15 and a 0 = 1 a_0 = 1 a 0 = 1 the background is mostly parallel to the momentum, 0.148 against 0.014 transverse, which is exactly the axis the signal occupies. Requiring 0.148 ( ε / 0.15 ) 2 < 2.25 × 10 − 6 0.148(\varepsilon/0.15)^2 < 2.25 \times10^{-6} 0.148 ( ε /0.15 ) 2 < 2.25 × 1 0 − 6 then gives w 0 > 270 λ w_0 > 270\lambda w 0 > 270 λ .
Therefore the two ends close on each other: at γ ≃ 1 \gamma\simeq1 γ ≃ 1 radiation is harmless, but the focusing background shares the signal axis and exceeds it by four to five orders of magnitude at any attainable waist, while at γ ≳ 10 2 \gamma\gtrsim10^2 γ ≳ 1 0 2 the axes separate and the radiative flip and the radiated fraction have overtaken the effect instead. No window is left between them. The flagship point compounds the problem from the other side. Setting a 0 = 75 a_0 = 75 a 0 = 75 at w 0 = 31 λ w_0 = 31\lambda w 0 = 31 λ , which is the 10% row of the waist table at γ = 1 \gamma= 1 γ = 1 , means a pulse power I π w 0 2 / 2 ≃ 1.2 × 10 17 W I\pi w_0^2/2 \simeq1.2 \times10^{17}\ \mathrm{W} I π w 0 2 /2 ≃ 1.2 × 1 0 17 W and about 10 kJ in 32 cycles, one to two orders above the largest systems in operation or under construction. Focused calculations were run only at N = 1 N = 1 N = 1 and a 0 ≤ 24 a_0 \leq24 a 0 ≤ 24 , so that point also lies outside the validated region. We do not claim an experimental window, and we state the budget in full because the alternative is to show the favorable half of each comparison.
Three further qualifications belong here. The obvious escape, a particle with a larger anomaly, is closed by the mass: the relevant parameter is a e a 0 ∝ a e m e / m a_e a_0 \propto a_e m_e/m a e a 0 ∝ a e m e / m , which is 1.16 × 10 − 3 a 0 1.16 \times10^{-3}a_0 1.16 × 1 0 − 3 a 0 for the electron against 9.8 × 10 − 4 a 0 9.8 \times10^{-4}a_0 9.8 × 1 0 − 4 a 0 for the proton and 5.6 × 10 − 6 a 0 5.6 \times10^{-6}a_0 5.6 × 1 0 − 6 a 0 for the muon, so the proton anomaly of 1.793 buys nothing. Extracting a e 2 a_e^2 a e 2 from a measured angle is no better than the absolute calibration of a 0 2 N a_0^2N a 0 2 N inside the interaction volume, which at 10 22 W c m − 2 10^{22}\ \mathrm{W\ cm}^{-2} 1 0 22 W cm − 2 is a few tens of percent, and the focal average was never taken. And the configuration needs a transversely polarized beam of γ ≃ 1 \gamma\simeq1 γ ≃ 1 –10 synchronized to the focus, for which laser-plasma sources of polarized electrons are the natural starting point [47 , 48 , 49 ] ; whether the required control of the impact parameter is realistic is an experimental question we have not addressed.
Scope and outlook
Every hypothesis behind Eq. (10) can be broken on purpose. Photon emission and radiation reaction spoil u ( + ∞ ) = u 0 u(+\infty) = u_0 u ( + ∞ ) = u 0 , and radiation reaction brings in a dimensionful time, so γ \gamma γ and ω \omega ω reenter. The closest published mechanism turns the spin toward the propagation axis of a circularly polarized field at first order in a e a_e a e , and it is emission that keeps the transverse spin phased there [50 ] . Spin-dependent radiative deflection acts in the same regime [51 ] . A standing wave adds a second null direction, so that the two little algebras together generate the whole Lorentz algebra and the ordering survives at g = 2 g = 2 g = 2 [52 , 53 ] , while a plasma gives k k k a mass, k 2 = ω p 2 k^2 = \omega_p^2 k 2 = ω p 2 , and a first-order rotation returns with weight ( n e / n c ) / ( a e a 0 ) (n_e/n_c)/(a_ea_0) ( n e / n c ) / ( a e a 0 ) , already of order 0.5 at n e / n c = 6 × 10 − 3 n_e/n_c = 6 \times10^{-3} n e / n c = 6 × 1 0 − 3 and a 0 = 10 a_0 = 10 a 0 = 10 . That rotation depends on the electron energy and on the carrier frequency, where the holonomy does not, and it disappears as v p h → c v_{\mathrm{ph}} \to c v ph → c [36 , 37 ] . In addition, self-interaction in a strong plane wave carries a spin dynamics of its own at the same order in α \alpha α [54 ] . We integrated none of these cases. The estimates come from the standard scalings collected in Refs. [5 , 46 ] .
However, a dressed anomaly leaves the structure intact but not the derivation: the cancellation of Eq. (8) takes a e a_e a e outside the integral, so a phase-dependent a e ( χ ( η ) ) a_e(\chi(\eta)) a e ( χ ( η )) would leave ∫ a e ( η ) n ^ ∧ d a ⊥ ≠ 0 \int a_e(\eta)\,\hat{n}\wedge d a_\perp\ne0 ∫ a e ( η ) n ^ ∧ d a ⊥ = 0 and restore a first-order term [34 , 55 ] . At χ ≃ 2 × 10 − 4 \chi\simeq2 \times10^{-4} χ ≃ 2 × 1 0 − 4 the dressing is far too weak to matter, but the statement is exact only for constant a e a_e a e , and outside the locally constant field approximation a local anomaly is not defined at all [20 ] .
Two uses survive the negative budget of Sec. VII. The identical zero at g = 2 g = 2 g = 2 , and the identical zero at any g g g for linear polarization, give structure-preserving spin integrators [56 , 57 ] a reference that needs no converged solution to compare against, and the tolerance scan of Sec. S4.3 shows what such a test looks like in practice. The classical result is also a well-posed question at the quantum level, where the exact Dirac–Pauli solutions in a plane wave [9 , 11 , 19 ] make it tractable. We expect an O ( a e 2 ) O(a_e^2) O ( a e 2 ) geometric term in the spin-flip amplitude of Ref. [34 ] at the order beyond the one written there, with a coefficient predicted by Eq. (10) and no free parameter.
Acknowledgments
Earlier stages of this work were presented at the 26th International Symposium on Spin Physics (SPIN 2025), Qingdao, China, 21–26 September 2025, and at the 1st Conference on Strong-Interaction Spin Physics, Qingdao, China, 26–31 July 2026.
This work was partially supported by the State Assignment of the Ministry of Education and Science of the Russian Federation (Project No. FZEN 2023-0006), the Nantong Science and Technology Plan Project (Grant Nos. JC2020137 and JC2020138), the Key Research and Development Program of Jiangsu Province of China (Grant No. BE2021013-1), the National Natural Science Foundation of Jiangsu Province of China (Grant No. BK20201438), and in part by the Natural Science Research Project of Jiangsu Provincial Institutions of Higher Education (Grant Nos. 1120KJA510002 and 20KJB510010).
N. S. A. and A. P. N. developed the analytic reduction and wrote the manuscript. A. P. N. performed the plane-wave computations. S. N. A. performed the focused-beam computations and the Maxwell residual tests. Q.-H. Q. supervised the project. All authors discussed the results and approved the final manuscript.
The authors declare no competing financial interest.
Data availability
The code that generates the data and the figures of this paper, together with the computed results themselves, is openly available in the Zenodo repository at Ref. [58 ] .
Supplemental Material
In the journal version this material is a separate file. It is reproduced here unchanged so that the preprint stands on its own. Sections, equations, figures and tables below carry the prefix “S”; “the main text” means Secs. I–VIII above, whose equation numbers carry no prefix.
Sections S1 and S2 carry the derivation of the holonomy law in the detail the main text leaves out, from the assumptions behind the Thomas–Bargmann–Michel–Telegdi (BMT) equation to the non-Abelian Stokes evaluation of the ordered exponential. Section S3 collects the invariance properties that make the result geometric, including the reason the storage-ring factor a e γ a_e\gamma a e γ does not apply. Section S4 documents the numerical protocol, the invariant control, the convergence of the null result in the integrator tolerance and the validation runs, Sec. S5 the focused-beam model and its limitations, Sec. S6 the window in which the classical treatment is meaningful, and Sec. S7 the mechanisms that break the theorem. Equations, figures, and tables carry the prefix “S”. Equation numbers without it refer to the main text. The anomaly is written a e = ( g − 2 ) / 2 a_e=(g-2)/2 a e = ( g − 2 ) /2 here and in the main text alike.
Every number quoted below is labeled either measured , meaning it was produced by one of the runs described in Secs. S4 and S5, or estimated , meaning it follows from a published scaling with no integration of our own. Nothing in between is reported, and where a quantity was neither measured nor estimated we say so: Sec. S3.4 is the one place where that happens.
The Thomas–BMT equation in light-front variables
Field and orbit
A null wave vector k μ k^\mu k μ with k 2 = 0 k^2=0 k 2 = 0 fixes the light-front phase η = k ⋅ x \eta=k\cdot x η = k ⋅ x in the metric ( + , − , − , − ) (+,-,-,-) ( + , − , − , − ) , and the dimensionless potential
a μ ( η ) = e A μ ( η ) m , k ⋅ a = 0 , (14) a^\mu(\eta)=\frac{eA^\mu(\eta)}{m},\qquad k\cdot a=0, \tag*{(14)} a μ ( η ) = m e A μ ( η ) , k ⋅ a = 0 , (14)
depends on η \eta η alone. The anomaly is written a e = ( g − 2 ) / 2 a_e=(g-2)/2 a e = ( g − 2 ) /2 , which leaves the symbol a a a for the potential and its components alone. The amplitude parameter is a 0 a_0 a 0 : in the convention fixed below the peak of ∣ a ⊥ ∣ \lvert a_\perp\rvert ∣ a ⊥ ∣ is a 0 / 1 + δ 2 a_0/\sqrt{1+\delta^2} a 0 / 1 + δ 2 at ellipticity δ \delta δ , which is a 0 a_0 a 0 for linear polarization and a 0 / 2 a_0/\sqrt{2} a 0 / 2 for circular polarization, the normalization holding the mean square rather than the peak fixed. The vector a ⊥ a_\perp a ⊥ carries components ( a 1 , a 2 ) = ( a x , a y ) (a_1,a_2)=(a_x,a_y) ( a 1 , a 2 ) = ( a x , a y ) along a right-handed polarization basis e ^ 1 × e ^ 2 = n ^ \hat e_1\times\hat e_2=\hat n e ^ 1 × e ^ 2 = n ^ .
Every sign below is referred to a single fixed convention, and all of them have been rechecked against it. The wedge product of two four-vectors is the mixed tensor
( p ∧ q ) μ ν = p μ q ν − q μ p ν , (p\wedge q)^\mu{}_\nu=p^\mu q_\nu-q^\mu p_\nu, ( p ∧ q ) μ ν = p μ q ν − q μ p ν ,
which acts on a four-vector S S S as p ( q ⋅ S ) − q ( p ⋅ S ) p(q\cdot S)-q(p\cdot S) p ( q ⋅ S ) − q ( p ⋅ S ) . The symbol ∧ \wedge ∧ never denotes a three-dimensional wedge below. The propagation direction is written n ^ \hat n n ^ both as a unit spatial vector and as the four-vector n ^ μ = 1 2 ( k − κ u ) μ \hat n^\mu=\frac{1}{2}(k-\kappa u)^\mu n ^ μ = 2 1 ( k − κ u ) μ that represents it in the rest frame of u 0 u_0 u 0 , and the same holds for a ⊥ a_\perp a ⊥ , whose four-vector form is a μ a^\mu a μ . A bold two-vector and its four-vector are interchangeable inside a wedge, where only the transverse components contribute. One consequence is used repeatedly: for spatial p p p and q q q and a rest-frame spin ζ \zeta ζ the lowered index flips a sign, ( p ∧ q ) ζ = ( p × q ) × ζ (p\wedge q)\zeta=(p\times q)\times\zeta ( p ∧ q ) ζ = ( p × q ) × ζ , so that e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 generates a right-handed rotation about n ^ \hat n n ^ and n ^ ∧ a ⊥ ′ \hat n\wedge a'_\perp n ^ ∧ a ⊥ ′ a right-handed rotation about n ^ × a ⊥ ′ \hat n\times a'_\perp n ^ × a ⊥ ′ .
Finiteness of the pulse enters once, through a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) = 0 a_\perp(+\infty)=a_\perp(-\infty)=0 a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) = 0 , which is Eq. (1) of the main text. The phase average of a ⊥ a_\perp a ⊥ is not constrained, and Sec. S3.5 returns to that point.
The scans of Secs. S4 and S5 use
a x = a 0 f ( η ) 1 + δ 2 cos ( η + φ 0 ) , a y = a 0 δ f ( η ) 1 + δ 2 sin ( η + φ 0 ) , (15) a_x=\frac{a_0f(\eta)}{\sqrt{1+\delta^2}}\cos(\eta+\varphi_0),\qquad
a_y=\frac{a_0\delta f(\eta)}{\sqrt{1+\delta^2}}\sin(\eta+\varphi_0), \tag*{(15)} a x = 1 + δ 2 a 0 f ( η ) cos ( η + φ 0 ) , a y = 1 + δ 2 a 0 δ f ( η ) sin ( η + φ 0 ) , (15)
with ellipticity δ \delta δ (δ = 0 \delta=0 δ = 0 linear, δ = 1 \delta=1 δ = 1 circular) and carrier-envelope phase (CEP) φ 0 \varphi_0 φ 0 . Two envelopes are used: a Gaussian f = exp ( − η 2 / 2 σ 2 ) f=\exp(-\eta^2/2\sigma^2) f = exp ( − η 2 /2 σ 2 ) with σ = π N / ln 2 = 3.7734 N \sigma=\pi N/\sqrt{\ln2}=3.7734N σ = π N / ln 2 = 3.7734 N , truncated at ∣ η ∣ ≤ 8 σ \lvert\eta\rvert\le8\sigma ∣ η ∣ ≤ 8 σ , and f = cos 2 ( π η / 2 L ) f=\cos^2(\pi\eta/2L) f = cos 2 ( π η /2 L ) with L = 8.6294 N L=8.6294N L = 8.6294 N on compact support. In both, N N N is the number of optical cycles in the full width at half maximum of the intensity envelope f 2 f^2 f 2 . The two are compared at equal N N N . The signed area then evaluates to
A = ∫ ( a x a y ′ − a y a x ′ ) d η = δ 1 + δ 2 a 0 2 ∫ f 2 d η , (16) \mathcal{A}=\int(a_xa_y'-a_ya_x')\,d\eta
=\frac{\delta}{1+\delta^2}a_0^2\int f^2\,d\eta, \tag*{(16)} A = ∫ ( a x a y ′ − a y a x ′ ) d η = 1 + δ 2 δ a 0 2 ∫ f 2 d η , (16)
with ∫ f 2 d η = σ π \int f^2d\eta=\sigma\sqrt{\pi} ∫ f 2 d η = σ π for the Gaussian and 0.75 L 0.75L 0.75 L for cos 2 \cos^2 cos 2 .
Because k μ F μ ν = 0 k_\mu F^{\mu\nu}=0 k μ F μν = 0 for a null plane wave, the light-front projection κ = k ⋅ u \kappa=k\cdot u κ = k ⋅ u is conserved, d / d τ = κ d / d η d/d\tau=\kappa\,d/d\eta d / d τ = κ d / d η , and the orbit is the Volkov solution [21 ]
u μ ( η ) = u 0 μ − a μ + k μ κ ( a ⋅ u 0 − 1 2 a ⋅ a ) , u ( ± ∞ ) = u 0 . (17) u^\mu(\eta)=u_0^\mu-a^\mu+\frac{k^\mu}{\kappa}\left(a\cdot u_0-\frac{1}{2}a\cdot a\right),\qquad
u(\pm\infty)=u_0. \tag*{(17)} u μ ( η ) = u 0 μ − a μ + κ k μ ( a ⋅ u 0 − 2 1 a ⋅ a ) , u ( ± ∞ ) = u 0 . (17)
In components, u ⊥ ( η ) = u 0 ⊥ − a ⊥ ( η ) u_\perp(\eta)=u_{0\perp}-a_\perp(\eta) u ⊥ ( η ) = u 0 ⊥ − a ⊥ ( η ) and u 0 , 3 = 1 2 [ κ ± ( 1 + u ⊥ 2 ) / κ ] u^{0,3}=\frac{1}{2}\left[\kappa\pm(1+u_\perp^2)/\kappa\right] u 0 , 3 = 2 1 [ κ ± ( 1 + u ⊥ 2 ) / κ ] , where κ = u 0 − u 3 \kappa=u^0-u^3 κ = u 0 − u 3 is the same conserved quantity, written for k k k normalized to unit frequency. For a head-on collision κ = γ ( 1 + β ) ≃ 2 γ \kappa=\gamma(1+\beta)\simeq2\gamma κ = γ ( 1 + β ) ≃ 2 γ . No integration is needed for the orbit, which removes one source of numerical error from everything that follows.
Assumptions
The spin is a classical four-vector S μ S^\mu S μ with S ⋅ u = 0 S\cdot u=0 S ⋅ u = 0 and S 2 = − 1 S^2=-1 S 2 = − 1 , transported by the covariant BMT equation [7 , 15 , 16 ] , Eq. (2) of the main text,
d S μ d τ = e m [ g 2 F μ ν S ν + ( g 2 − 1 ) u μ ( S λ F λ ν u ν ) ] . (18) \frac{dS^\mu}{d\tau}=\frac{e}{m}\left[\frac{g}{2}F^{\mu\nu}S_\nu+\left(\frac{g}{2}-1\right)u^\mu\left(S_\lambda F^{\lambda\nu}u_\nu\right)\right]. \tag*{(18)} d τ d S μ = m e [ 2 g F μν S ν + ( 2 g − 1 ) u μ ( S λ F λ ν u ν ) ] . (18)
Its laboratory form is d S / d t = Ω × S d\mathbf{S}/dt=\mathbf{\Omega}\times\mathbf{S} d S / d t = Ω × S with
Ω = − e m [ ( a e + 1 γ ) B − a e γ γ + 1 ( β ⋅ B ) β − ( a e + 1 γ + 1 ) β × E ] , (19) \mathbf{\Omega}=-\frac{e}{m}\left[\left(a_e+\frac{1}{\gamma}\right)\mathbf{B}-\frac{a_e\gamma}{\gamma+1}(\boldsymbol{\beta}\cdot\mathbf{B})\boldsymbol{\beta}-\left(a_e+\frac{1}{\gamma+1}\right)\boldsymbol{\beta}\times\mathbf{E}\right], \tag*{(19)} Ω = − m e [ ( a e + γ 1 ) B − γ + 1 a e γ ( β ⋅ B ) β − ( a e + γ + 1 1 ) β × E ] , (19)
and the coefficients of Eq. (19) were used to check those of Eq. (18) term by term. At γ = 1 \gamma=1 γ = 1 the first coefficient is a e + 1 = g / 2 a_e+1=g/2 a e + 1 = g /2 , as it must be for an electron at rest. The Stern–Gerlach force is dropped: relative to the Lorentz force it carries a factor ℏ k / m c \hbar k/mc ℏ k / m c , which is ℏ ω / m c 2 ≃ 3 × 10 − 6 \hbar\omega/mc^2\simeq3\times10^{-6} ℏ ω / m c 2 ≃ 3 × 1 0 − 6 for Ti:sapphire light in the rest frame of a slow electron, so that at the accuracy of interest here the orbit is insensitive to the spin. Radiation reaction goes with it, bounded from published scalings in Sec. S6 and never integrated, and so does photon emission, which is the same statement at the quantum level. An electric dipole moment is set to zero. The Stern–Gerlach term is the mildest of the four, by the factor just quoted. The other three are genuine restrictions.
Within the plane-wave sector the classical transport law is not itself an approximation, which limits the damage. Chakrabarti solved the Dirac–Pauli equation exactly for fields of the form F = F ( k ⋅ x ) F=F(k\cdot x) F = F ( k ⋅ x ) with k 2 = 0 k^2=0 k 2 = 0 and showed that the polarization four-vector obeys the BMT equation exactly, with an arbitrary anomalous magnetic moment and an arbitrary electric dipole moment [9 ] . Barducci and Giachetti later constructed the eigenfunctions and the Green function of the same problem [19 ] . The classical spin transport used below is therefore the exact one-particle content of the Dirac–Pauli theory in this background, and our treatment omits radiation while keeping the spin kinematics exact.
A single antisymmetric generator
Since u ⋅ S = 0 u\cdot S=0 u ⋅ S = 0 , the right-hand side of Eq. (18) can be brought to antisymmetric form. The completion is unique: two antisymmetric matrices that act identically on the three-dimensional space orthogonal to u u u differ by an antisymmetric matrix with a three-dimensional kernel. The rank of an antisymmetric matrix is even. That difference therefore vanishes. Writing f = ( e / m ) F = k ∧ a ′ f=(e/m)F=k\wedge a' f = ( e / m ) F = k ∧ a ′ and w μ = f μ ν u ν w^\mu=f^{\mu\nu}u_\nu w μ = f μν u ν ,
d S μ d η = 1 κ Ω μ ν S ν , Ω = ( 1 + a e ) k ∧ a ′ + a e ( u ∧ w ) , (20) \frac{dS^\mu}{d\eta}=\frac{1}{\kappa}\Omega^\mu{}_\nu S^\nu,\qquad\Omega=(1+a_e)k\wedge a'+a_e(u\wedge w), \tag*{(20)} d η d S μ = κ 1 Ω μ ν S ν , Ω = ( 1 + a e ) k ∧ a ′ + a e ( u ∧ w ) , (20)
which is Eq. (3) of the main text. The rest of Secs. S1 and S2 is written for a head-on geometry normalized to κ = 1 \kappa=1 κ = 1 , and the restriction is lifted here, one paragraph before it is imposed. A null rotation N N N in the little group I S O ( 2 ) k ISO(2)_k I S O ( 2 ) k of k k k leaves k k k fixed and sends e i ↦ e i + c i k e_i\mapsto e_i+c_i k e i ↦ e i + c i k , so that N f N − 1 = k ∧ ( a ′ + c k ) = f NfN^{-1}=k\wedge(a'+ck)=f N f N − 1 = k ∧ ( a ′ + c k ) = f : the background is invariant under I S O ( 2 ) k ISO(2)_k I S O ( 2 ) k , which is stronger than gauge equivalence, and since I S O ( 2 ) k ISO(2)_k I S O ( 2 ) k acts transitively on the mass shell at fixed κ \kappa κ , any u 0 u_0 u 0 can be carried to ( k + k ˉ ) / 2 (k+\bar{k})/2 ( k + k ˉ ) /2 . The net map of the original geometry is then the conjugate N − 1 Λ n e t N N^{-1}\Lambda_{\mathrm{net}}N N − 1 Λ net N , and a rotation angle is a conjugation invariant. This matters because Eq. (21) below, and with it the reduction of Sec. S2.2 and Eq. (7) of the main text, is false as written when u 0 ⊥ ≠ 0 u_{0\perp}\ne0 u 0 ⊥ = 0 : the general form of the reduced generator carries k k k , M = 1 2 ( k / κ − k ˉ κ ) ∧ a ′ M=\frac{1}{2}(k/\kappa-\bar{k}\kappa)\wedge a' M = 2 1 ( k / κ − k ˉ κ ) ∧ a ′ , in which κ \kappa κ is the parameter of η \eta η and nothing else. Section S3.2 gives the measured values.
Expanding u ∧ w u\wedge w u ∧ w in the basis { k , k ˉ , e 1 , e 2 } \{k,\bar{k},e_1,e_2\} { k , k ˉ , e 1 , e 2 } with k ⋅ k ˉ = 2 k\cdot\bar{k}=2 k ⋅ k ˉ = 2 and e i ⋅ e j = − δ i j e_i\cdot e_j=-\delta_{ij} e i ⋅ e j = − δ ij , and with u 0 = ( k + k ˉ ) / 2 u_0=(k+\bar{k})/2 u 0 = ( k + k ˉ ) /2 ,
u ∧ w = U 2 ( k ˉ ∧ k ) − 1 2 k ˉ ∧ a ′ − ( 1 2 + ∣ a ⊥ ∣ 2 2 ) k ∧ a ′ + U k ∧ a + W ( e 1 ∧ e 2 ) , (21) u\wedge w=\frac{\mathcal{U}}{2}(\bar{k}\wedge k)-\frac{1}{2}\bar{k}\wedge a'
-\left(\frac{1}{2}+\frac{|a_\perp|^2}{2}\right)k\wedge a'+\mathcal{U}k\wedge a+\mathcal{W}(e_1\wedge e_2), \tag*{(21)} u ∧ w = 2 U ( k ˉ ∧ k ) − 2 1 k ˉ ∧ a ′ − ( 2 1 + 2 ∣ a ⊥ ∣ 2 ) k ∧ a ′ + U k ∧ a + W ( e 1 ∧ e 2 ) , (21)
with
U = a ⊥ ⋅ a ⊥ ′ , W = a 1 a 2 ′ − a 2 a 1 ′ . (22) \mathcal{U}=a_\perp\cdot a'_\perp,\qquad\mathcal{W}=a_1a'_2-a_2a'_1. \tag*{(22)} U = a ⊥ ⋅ a ⊥ ′ , W = a 1 a 2 ′ − a 2 a 1 ′ . (22)
The symbol χ \chi χ is reserved throughout for the quantum nonlinearity parameter of Sec. S6.
However, two of the five terms in Eq. (21) lie outside the little algebra i s o ( 2 ) − \mathfrak{iso}(2)_- iso ( 2 ) − of k k k : the boost k ˉ ∧ k \bar{k}\wedge k k ˉ ∧ k and the conjugate null rotation k ˉ ∧ a ′ \bar{k}\wedge a' k ˉ ∧ a ′ . Together with k ∧ e i k\wedge e_i k ∧ e i they generate a nonabelian s l ( 2 , R ) \mathfrak{sl}(2,\mathbb{R}) sl ( 2 , R ) already for linear polarization. Hence the η \eta η ordering in Eq. (20) is not degenerate at g ≠ 2 g\ne2 g = 2 , and the one-line argument of Sec. S2.1 does not extend to it.
Derivation of the holonomy law
The g = 2 g=2 g = 2 propagator
At a e = 0 a_e=0 a e = 0 the generator is Ω = k ∧ a ′ = a 1 ′ N 1 + a 2 ′ N 2 \Omega=k\wedge a'=a'_1N_1+a'_2N_2 Ω = k ∧ a ′ = a 1 ′ N 1 + a 2 ′ N 2 with N i = k ∧ e i N_i=k\wedge e_i N i = k ∧ e i . Direct multiplication gives
( N i N j ) μ ν = δ i j k μ k ν , (23) (N_iN_j)^\mu{}_\nu=\delta_{ij}k^\mu k_\nu, \tag*{(23)} ( N i N j ) μ ν = δ ij k μ k ν , (23)
which is symmetric in ( i j ) (ij) ( ij ) , so [ N 1 , N 2 ] = 0 [N_1,N_2]=0 [ N 1 , N 2 ] = 0 and N i 3 = 0 N_i^3=0 N i 3 = 0 . The generator commutes with itself at different phases, the ordering collapses, and
Λ 0 ( η ) = exp [ a 1 ( η ) N 1 + a 2 ( η ) N 2 ] , Λ 0 ( + ∞ ) = 1 , (24) \Lambda_0(\eta)=\exp[a_1(\eta)N_1+a_2(\eta)N_2],\qquad\Lambda_0(+\infty)=\mathbb{1}, \tag*{(24)} Λ 0 ( η ) = exp [ a 1 ( η ) N 1 + a 2 ( η ) N 2 ] , Λ 0 ( + ∞ ) = 1 , (24)
by Eq. (1) of the main text. No property of u 0 u_0 u 0 enters, since u u u does not appear in Ω \Omega Ω at a e = 0 a_e=0 a e = 0 at all. This is the statement that the asymptotic spin index of a Volkov state is conserved [21 ] , and for linear polarization it is the vanishing of the electric pulse area found by Aleksandrov et al. [31 ] .
Interaction picture and the exact reduction
Since ( 1 + a e ) f + a e u ∧ w = f + a e ( f + u ∧ w ) (1+a_e)f+a_e\,u\wedge w=f+a_e\left(f+u\wedge w\right) ( 1 + a e ) f + a e u ∧ w = f + a e ( f + u ∧ w ) , factoring the evolution as U = Λ 0 V U=\Lambda_0V U = Λ 0 V gives
d V d η = a e M ( η ) V , M = Ad Λ 0 − 1 ( f + u ∧ w ) . (25) \frac{dV}{d\eta}=a_eM(\eta)V,\qquad M=\operatorname{Ad}_{\Lambda_0^{-1}}\left(f+u\wedge w\right). \tag*{(25)} d η d V = a e M ( η ) V , M = Ad Λ 0 − 1 ( f + u ∧ w ) . (25)
Write A = k ∧ a ⊥ = a 1 N 1 + a 2 N 2 A=k\wedge a_\perp=a_1N_1+a_2N_2 A = k ∧ a ⊥ = a 1 N 1 + a 2 N 2 , so that Λ 0 = e A \Lambda_0=e^A Λ 0 = e A and M = e − ad A X M=e^{-\operatorname{ad}_A}X M = e − ad A X with X = f + u ∧ w X=f+u\wedge w X = f + u ∧ w . The series terminates after three terms. Nilpotency of ad A \operatorname{ad}_A ad A is what stops it, and on the whole of s o ( 1 , 3 ) \mathfrak{so}(1,3) so ( 1 , 3 ) the operator A A A raises the grading k ˉ ∧ e i → { k ˉ ∧ k , e 1 ∧ e 2 } → k ∧ e i → 0 \bar{k}\wedge e_i\to\{\bar{k}\wedge k,e_1\wedge e_2\}\to k\wedge e_i\to0 k ˉ ∧ e i → { k ˉ ∧ k , e 1 ∧ e 2 } → k ∧ e i → 0 , so ad A 3 = 0 \operatorname{ad}_A^3=0 ad A 3 = 0 and
M = X − [ A , X ] + 1 2 [ A , [ A , X ] ] . M=X-[A,X]+\frac{1}{2}[A,[A,X]]. M = X − [ A , X ] + 2 1 [ A , [ A , X ]] .
Nilpotency of the N i N_i N i alone would give only ad A 5 = 0 \operatorname{ad}_A^5=0 ad A 5 = 0 .
Table S1 carries the three terms out at one numerical point and shows where each cancellation happens. Everything along the boost k ˉ ∧ k \bar{k}\wedge k k ˉ ∧ k and along e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 is removed by the first commutator, with no input beyond the algebra. The two algebraic identities
U a 1 − W a 2 = ∣ a ⊥ ∣ 2 a 1 ′ , U a 2 + W a 1 = ∣ a ⊥ ∣ 2 a 2 ′ , (26) Ua_1-Wa_2=|a_\perp|^2a'_1,\qquad Ua_2+Wa_1=|a_\perp|^2a'_2, \tag*{(26)} U a 1 − W a 2 = ∣ a ⊥ ∣ 2 a 1 ′ , U a 2 + W a 1 = ∣ a ⊥ ∣ 2 a 2 ′ , (26)
which follow from Eq. (22) by inspection, do something else: they collapse the residual null rotations, a combination of k ∧ a ′ k\wedge a' k ∧ a ′ , k ∧ a ⊥ k\wedge a_\perp k ∧ a ⊥ and k ∧ a ~ k\wedge\tilde a k ∧ a ~ with a ~ = a 2 e 1 − a 1 e 2 \tilde a=a_2e_1-a_1e_2 a ~ = a 2 e 1 − a 1 e 2 , into the single term 1 2 k ∧ a ′ \frac{1}{2}k\wedge a' 2 1 k ∧ a ′ . What survives is
M ( η ) = 1 2 ( k − k ˉ ) ∧ a ′ ( η ) = n ^ ∧ a ′ ( η ) , (27) M(\eta)=\frac{1}{2}(k-\bar{k})\wedge a'(\eta)=\hat n\wedge a'(\eta), \tag*{(27)} M ( η ) = 2 1 ( k − k ˉ ) ∧ a ′ ( η ) = n ^ ∧ a ′ ( η ) , (27)
equivalently, for the rest-frame spin ζ \zeta ζ and in the convention of Sec. S1,
d ζ d η = a e ( n ^ × a ⊥ ′ ) × ζ , (28) \frac{d\zeta}{d\eta}=a_e\left(\hat n\times a'_\perp\right)\times\zeta, \tag*{(28)} d η d ζ = a e ( n ^ × a ⊥ ′ ) × ζ , (28)
which is Eq. (7) of the main text. We evaluated the components of Eq. (20) on a grid in η \eta η and compared them with Eq. (27) . The largest discrepancy is 2.7 × 10 − 15 2.7\times10^{-15} 2.7 × 1 0 − 15 (measured).
M M M is a pure rotation of the rest frame about an axis n ^ × a ⊥ ′ \hat n\times a'_\perp n ^ × a ⊥ ′ lying in the polarization plane. For linear polarization that axis is fixed, the rotations at different phases commute, and V ( + ∞ ) = exp [ a e Δ a 1 n ^ ∧ e 1 ] = 1 V(+\infty)=\exp[a_e\Delta a_1\hat n\wedge e_1]=\mathbb{1} V ( + ∞ ) = exp [ a e Δ a 1 n ^ ∧ e 1 ] = 1 exactly at any g g g and any a 0 a_0 a 0 . For u 0 ⊥ ≠ 0 u_{0\perp}\ne0 u 0 ⊥ = 0 the conjugation of Sec. S1 carries the conclusion over unchanged, since N − 1 1 N = 1 N^{-1}\mathbb{1}N=\mathbb{1} N − 1 1 N = 1 . This is the mechanism behind the vanishing quoted in the main text. It is perturbative in nothing.
Cancellation of the first order in the anomaly holds for every polarization, and its cause is the one that also makes the whole construction geometric. M d η = n ^ ∧ d a ⊥ M\,d\eta=\hat n\wedge da_\perp M d η = n ^ ∧ d a ⊥ is a one-form on the polarization plane whose coefficients are constants, so its integral along a path depends on the endpoints alone, and Eq. (1) of the main text says that the two endpoints coincide. Hence a e ∫ M d η = a e n ^ ∧ Δ a ⊥ = 0 a_e\int M\,d\eta=a_e\hat n\wedge\Delta a_\perp=0 a e ∫ M d η = a e n ^ ∧ Δ a ⊥ = 0 , exactly, at any a 0 a_0 a 0 and for any curve. A weaker fact is sometimes offered in place of this one and does not do the work. Λ n e t \Lambda_{\rm net} Λ net lies in the stabilizer S O ( 3 ) u 0 SO(3)_{u_0} S O ( 3 ) u 0 of u 0 u_0 u 0 , which follows from u ( ± ∞ ) = u 0 u(\pm\infty)=u_0 u ( ± ∞ ) = u 0 and the conservation of S ⋅ u S\cdot u S ⋅ u , but membership in S O ( 3 ) u 0 SO(3)_{u_0} S O ( 3 ) u 0 confines the first-order coefficient to a three-dimensional space and constrains it no further. Section S5.4 makes that concrete, since in a focused beam the net map still lies in S O ( 3 ) SO(3) S O ( 3 ) once the boost is subtracted, while its first-order coefficient r 1 r_1 r 1 is nonzero and grows as ε 2 \varepsilon^2 ε 2 .
Non-Abelian Stokes evaluation and the order in the anomaly
The geometric statement can now be given in full. The base is the polarization plane R 2 \mathbb{R}^2 R 2 with coordinates a ⊥ = ( a 1 , a 2 ) a_\perp=(a_1,a_2) a ⊥ = ( a 1 , a 2 ) , the bundle is the trivial principal bundle R 2 × S O ( 3 ) u 0 \mathbb{R}^2\times SO(3)_{u_0} R 2 × S O ( 3 ) u 0 whose structure group is the stabilizer of u 0 u_0 u 0 in S O ( 1 , 3 ) SO(1,3) S O ( 1 , 3 ) , and M M M lies in the corresponding Lie algebra because M u 0 = 0 Mu_0=0 M u 0 = 0 . The connection one-form is
ω = a e ( G 1 d a 1 + G 2 d a 2 ) , G i = n ^ ∧ e i , (29) \omega=a_e(G_1\,da_1+G_2\,da_2),\qquad G_i=\hat n\wedge e_i, \tag*{(29)} ω = a e ( G 1 d a 1 + G 2 d a 2 ) , G i = n ^ ∧ e i , (29)
with coefficients independent of η \eta η and of $a_\perp alike. The path is C : η ↦ a ⊥ ( η ) \mathcal{C}:\eta\mapsto a_\perp(\eta) C : η ↦ a ⊥ ( η ) , and it closes for exactly one reason, Eq. (1) of the main text. That is the only place where finiteness of the pulse is used. The net map is the holonomy of ω \omega ω around that loop,
Λ n e t = Hol ω ( C ) = P exp ∮ C ω , (30) \Lambda_{\rm net}=\operatorname{Hol}_{\omega}(\mathcal{C})=\mathcal{P}\exp\oint_{\mathcal{C}}\omega, \tag*{(30)} Λ net = Hol ω ( C ) = P exp ∮ C ω , (30)
and it depends on C \mathcal{C} C alone, with no reference to the parameterization. No adiabatic parameter appears here and no eigenspace is being followed, so this is ordinary non-Abelian parallel transport, and the resemblance to the geometric phases of [25 , 26 ] is an analogy. The one-form is also exact, but exactness is not what does the work, since in the non-Abelian case an exact form can still have nontrivial holonomy and reparameterization invariance holds for any one-form whatsoever. Constant coefficients are what matter.
The curvature of ω \omega ω is constant over the plane,
[ G 1 , G 2 ] = e 1 ∧ e 2 , F = d ω + ω ∧ ω = a e 2 ( e 1 ∧ e 2 ) d a 1 ∧ d a 2 , (31) [G_1,G_2]=e_1\wedge e_2,\qquad F=d\omega+\omega\wedge\omega=a_e^2(e_1\wedge e_2)\,da_1\wedge da_2, \tag*{(31)} [ G 1 , G 2 ] = e 1 ∧ e 2 , F = d ω + ω ∧ ω = a e 2 ( e 1 ∧ e 2 ) d a 1 ∧ d a 2 , (31)
and e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 generates a right-handed rotation about n ^ \hat n n ^ . The version of the non-Abelian Stokes theorem that applies is the surface-ordered one [27 , 28 ] , in which the path-ordered exponential around C \mathcal{C} C becomes a surface-ordered exponential of the curvature carried back to a base point along a mesh of paths. Constancy of F F F in the coordinates of the base does not make that ordering trivial, because what is ordered is not F F F but its adjoint transport along the mesh, and that is not constant. The theorem therefore delivers the leading term and nothing more: a rotation about n ^ \hat n n ^ through a e 2 a_e^2 a e 2 times the oriented area enclosed by C \mathcal{C} C , with the overall sign, which turns on the orientation convention relating the boundary to the surface ordering, taken below from the Magnus expansion. One further caution belongs here. The area is algebraic and counted with multiplicity, so that for an N N N -cycle envelope, where C C C winds around the origin roughly N N N times, A \mathcal{A} A counts 2 N 2N 2 N times the area of one turn.
Table S1. The three surviving terms of M = e − a e a d A X M=e^{-a_{\mathrm e}a_{\mathrm d}}A X M = e − a e a d A X with X = f + u ∧ w X=f+u\wedge w X = f + u ∧ w , decomposed in the bivector basis of s o ( 1 , 3 ) \mathfrak{so}(1,3) so ( 1 , 3 ) and evaluated at a ⊥ = ( 0.7 , − 1.3 ) a_\perp=(0.7,-1.3) a ⊥ = ( 0.7 , − 1.3 ) , a ⊥ ′ = ( 0.4 , 0.9 ) a'_\perp=(0.4,0.9) a ⊥ ′ = ( 0.4 , 0.9 ) for the head-on geometry at κ = 1 \kappa=1 κ = 1 . The entries are exact algebra evaluated at one point. The two columns outside the polarization plane, k ˉ ∧ k \bar{k}\wedge k k ˉ ∧ k and e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 , are emptied by the first commutator alone; the identities of Eq. (26) act only on the two k ∧ e i k\wedge e_i k ∧ e i columns. The sum reproduces 1 2 ( k − k ˉ ) ∧ a ′ = n ^ ∧ a ′ \frac{1}{2}(k-\bar{k})\wedge a'=\hat{n}\wedge a' 2 1 ( k − k ˉ ) ∧ a ′ = n ^ ∧ a ′ with a ′ = ( 0.4 , 0.9 ) a'=(0.4,0.9) a ′ = ( 0.4 , 0.9 ) .
The second Magnus term [29 , 30 ] gives the same leading result in three lines, and it fixes the sign. The rotation vector of Eq. (28) is m ( η ) = a e n ^ × a ⊥ ′ ( η ) \boldsymbol{m}(\eta)=a_{\mathrm e}\hat{n}\times a'_\perp(\eta) m ( η ) = a e n ^ × a ⊥ ′ ( η ) , which is a e a ⊥ ′ a_{\mathrm e}a'_\perp a e a ⊥ ′ turned through a right angle about n ^ \hat{n} n ^ . Turning both arguments of a cross product of planar vectors through the same right angle leaves its n ^ \hat{n} n ^ component alone, so the double integral may be written with a ⊥ ′ a'_\perp a ⊥ ′ directly,
Θ n e t = a e 2 2 ∬ η 1 < η 2 [ a ⊥ ′ ( η 2 ) × a ⊥ ′ ( η 1 ) ] ⋅ n ^ d η 1 d η 2 . (32) \Theta_{\mathrm{net}}=\frac{a_{\mathrm e}^{2}}{2}\iint_{\eta_1<\eta_2}\left[a'_\perp(\eta_2)\times a'_\perp(\eta_1)\right]\cdot\hat{n}\,d\eta_1d\eta_2. \tag*{(32)} Θ net = 2 a e 2 ∬ η 1 < η 2 [ a ⊥ ′ ( η 2 ) × a ⊥ ′ ( η 1 ) ] ⋅ n ^ d η 1 d η 2 . (32)
Integrating the inner variable and using a ⊥ ( − ∞ ) = 0 a_\perp(-\infty)=0 a ⊥ ( − ∞ ) = 0 converts the double integral into − ∫ ( a x a y ′ − a y a x ′ ) d η -\int(a_xa'_y-a_ya'_x)\,d\eta − ∫ ( a x a y ′ − a y a x ′ ) d η , so
Θ n e t = − 1 2 a e 2 A + O ( a e 4 ) , (33) \Theta_{\mathrm{net}}=-\frac{1}{2}a_{\mathrm e}^{2}\mathcal{A}+O(a_{\mathrm e}^{4}), \tag*{(33)} Θ net = − 2 1 a e 2 A + O ( a e 4 ) , (33)
in the triad e ^ 1 × e ^ 2 = n ^ \hat{e}_1\times\hat{e}_2=\hat{n} e ^ 1 × e ^ 2 = n ^ . This is Eq. (10) of the main text, and A \mathcal{A} A is twice the algebraic area enclosed by C C C .
The order of the remainder is a theorem, and the proof takes two lines. Grade s o ( 3 ) u 0 \mathfrak{so}(3)_{u_0} so ( 3 ) u 0 by calling G 1 G_1 G 1 and G 2 G_2 G 2 odd and e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 even. The bracket respects that grading, since [odd, odd] = even and [even, odd] = odd. Every term of the Magnus series is an integral of n n n -fold nested brackets of M M M , and M M M is odd, so the n n n th term has parity n n n : terms with n n n even lie along e 1 ∧ e 2 e_1\wedge e_2 e 1 ∧ e 2 , and terms with n n n odd lie in the plane spanned by G 1 G_1 G 1 and G 2 G_2 G 2 . The consequences follow immediately. The component of the exponent along n ^ \hat{n} n ^ carries only even powers of a e a_{\mathrm e} a e , which establishes the O ( a e 4 ) O(a_{\mathrm e}^{4}) O ( a e 4 ) in Eq. (33) ; the leading correction to the axis is the third-order in-plane term, so at fixed curve the axis is tilted out of n ^ \hat{n} n ^ by O ( a e ) O(a_{\mathrm e}) O ( a e ) ; and the change in the projection of the spin on n ^ \hat{n} n ^ is quadratic in that tilt and therefore O ( a e 6 ) O(a_{\mathrm e}^{6}) O ( a e 6 ) . Section S3.4 uses the last of the three. Without it that statement would rest on nothing.
Equation (S21) below sharpens the remainder further, since the relative correction there is governed by a e a 0 a_{\mathrm e}a_0 a e a 0 , so the informative form is Θ n e t = − 1 2 a e 2 A [ 1 + O ( ( a e a 0 ) 2 ) ] \Theta_{\mathrm{net}}=-\frac{1}{2}a_{\mathrm e}^{2}\mathcal{A}[1+O((a_{\mathrm e}a_0)^2)] Θ net = − 2 1 a e 2 A [ 1 + O (( a e a 0 ) 2 )] . The Magnus series itself, on the other hand, has a guaranteed radius of convergence only while a e ∫ ∥ M ∥ d η = a e length ( C ) < π a_{\mathrm e}\int\lVert M\rVert\,d\eta=a_{\mathrm e}\operatorname{length}(C)<\pi a e ∫ ∥ M ∥ d η = a e length ( C ) < π , and the working point of Sec. S6 violates that bound outright: at a 0 = 75 a_0=75 a 0 = 75 and N = 32 N=32 N = 32 the curve has length ≃ 2 π N a 0 / 2 ≃ 1.1 × 10 4 \simeq2\pi Na_0/\sqrt{2}\simeq1.1\times10^4 ≃ 2 π N a 0 / 2 ≃ 1.1 × 1 0 4 , so a e length ( C ) ≃ 12 ≫ π a_{\mathrm e}\operatorname{length}(C)\simeq12\gg\pi a e length ( C ) ≃ 12 ≫ π . Hence nothing quoted at that point rests on the series. What supports Eq. (33) is the resummation of Sec. S2.4, which sums all orders in a e a 0 a_{\mathrm e}a_0 a e a 0 for a circular plateau, together with the measured ratios of Sec. S4.5.
One definition has to be pinned down before Eq. (33) can be read literally. Its right-hand side is an unbounded real number growing as a e 2 a 0 2 N a_{\mathrm e}^{2}a_0^{2}N a e 2 a 0 2 N , whereas the angle extracted from an S O ( 3 ) SO(3) S O ( 3 ) matrix lies in [ 0 , π ] [0,\pi] [ 0 , π ] , is defined modulo 2 π 2\pi 2 π , and carries a sign ambiguity that goes with a reversal of the axis. We define Θ n e t \Theta_{\mathrm{net}} Θ net as the continuous lift in a e a_{\mathrm e} a e from a e = 0 a_{\mathrm e}=0 a e = 0 , equivalently as the angle accumulated along the path, and the measured quantity is its reduction modulo 2 π 2\pi 2 π . The wrapping is not hypothetical. For a four-cycle circular plateau with one-cycle fronts at a 0 = 3 a_0=3 a 0 = 3 and a e = 0.2 a_{\mathrm e}=0.2 a e = 0.2 , where A = 268.61 \mathcal{A}=268.61 A = 268.61 and a e a 0 = 0.6 a_{\mathrm e}a_0=0.6 a e a 0 = 0.6 is not small, Eq. (34) predicts 4.96 rad and the leading order Eq. (33) gives 5.37 rad; the angle read from the matrix is 1.307 rad, which is 2 π 2\pi 2 π minus 4.976 rad, and the extracted axis changes sign (measured; code/holonomy_scan.json, stage wrap). Section S4.5 drops the affected points from the scaling table for the same reason.
For linear polarization C C C is a segment and A = 0 \mathcal{A}=0 A = 0 , and the vanishing there is exact at every order in a e a_{\mathrm e} a e and a 0 a_0 a 0 by the fixed-axis argument of Sec. S2.2, well beyond the order of Eq. (33) . For a curve traced out and back, such as a y ∝ a x 2 a_y\propto a_x^2 a y ∝ a x 2 , the enclosed area is zero and the ordered exponential is the identity for the same reason. We measured ≤ 9.95 × 10 − 12 \leq9.95\times10^{-12} ≤ 9.95 × 1 0 − 12 rad over the 18 cases of the FIG. 1 grid, with anomalies up to a e = 0.6 a_{\mathrm e}=0.6 a e = 0.6 . The covariant cross-check of Sec. S4 returns ≤ 7.7 × 10 − 11 \leq7.7\times10^{-11} ≤ 7.7 × 1 0 − 11 rad on four cases of its own, at its coarser floor.
Resummation of the axis tilt for a circular plateau
Let a ⊥ a_\perp a ⊥ trace a circle of radius a 0 a_0 a 0 at unit rate in η \eta η over a phase interval Δ η \Delta\eta Δ η , so that A = a 0 2 Δ η \mathcal{A}=a_0^2\Delta\eta A = a 0 2 Δ η . Then n ^ × a ⊥ ′ \hat{n}\times a'_\perp n ^ × a ⊥ ′ has fixed magnitude a 0 a_0 a 0 and rotates about n ^ \hat{n} n ^ at unit rate, and Eq. (28) becomes the magnetic-resonance problem for a rotating field of strength a e a 0 a_{\mathrm e}a_0 a e a 0 on exact resonance. In the frame corotating about n ^ \hat{n} n ^ the rotation vector is constant with magnitude 1 + a e 2 a 0 2 \sqrt{1+a_{\mathrm e}^{2}a_0^{2}} 1 + a e 2 a 0 2 . Transforming back subtracts one turn per unit phase, and the net rotation about n ^ \hat{\boldsymbol{n}} n ^ has magnitude
Θ c i r c = ( 1 + a e 2 a 0 2 − 1 ) Δ η (34) \Theta_{\mathrm{circ}}=\left(\sqrt{1+a_e^2a_0^2}-1\right)\Delta\eta \tag*{(34)} Θ circ = ( 1 + a e 2 a 0 2 − 1 ) Δ η (34)
Eq. (12) of the main text, with the sense of rotation of Eq. (S20). Expanding for a e a 0 ≪ 1 a_ea_0\ll1 a e a 0 ≪ 1 returns 1 2 a e 2 a 0 2 Δ η = 1 2 a e 2 A \frac{1}{2}a_e^2a_0^2\Delta\eta=\frac{1}{2}a_e^2\mathcal{A} 2 1 a e 2 a 0 2 Δ η = 2 1 a e 2 A . At a 0 = 1 a_0=1 a 0 = 1 and a e = 0.05 a_e=0.05 a e = 0.05 the integrator gives 2.3193 × 10 − 2 2.3193\times10^{-2} 2.3193 × 1 0 − 2 , the resummation 2.3193 × 10 − 2 2.3193\times10^{-2} 2.3193 × 1 0 − 2 , and the area law 2.3208 × 10 − 2 2.3208\times10^{-2} 2.3208 × 1 0 − 2 (measured).
The word exact does not belong to Eq. (S21), and three conditions say why. An unbroken circle violates a ⊥ ( ± ∞ ) = 0 a_\perp(\pm\infty)=0 a ⊥ ( ± ∞ ) = 0 and is therefore not an admissible pulse, so Eq. (S21) describes a plateau and not a pulse. Composing the corotating rotation with the compensating turn returns a single angle only for an integer number of turns; away from that the two axes differ and the result is a rotation whose angle is not the difference of the two. And for a pulse with fronts the phase interval has to be read as A / a 0 2 \mathcal{A}/a_0^2 A / a 0 2 , which is the form printed in the main text; reading it as the plateau length instead costs 19% at a 0 = 1 a_0=1 a 0 = 1 and a e = 0.05 a_e=0.05 a e = 0.05 . What survives all three is an asymptotic improvement, not an identity: at a e a 0 = 0.3 a_ea_0=0.3 a e a 0 = 0.3 Eq. (S21) is still off by 0.1%. The corresponding statement about the axis is cleaner, and it is the one the section is named for. In the corotating frame the rotation vector is tilted from n ^ \hat{\boldsymbol{n}} n ^ by arctan ( a e a 0 ) \arctan(a_ea_0) arctan ( a e a 0 ) , which is the O ( a e ) O(a_e) O ( a e ) tilt that the grading argument of Sec. S2.3 predicts, resummed in a e a 0 a_ea_0 a e a 0 .
Note the convention. The radius of the circle in Eq. (S21) is a 0 a_0 a 0 , whereas in the ellipticity-normalized parameterization of Eq. (S2) a circularly polarized pulse has radius a 0 / 2 a_0/\sqrt{2} a 0 / 2 . The two differ by a factor of two in A \mathcal{A} A , and the numbers quoted elsewhere in this document use Eq. (S2).
The expansion parameter is a e a 0 a_ea_0 a e a 0 , not a e a_e a e . With the physical anomaly a e a 0 = 1.16 × 10 − 3 a 0 a_ea_0=1.16\times10^{-3}a_0 a e a 0 = 1.16 × 1 0 − 3 a 0 , so Eq. (S20) is accurate to 10 − 2 10^{-2} 1 0 − 2 up to a 0 ∼ 100 a_0\sim100 a 0 ∼ 100 . Above a e a 0 ∼ 1 a_ea_0\sim1 a e a 0 ∼ 1 the net rotation becomes linear in a e a_e a e , which no laboratory field reaches.
Invariance properties
Reparameterization of the phase
Equation (S17) contains η \eta η only through a ⊥ ( η ) a_\perp(\eta) a ⊥ ( η ) . Under a monotone reparameterization η → s ( η ) \eta\to s(\eta) η → s ( η ) that leaves the curve C \mathcal{C} C unchanged, the one-form n ^ ∧ d a ⊥ \hat{\boldsymbol{n}}\wedge d\boldsymbol{a}_\perp n ^ ∧ d a ⊥ is unchanged and so is Λ n e t \Lambda_{\mathrm{net}} Λ net . The statement holds by construction, so the numerical test is a control on the code and not evidence for the theorem: net rotations for one curve traced at different rates agree to eight or nine digits over eight runs, including at the inflated value a e = 0.25 a_e=0.25 a e = 0.25 , where the leading order of Eq. (S20) is far from sufficient (measured).
What drops out is the rate, and only the rate, so the carrier frequency, a chirp and the envelope shape have to be handled with care, because in general they change C \mathcal{C} C itself. The correct statement is that everything about the field enters through the curve C \mathcal{C} C exactly, and at leading order through A \mathcal{A} A alone. Lengthening the pulse at fixed a 0 a_0 a 0 enlarges the curve, adds turns and raises A \mathcal{A} A , so Θ n e t ∝ N \Theta_{\mathrm{net}}\propto N Θ net ∝ N . Changing the envelope at fixed N N N and fixed a 0 a_0 a 0 likewise changes A \mathcal{A} A , and Sec. S4.5 measures by how much: A = 334.412 \mathcal{A}=334.412 A = 334.412 for the Gaussian against 323.603 323.603 323.603 for cos 2 \cos^2 cos 2 at a 0 = 10 a_0=10 a 0 = 10 and N = 1 N=1 N = 1 , a difference of 3.2% that propagates straight into Θ n e t \Theta_{\mathrm{net}} Θ net . A quadratic chirp on a symmetric envelope leaves A \mathcal{A} A alone by parity, since the chirp-dependent part of φ ′ \varphi' φ ′ in A ∝ ∫ f 2 φ ′ d η \mathcal{A}\propto\int f^2\varphi'\,d\eta A ∝ ∫ f 2 φ ′ d η is odd about the center of an even f 2 f^2 f 2 . That is a control on the code and not a law.
Independence of the electron energy
The null rotation generated by N 1 = k ∧ e 1 N_1=k\wedge e_1 N 1 = k ∧ e 1 leaves k k k and κ \kappa κ fixed, shifts u 0 ⊥ u_{0\perp} u 0 ⊥ by − c κ e 1 -c\kappa e_1 − c κ e 1 for a parameter c c c , and changes the polarization vectors only by terms ∝ k μ \propto k^\mu ∝ k μ , so that the field tensor itself comes back unchanged, N f N − 1 = f NfN^{-1}=f N f N − 1 = f , which is stronger than gauge equivalence. Choosing c = u 0 ⊥ / k c=u_{0\perp}/k c = u 0 ⊥ / k brings any geometry to head-on, the net map is conjugated, Λ n e t → N Λ n e t N − 1 \Lambda_{\mathrm{net}}\to N\Lambda_{\mathrm{net}}N^{-1} Λ net → N Λ net N − 1 , and a rotation angle is a conjugation invariant. That disposes of the transverse momentum. The light-front projection κ \kappa κ needs a different argument, because κ = k ⋅ u \kappa=k\cdot u κ = k ⋅ u is a Lorentz invariant and no boost removes it. It appears solely through the scale chosen for k k k , hence as a rescaling η → η / κ \eta\to\eta/\kappa η → η / κ of the phase, to which the holonomy is insensitive by Sec. S3.1. Thus the inputs that remain are the curve C \mathcal{C} C and the number a e a_e a e , and γ \gamma γ is not among them.
Measured values agree. For circular polarization at a 0 = 5 a_0=5 a 0 = 5 and N = 1 N=1 N = 1 the net rotation is 5.621424 × 10 − 5 5.621424\times10^{-5} 5.621424 × 1 0 − 5 rad at γ = 1 \gamma=1 γ = 1 , 10 10 10 , 10 2 10^2 1 0 2 , 10 3 10^3 1 0 3 and 5.621423 × 10 − 5 5.621423\times10^{-5} 5.621423 × 1 0 − 5 at γ = 10 4 \gamma=10^4 γ = 1 0 4 , six matching digits over four decades in energy. With the anomaly switched on, an initial transverse momentum up to u 0 ⊥ = 17 u_{0\perp}=17 u 0 ⊥ = 17 leaves the angle at 1.1475916 × 10 − 5 1.1475916\times10^{-5} 1.1475916 × 1 0 − 5 , and a copropagating geometry with κ = 0.025 \kappa=0.025 κ = 0.025 gives the same. The g = 2 g=2 g = 2 null test of the main text was run to u 0 ⊥ = 15 u_{0\perp}=15 u 0 ⊥ = 15 . Only the rotation axis moves, and it tracks the propagation direction of the wave in the electron rest frame to four digits (measured).
Why the storage-ring factor a e γ a_e\gamma a e γ does not apply
In a storage ring the field is static in the laboratory, and the ratio of the anomalous precession frequency to the cyclotron frequency is a e γ a_e\gamma a e γ [16 , 38 , 39 ] . That ratio compares rates at one instant of laboratory time, and the enhancement comes from the rest-frame field being γ \gamma γ times the laboratory field at fixed B B B . Neither ingredient survives here. The strength of a plane wave in the rest frame is measured by the invariant a 0 a_0 a 0 , which no boost changes, and the g = 2 g=2 g = 2 part of the precession, which is what a e γ a_e\gamma a e γ multiplies, gives a net rotation of exactly zero. The product is a e γ ⋅ 0 = 0 a_e\gamma\cdot0 = 0 a e γ ⋅ 0 = 0 .
This is a question we have been asked more than once, so we state the practical form: the enhancement is real, but it applies to a quantity that vanishes identically in this geometry, and the surviving quantity is O ( a e 2 ) O(a_e^2) O ( a e 2 ) with no compensating factor, so that the one configuration in this paper where a e γ a_e\gamma a e γ does return is the static field superposed on the wave in § 3.7 , where the added field tensor no longer lies in k ∧ ( ⋅ ) k\wedge(\cdot) k ∧ ( ⋅ ) and the argument of Sec. S2.1 has nothing to act on.
Helicity
The grading argument of Sec. S2.3 settles the structure here. At fixed curve the rotation axis of Λ n e t \Lambda_{\mathrm{net}} Λ net departs from n ^ \hat{\boldsymbol{n}} n ^ by O ( a e ) O(a_e) O ( a e ) , the projection of the spin on the propagation direction changes at O ( a e 2 ) O(a_e^2) O ( a e 2 ) , and the observable is therefore the rotation of a transverse polarization. That fixes the experimental configuration, since a longitudinally polarized beam feels nothing at any order within reach. Section S2.4 adds one qualification: the natural expansion parameter is a e a 0 a_ea_0 a e a 0 , so at large a 0 a_0 a 0 the tilt should be read as O ( a e a 0 ) O(a_ea_0) O ( a e a 0 ) .
The exponent is measured, and not left as an inference from the grading. With the spin started along n ^ \hat{\boldsymbol{n}} n ^ , at γ = 1 \gamma=1 γ = 1 , N = 1 N=1 N = 1 and circular polarization, and with the anomaly inflated so that a e a 0 a_ea_0 a e a 0 covers 0.03 to 0.5, the transported triad gives Δ S ∥ = R 33 − 1 = − 1.80 × 10 − 3 ( a e a 0 ) 6 \Delta S_{\parallel}=R_{33}-1=-1.80\times10^{-3}(a_ea_0)^6 Δ S ∥ = R 33 − 1 = − 1.80 × 1 0 − 3 ( a e a 0 ) 6 , with local log-log exponents of 5.98 to 6.00 at the small-anomaly end and a coefficient that repeats to 0.5% over a tenfold range in a 0 a_0 a 0 . The mechanism was checked factor by factor: Θ n e t ∝ a e 2.000 \Theta_{\mathrm{net}}\propto a_e^{2.000} Θ net ∝ a e 2.000 , the axis tilt ψ ∝ a e 0.996 \psi\propto a_e^{0.996} ψ ∝ a e 0.996 , and Δ S ∥ = − ( 1 − cos Θ n e t ) sin 2 ψ \Delta S_{\parallel}=-(1-\cos\Theta_{\mathrm{net}})\sin^2\psi Δ S ∥ = − ( 1 − cos Θ net ) sin 2 ψ to 5 × 10 − 3 5\times10^{-3} 5 × 1 0 − 3 in relative terms. The orthogonality error of the transported triad is 1.5 × 10 − 14 1.5\times10^{-14} 1.5 × 1 0 − 14 , two decades below the smallest measured Δ S ∥ \Delta S_{\parallel} Δ S ∥ . Two limits are worth naming. The exponent belongs to the helicity state, since for a spin that does not start along n ^ \hat{\boldsymbol{n}} n ^ the projection on n ^ \hat{\boldsymbol{n}} n ^ changes at O ( a e 3 ) O(a_e^3) O ( a e 3 ) , measured exponent 3.00. The coefficient was measured at one pulse length, one Lorentz factor and one ellipticity, so it does not transfer to the working point of § 3.6 , where Θ n e t \Theta_{\mathrm{net}} Θ net is already 0.4 rad and the fit window used here has been left behind (code/spin_magnitude.json, key t7).
The net operator itself does not depend on the initial spin direction, since we transport a full orthonormal triad, and only the projection onto a chosen initial direction does.
The pulse condition, stated precisely
Maxwell’s equations and finiteness of the pulse require a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) a_{\perp}(+\infty)=a_{\perp}(-\infty) a ⊥ ( + ∞ ) = a ⊥ ( − ∞ ) , which is the absence of a dc component in the transverse electric field. They do not require ∫ a ⊥ d η = 0 \int a_{\perp}\,d\eta=0 ∫ a ⊥ d η = 0 , which is a condition on a net displacement. Forcing the second condition by subtracting a pedestal c f f ( η ) c_f f(\eta) c f f ( η ) with c f = e − σ 2 / 2 cos φ 0 c_f=e^{-\sigma^2/2}\cos\varphi_0 c f = e − σ 2 /2 cos φ 0 changes the net rotation by 2.3 × 10 − 5 2.3\times10^{-5} 2.3 × 1 0 − 5 in relative terms at a 0 = 10 a_0=10 a 0 = 10 , N = 1 N=1 N = 1 , δ = 1 \delta=1 δ = 1 (2.2485550 × 10 − 4 2.2485550\times10^{-4} 2.2485550 × 1 0 − 4 against 2.2484989 × 10 − 4 2.2484989\times10^{-4} 2.2484989 × 1 0 − 4 , measured). A pulse deliberately built with ∫ a ⊥ d η ≠ 0 \int a_{\perp}\,d\eta\ne0 ∫ a ⊥ d η = 0 obeys Eq. (33) to 10 − 5 10^{-5} 1 0 − 5 in relative terms at A = − 0.14159 A=-0.14159 A = − 0.14159 .
A pulse that violates the first condition is a different matter. It breaks the theorem at g = 2 g=2 g = 2 , and § 3.7 takes that case up.
Carrier-envelope phase
For strictly circular polarization a CEP shift is a rigid rotation of C C C about the origin. The connection of Eq. (27) is equivariant under rotations about n ^ \hat{\boldsymbol{n}} n ^ , and the holonomy angle is a class function, so the CEP drops out at every order in a e a_e a e , and the ten matching digits measured at a e = 0.25 a_e=0.25 a e = 0.25 are again a control on the code. For elliptical polarization a CEP-odd part appears only in higher orders of the holonomy: 8 × 10 − 5 8\times10^{-5} 8 × 1 0 − 5 in relative terms at a e = 0.25 a_e=0.25 a e = 0.25 , which extrapolates to ∼ 2 × 10 − 9 \sim2\times10^{-9} ∼ 2 × 1 0 − 9 relative and ∼ 10 − 14 \sim10^{-14} ∼ 1 0 − 14 rad absolute at the physical anomaly. The largest CEP-dependent part seen anywhere in the plane-wave scans is 2.4 × 10 − 10 2.4\times10^{-10} 2.4 × 1 0 − 10 rad, at δ = 0 \delta=0 δ = 0 and γ = 5000 \gamma=5000 γ = 5000 , consistent with the energy-dependent floor of Sec. S4.2 at that γ \gamma γ .
Numerical verification
Method
Two independent implementations were used. The first (code/spin_magnitude.py) works in the laboratory frame, writes the BMT equation componentwise in the form
d S μ d η = ( 1 + a e ) [ n μ ( a ′ ⋅ S ) − a ′ μ ( n ⋅ S ) ] + a e u μ [ ( n ⋅ S ) ( a ′ ⋅ u ) − ( a ′ ⋅ S ) ] , (35) \frac{dS^\mu}{d\eta}=(1+a_e)\left[n^\mu(a'\cdot S)-a'^\mu(n\cdot S)\right]+a_e\,u^\mu\left[(n\cdot S)(a'\cdot u)-(a'\cdot S)\right], \tag*{(35)} d η d S μ = ( 1 + a e ) [ n μ ( a ′ ⋅ S ) − a ′ μ ( n ⋅ S ) ] + a e u μ [ ( n ⋅ S ) ( a ′ ⋅ u ) − ( a ′ ⋅ S ) ] , (35)
with n μ = k μ / ( k ⋅ u 0 ) n^\mu=k^\mu/(k\cdot u_0) n μ = k μ / ( k ⋅ u 0 ) , and evaluates the orbit from Eq. (17) analytically at every step. The second (code/indep-check*.py) assembles the generator from covariant bivectors in a different coordinate system, with the wave along − z -z − z and the electron initially at rest. The two were written from the same equations but not from the same code, and they agree on every conclusion drawn below. They do not agree on how small a zero is: the second implementation reads the angle from the trace of the rotation matrix, an operation that is badly conditioned near the identity, and its floor is four to seven orders of magnitude coarser, with ∥ R − I ∥ \lVert R-\mathbb{I}\rVert ∥ R − I ∥ reaching 2.1 × 10 − 9 2.1\times10^{-9} 2.1 × 1 0 − 9 and the extracted angle reaching 6.6 × 10 − 6 6.6\times10^{-6} 6.6 × 1 0 − 6 rad on runs whose exact answer is zero. Every bound quoted below, and in particular the g = 2 g=2 g = 2 bound of Table S4 , comes from the first implementation. The covariant floor appears twice, in Sec. S2.3 and in the last column of Table S4 , and is named as such in both places.
TABLE S4. Validation of the plane-wave results. All bounds are measured. Cases counted are independent parameter combinations; the scan covers two envelopes, a 0 a_0 a 0 from 0.1 to 10, N N N from 0.5 to 16, Lorentz factors from 1 to 5 × 10 3 5 \times10^3 5 × 1 0 3 , three carrier-envelope phases and ellipticities from 0 to 1. The zero-area row is measured on the 18 such configurations of the grid behind Figure 1 (code/holonomy_scan.json); the value in parentheses there is the covariant cross-check of Sec. S4 on four cases of its own, at the coarser floor described in that section. Source data: code/spin_magnitude.json (keys t1, t3), code/holonomy_scan.json and code/indep_check*.py.
Table S4.
Table S2. Measured bounds on the exact invariants at the end of the pulse, taken over all configurations. The first two columns come from code/spin_magnitude.json; the last is the same quantity over the wider grid behind Figure 1 of the main text (code/holonomy_scan.json), which reaches larger A A A and is correspondingly coarser. All entries measured.
Both integrate with the DOP853 explicit Runge–Kutta scheme, the first at rtol = 10 − 13 = 10^{-13} = 1 0 − 13 and atol = 10 − 16 = 10^{-16} = 1 0 − 16 , the second at rtol = 10 − 12 = 10^{-12} = 1 0 − 12 and atol = 10 − 14 = 10^{-14} = 1 0 − 14 . Three orthonormal rest-frame four-vectors are transported together with the orbit. At the end the reference triad is rebuilt from the initial one, and the net rotation is read off the resulting 3 × 3 3 \times3 3 × 3 matrix as a rotation vector. Reading the angle from a transported frame is what makes the result independent of the initial polarization.
An earlier version of the plane-wave code integrated the four-velocity along with the spin. That version wasted about half its steps enforcing an identity that Eq. (17) already satisfies exactly, and it lost two digits at γ = 10 3 \gamma= 10^{3} γ = 1 0 3 . The runs reported here use the closed-form orbit.
Invariants and the resolution floor
The exact invariants are u ⋅ u = 1 u \cdot u = 1 u ⋅ u = 1 , S ⋅ u = 0 S \cdot u = 0 S ⋅ u = 0 , S ⋅ S = − 1 S \cdot S = -1 S ⋅ S = − 1 , and orthogonality of the transported triad, and Table S2 collects the measured bounds. One qualification belongs with them. They are evaluated at the end of the pulse, and nothing is recorded along the trajectory, so an excursion in the interior that came back would not appear.
Medians lie far below those maxima, 5.6 × 10 − 15 5.6 \times10^{-15} 5.6 × 1 0 − 15 for ∣ S ⋅ u ∣ |S \cdot u| ∣ S ⋅ u ∣ and 4.4 × 10 − 15 4.4 \times10^{-15} 4.4 × 1 0 − 15 for ∣ S ⋅ S + 1 ∣ |S \cdot S + 1| ∣ S ⋅ S + 1∣ at γ = 1 \gamma= 1 γ = 1 . Table S2 reports worst cases over the whole scan. Truncation of the Gaussian envelope enters at the same level, ∣ a ⊥ ( ± η max ) ∣ ≤ 1.24 × 10 − 13 |a_{\perp}(\pm\eta_{\max})| \leq1.24 \times10^{-13} ∣ a ⊥ ( ± η m a x ) ∣ ≤ 1.24 × 1 0 − 13 , and vanishes identically for cos 2 \cos^{2} cos 2 .
The invariant error grows with energy, because the laboratory components of S S S and u u u scale as γ \gamma γ while the invariants are built from cancellations among them. A cancellation estimate would give γ 2 ϵ m a c h \gamma^{2}\epsilon_{\mathrm{mach}} γ 2 ϵ mach , and the measured growth is slower than that: from 1.65 × 10 − 12 1.65 \times10^{-12} 1.65 × 1 0 − 12 at γ = 1 \gamma= 1 γ = 1 to 3.6 × 10 − 7 3.6 \times10^{-7} 3.6 × 1 0 − 7 at γ = 5 × 10 3 \gamma= 5 \times10^{3} γ = 5 × 1 0 3 , an effective exponent of 1.44 over the range, approaching 2 only above γ ≃ 10 2 \gamma\simeq10^{2} γ ≃ 1 0 2 . The net rotation at g = 2 g = 2 g = 2 behaves differently, and the difference matters for how the null results are read, since that quantity does not grow with γ \gamma γ at all but stays flat at 2.3 2.3 2.3 –2.6 × 10 − 13 2.6 \times10^{-13} 2.6 × 1 0 − 13 rad across the whole range, being set by the extraction of a rotation vector from a matrix close to the identity. It scales with intensity instead, the median
Figure S1. Measured net spin rotation against the Lorentz factor. Filled circles are the g = 2 g = 2 g = 2 scan and filled squares the linear-polarization scan at inflated anomaly, both of which are zero by Eq. (24) and by the fixed-axis argument of Sec. S2.2, so what is plotted for them is round-off. The g = 2 g = 2 g = 2 maximum is flat at 2.3 2.3 2.3 –2.6 × 10 − 13 2.6 \times10^{-13} 2.6 × 1 0 − 13 rad, set by the extraction of the rotation vector from an S O ( 3 ) SO(3) S O ( 3 ) matrix, and the growth with γ \gamma γ is carried by the anomaly-inflated scan and by the measured invariant error ∥ R T R − I ∥ \lVert R^{\mathsf{T}}R-\mathbb{I}\rVert ∥ R T R − I ∥ of the same runs (grey triangles), which rises from 1.65 × 10 − 12 1.65 \times10^{-12} 1.65 × 1 0 − 12 at γ = 1 \gamma= 1 γ = 1 to 3.6 × 10 − 7 3.6 \times10^{-7} 3.6 × 1 0 − 7 at γ = 5 × 10 3 \gamma= 5 \times10^{3} γ = 5 × 1 0 3 . The dashed line is the γ 2 \gamma^{2} γ 2 cancellation estimate anchored on the measured value at γ = 1 \gamma= 1 γ = 1 , and the dotted line is the power law that error actually follows; the data fall below the estimate by up to two decades, the effective exponent being 1.44. Open symbols are the circularly polarized case at a 0 = 5 a_{0} = 5 a 0 = 5 , N = 1 N = 1 N = 1 , flat to six digits over four decades in γ \gamma γ . That estimate meets the signal at the cross near γ ≃ 5.8 × 10 3 \gamma\simeq5.8 \times10^{3} γ ≃ 5.8 × 1 0 3 , which marks where the laboratory formulation stops resolving it.
rising roughly as a 0 a_{0} a 0 , from 7.4 × 10 − 16 7.4 \times10^{-16} 7.4 × 1 0 − 16 at a 0 = 0.1 a_{0} = 0.1 a 0 = 0.1 through 5.0 × 10 − 15 5.0 \times10^{-15} 5.0 × 1 0 − 15 at a 0 = 1 a_{0} = 1 a 0 = 1 to 5.0 × 10 − 14 5.0 \times10^{-14} 5.0 × 1 0 − 14 at a 0 = 10 a_{0} = 10 a 0 = 10 , which is the signature of accumulated round-off. Figure S1 separates the two behaviors. The precision scans were run at γ = 1 \gamma= 1 γ = 1 , which Sec. S3.2 shows to be equivalent.
Convergence in the integrator tolerance
A null result is worth only as much as the evidence that the step size has stopped mattering, so the plane-wave zero was rerun over the integrator tolerance at a e = 0 a_{e} = 0 a e = 0 , with rtol from 10 − 9 10^{-9} 1 0 − 9 to 10 − 14 10^{-14} 1 0 − 14 and atol from 10 − 3 10^{-3} 1 0 − 3 to 10 − 12 10^{-12} 1 0 − 12 . Table S3 gives three configurations, and they read in opposite directions.
At a 0 = 10 a_{0} = 10 a 0 = 10 the zero has reached the arithmetic. Tightening the tolerance by two decades from 10 − 12 10^{-12} 1 0 − 12 to 10 − 14 10^{-14} 1 0 − 14 does not lower Θ n e t \Theta_{\mathrm{net}} Θ net , which instead fluctuates over ( 1.0 (1.0 ( 1.0 –2.6 ) × 10 − 13 2.6) \times10^{-13} 2.6 ) × 1 0 − 13 while the number of right-hand-side evaluations triples, and that is the evidence that the vanishing is a property of the equations. The same fact costs us two digits of the headline number. At r t o l = 10 − 12 \mathrm{rtol} = 10^{-12} rtol = 1 0 − 12 the identical computation returns 2.0 × 10 − 13 2.0 \times10^{-13} 2.0 × 1 0 − 13 and at 10 − 14 10^{-14} 1 0 − 14 it returns 2.4 × 10 − 13 2.4 \times10^{-13} 2.4 × 1 0 − 13 , a spread of 22.5% below the value 2.57 × 10 − 13 2.57 \times10^{-13} 2.57 × 1 0 − 13 obtained at the working tolerance, so three significant figures are not secured and the defensible statement is the bound ≤ 3 × 10 − 13 \le3 \times10^{-13} ≤ 3 × 1 0 − 13 rad used in Table S4 . At a 0 ≲ 1 a_0 \lesssim1 a 0 ≲ 1 the picture inverts: the measured value still falls monotonically with the tolerance, from 4.4 × 10 − 14 4.4 \times10^{-14} 4.4 × 1 0 − 14 to 2.3 × 10 − 16 2.3 \times10^{-16} 2.3 × 1 0 − 16 , so over that part of the grid the reported zero is an upper bound set by truncation and not a round-off floor. Little room is left in either direction, since SciPy clamps rtol below 2.22 × 10 − 14 2.22 \times10^{-14} 2.22 × 1 0 − 14 and the working value already sits one decade above that.
Table S3. Measured ∣ Θ n e t ∣ |\Theta_{\mathrm{net}}| ∣ Θ net ∣ in radians at g = 2 g = 2 g = 2 against the relative tolerance of the DOP853 integrator, at a t o l = 10 − 3 r t o l \mathrm{atol} = 10^{-3}\mathrm{rtol} atol = 1 0 − 3 rtol . The working tolerance of the runs reported elsewhere is r t o l = 10 − 13 \mathrm{rtol} = 10^{-13} rtol = 1 0 − 13 . Each row is at one carrier-envelope phase, φ 0 = π / 2 \varphi_0 = \pi/2 φ 0 = π /2 for the first and φ 0 = 0.7 \varphi_0 = 0.7 φ 0 = 0.7 for the other two. SciPy raises any relative tolerance below 100 machine epsilons, so the last column is executed at 2.22 × 10 − 14 2.22 \times10^{-14} 2.22 × 1 0 − 14 . Sixteen of the eighteen entries reproduce an earlier run to the digits printed. The two that do not are both in the loosest column, 10 − 9 10^{-9} 1 0 − 9 : the third row differs by rounding, and the second returns 9.8 × 10 − 14 9.8 \times10^{-14} 9.8 × 1 0 − 14 against 3.0 × 10 − 13 3.0 \times10^{-13} 3.0 × 1 0 − 13 before, a factor of three whose original configuration could not be recovered. Both cells carry the recomputed value, and neither enters any bound quoted here. Source: code/spin_magnitude.json, key t6.
Validation runs
Table S4 collects the tests. The entry that carries the most weight is the second: the vanishing at g = 2 g = 2 g = 2 survives when the anomaly is switched on and inflated by four orders of magnitude above its physical value, which rules out the reading “the effect is there but small.”
Scaling laws
How well Eq. (33) reproduces the measured angle away from linear polarization depends on which scan is meant, and the two answers differ by two digits. Over the wide grid behind Figure 1 of the main text, which mixes envelopes, ellipticities, chirps and polarization gating, the relative residual has median 8.0 × 10 − 5 8.0 \times10^{-5} 8.0 × 1 0 − 5 and maximum 9.1 × 10 − 4 9.1 \times10^{-4} 9.1 × 1 0 − 4 , that is four digits typically and three at worst. Along a single parameter axis the agreement is the five digits quoted in the main text, and Table S5 gives the ratio of measurement to prediction axis by axis. The ratio Θ n e t / a e 2 \Theta_{\mathrm{net}}/a_e^2 Θ net / a e 2 stays constant to 0.8% across a 26-fold range in the anomaly, which is the direct statement that the effect is quadratic. Points with a e ≥ 0.1 a_e \ge0.1 a e ≥ 0.1 are excluded here. Beyond that value Θ n e t \Theta_{\mathrm{net}} Θ net exceeds π \pi π and the extracted rotation vector wraps, in the sense of Sec. S2.3. For a circularly polarized Gaussian pulse the law evaluates to Θ n e t = 2.2486 × 10 − 6 a 0 2 N \Theta_{\mathrm{net}} = 2.2486 \times10^{-6}a_0^2N Θ net = 2.2486 × 1 0 − 6 a 0 2 N rad, and for cos 2 \cos^2 cos 2 2.1759 × 10 − 6 a 0 2 N 2.1759 \times10^{-6}a_0^2N 2.1759 × 1 0 − 6 a 0 2 N rad. The 3.2% difference is entirely ∫ f 2 d η \int f^2 d\eta ∫ f 2 d η : at a 0 = 10 a_0 = 10 a 0 = 10 and N = 1 N = 1 N = 1 the two areas are A = 334.412 \mathcal{A} = 334.412 A = 334.412 and 323.603 (measured).
TABLE S5. Measured net rotation divided by − 1 2 a e 2 A -\frac{1}{2}a_e^2 A − 2 1 a e 2 A , along one parameter axis at a time. Circular polarization, Gaussian envelope, γ = 10 3 \gamma= 10^3 γ = 1 0 3 ; the fixed values are a 0 = 10 a_0 = 10 a 0 = 10 and N = 4 N = 4 N = 4 , except in the first row, where N = 1 N = 1 N = 1 . The last row instead lists Θ n e t / a e 2 \Theta_{\rm net}/a_e^2 Θ net / a e 2 , which Eq. (33) requires to equal A / 2 = 668.8 A/2 = 668.8 A /2 = 668.8 . Source: code/spin_magnitude.json, key t3.
Table S5.
Finite focusing
Field model
Equation (S20) rests on A = A ( η ) A = A(\eta) A = A ( η ) , which no focused beam satisfies. The field used below is a Gaussian pulse built from the Lax–Louissell–McKnight (LLM) expansion in the diffraction parameter ε = 1 / ( k w 0 ) = λ / ( 2 π w 0 ) \varepsilon= 1/(kw_0) = \lambda/(2\pi w_0) ε = 1/ ( k w 0 ) = λ / ( 2 π w 0 ) , in the form standard in the strong-field literature [43 , 44 , 59 ] . With s = ε 2 ( X 2 + Y 2 ) s = \varepsilon^2(X^2 + Y^2) s = ε 2 ( X 2 + Y 2 ) , ζ = 2 ε 2 Z \zeta= 2\varepsilon^2 Z ζ = 2 ε 2 Z and F = 1 / ( 1 + i ζ ) F = 1/(1 + i\zeta) F = 1/ ( 1 + i ζ ) , the transverse complex amplitude is carried to third order,
ψ = ψ 0 + ε 2 ψ 2 , ψ 0 = F e − F s , (36) \psi= \psi_0 + \varepsilon^2\psi_2,\qquad\psi_0 = F e^{-Fs}, \tag*{(36)} ψ = ψ 0 + ε 2 ψ 2 , ψ 0 = F e − F s , (36)
ψ 2 = e − F s [ c ψ F 2 + ( 2 − c ψ ) F 3 s − F 4 s 2 ] . (37) \psi_2 = e^{-Fs}\left[c_\psi F^2 + (2-c_\psi)F^3s - F^4s^2\right]. \tag*{(37)} ψ 2 = e − F s [ c ψ F 2 + ( 2 − c ψ ) F 3 s − F 4 s 2 ] . (37)
We rederived ψ 2 \psi_2 ψ 2 rather than copying it: substitution into the wave equation gives s ψ 2 , s s + ψ 2 , s + i ψ 2 , ζ = − ∂ ζ 2 ψ 0 s\psi_{2,ss}+\psi_{2,s}+i\psi_{2,\zeta}=-\partial_\zeta^2\psi_0 s ψ 2 , ss + ψ 2 , s + i ψ 2 , ζ = − ∂ ζ 2 ψ 0 , whose general solution is Eq. (37) with one free constant c ψ c_\psi c ψ , multiplying ∂ ζ ψ 0 \partial_\zeta\psi_0 ∂ ζ ψ 0 , a homogeneous paraxial solution. That constant is the O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) waist-shift ambiguity built into the expansion. We work at c ψ = 0 c_\psi= 0 c ψ = 0 , where the correction vanishes on axis and a 0 a_0 a 0 keeps its meaning as the peak potential at focus. The spread over c ψ c_\psi c ψ is carried as a systematic error.
The longitudinal component is obtained by closing the Coulomb gauge ∇ ⋅ a = 0 \nabla\cdot\mathbf{a} = 0 ∇ ⋅ a = 0 exactly, by the fixed-point iteration W = − i ( D y + ∂ Z W ) W = -i(D_y + \partial_Z W) W = − i ( D y + ∂ Z W ) with W 0 = i D y W_0 = iD_y W 0 = i D y and D = α y χ X + α y ψ Y D = \alpha_y\chi_X + \alpha_y\psi_Y D = α y χ X + α y ψ Y . Two iterations give a longitudinal field complete through ε 3 \varepsilon^3 ε 3 and through ε / σ 2 \varepsilon/\sigma^2 ε / σ 2 . Then E = − ∂ t a \mathbf{E} = -\partial_t\mathbf{a} E = − ∂ t a and B = ∇ × a \mathbf{B} = \nabla\times\mathbf{a} B = ∇ × a . At retained order 1 the measured ratio E z / E ⊥ E_z/E_\perp E z / E ⊥ grows as ε 1.07 \varepsilon^{1.07} ε 1.07 from 0.0174 at ε = 0.025 \varepsilon= 0.025 ε = 0.025 , and with ψ 2 \psi_2 ψ 2 retained it reaches 0.276 at ε = 0.30 \varepsilon= 0.30 ε = 0.30 , so the model carries a longitudinal field of the right order.
Maxwell residuals
Two of Maxwell’s equations hold identically, because the fields are built from a potential: ∇ ⋅ B \nabla\cdot B ∇ ⋅ B and ∇ × E + ∂ t B \nabla\times E+\partial_t B ∇ × E + ∂ t B are machine zero at all sampled points. The other two are measured on 48 points fixed in the self-similar variables ( ρ / ω 0 , z / ρ , η / σ ) (\rho/\omega_0,z/\rho,\eta/\sigma) ( ρ / ω 0 , z / ρ , η / σ ) , so that different ε \varepsilon ε are compared at the same place in the beam, with residuals normalized to a 0 a_0 a 0 . Table S6 gives them for a quasimonochromatic run, σ = 100 \sigma= 100 σ = 100 , which isolates the LLM hierarchy from the envelope.
TABLE S6. Maxwell residuals of the field model, normalized to a 0 a_0 a 0 , for a quasimonochromatic pulse (σ = 100 \sigma= 100 σ = 100 ) at retained order 1 and order 3 in ε \varepsilon ε . ∇ ⋅ B = 0 \nabla\cdot B = 0 ∇ ⋅ B = 0 and Faraday’s law are satisfied identically and are not listed. All entries measured. Source: code/focused_beam.json, stage field.
Table S6.
The Gauss residual falls as ε 5.5 \varepsilon^{5.5} ε 5.5 , a consequence of the exact Coulomb closure, and the Ampère residual as ε 3.5 \varepsilon^{3.5} ε 3.5 at order 1. Keeping ψ 2 \psi_2 ψ 2 lowers the Ampère residual by a factor 4.6 to 11 over ε = 0.1 \varepsilon= 0.1 ε = 0.1 –0.2 0.2 0.2 , which is the hierarchy working as intended.
For a physical pulse the picture is worse and we state it plainly. At σ = 3.77 \sigma= 3.77 σ = 3.77 (N = 1 N = 1 N = 1 ) the residuals fall only as ε 1.4 \varepsilon^{1.4} ε 1.4 for Gauss and ε 2 \varepsilon^2 ε 2 for Ampère, and at ε = 0.15 \varepsilon= 0.15 ε = 0.15 they reach 1.4 × 10 − 3 1.4 \times10^{-3} 1.4 × 1 0 − 3 and 7.2 × 10 − 3 7.2 \times10^{-3} 7.2 × 1 0 − 3 of a 0 a_0 a 0 . The finite spectral width produces terms ∝ ε p / σ q \propto\varepsilon^{p}/\sigma^q ∝ ε p / σ q from the product of the z z z -dependent mode with the envelope in η \eta η , and no finite order in ε \varepsilon ε removes them, so this is a property of the pulse and not a defect of the LLM hierarchy. An isodiffracting envelope would remove them, and we did not try one. The practical consequence appears in Sec. S5.6: any measured quantity below 10 − 3 a 0 10^{-3}a_0 1 0 − 3 a 0 is a property of the model.
In the opposite limit the model reduces correctly. At ε ≤ 10 − 2 \varepsilon\le10^{-2} ε ≤ 1 0 − 2 the field agrees with the exact plane wave of Eq. (15) to 1.1 × 10 − 16 1.1 \times10^{-16} 1.1 × 1 0 − 16 at five phases (measured).
Dynamics and observable
The Lorentz and BMT equations are integrated together in laboratory time,
x ˙ = β , u ˙ = E + β × B , \dot{x}=\beta,\qquad\dot{u}=E+\beta\times B, x ˙ = β , u ˙ = E + β × B ,
S ˙ μ = 1 γ [ ( 1 + a e ) ( f S ) μ + a e u μ ( S ⋅ f u ) ] , (38) \dot{S}^{\mu}=\frac{1}{\gamma}\left[(1+a_e)(fS)^{\mu}+a_e u^{\mu}(S\cdot f u)\right], \tag*{(38)} S ˙ μ = γ 1 [ ( 1 + a e ) ( f S ) μ + a e u μ ( S ⋅ f u ) ] , (38)
with u 0 = 1 + u 2 u^0=\sqrt{1+u^2} u 0 = 1 + u 2 imposed algebraically, so u ⋅ u = 1 u\cdot u=1 u ⋅ u = 1 holds exactly. Working tolerances are rtol = 10 − 11 =10^{-11} = 1 0 − 11 and atol = 10 − 13 =10^{-13} = 1 0 − 13 , with rtol = 10 − 13 =10^{-13} = 1 0 − 13 as the convergence reference. Along a trajectory sampled at 401 points we measured ∣ u ⋅ u − 1 ∣ ≤ 2.8 × 10 − 14 |u\cdot u-1|\le2.8\times10^{-14} ∣ u ⋅ u − 1∣ ≤ 2.8 × 1 0 − 14 , ∣ S ⋅ u ∣ ≤ 2.8 × 10 − 12 |S\cdot u|\le2.8\times10^{-12} ∣ S ⋅ u ∣ ≤ 2.8 × 1 0 − 12 and ∣ S ⋅ S + 1 ∣ ≤ 2.3 × 10 − 12 |S\cdot S+1|\le2.3\times10^{-12} ∣ S ⋅ S + 1∣ ≤ 2.3 × 1 0 − 12 , six to seven orders below every quantity reported here.
The observable is therefore the spin rotation relative to the momentum. Three rest-frame four-vectors are transported, a reference triad is constructed at the exit by boosting the entrance triad along u 0 → u f u_0\to u_f u 0 → u f , and the rotation vector r r r is read from R i j = ⟨ e ~ i , e j ( t f ) ⟩ R_{ij}=\langle\tilde e_i,e_j(t_f)\rangle R ij = ⟨ e ~ i , e j ( t f )⟩ . Removing that boost removes with it the kinematic transport that accompanies ponderomotive deflection, which would otherwise dominate everything. In the plane-wave limit u f = u 0 u_f = u_0 u f = u 0 , the boost is the identity, and r r r reduces to the net rotation of Sec. S4.
Table S7. Focused Gaussian pulse: circular polarization, a 0 = 1 a_0 = 1 a 0 = 1 , N = 1 N = 1 N = 1 , φ 0 = 0 \varphi_0 = 0 φ 0 = 0 , head-on collision at γ = 10 \gamma= 10 γ = 10 . The plane-wave reference is − 1 2 A = − 1.6720609 -\frac{1}{2}\mathcal{A} = -1.6720609 − 2 1 A = − 1.6720609 . All entries measured. Source: code/focused_beam.json, stage eps.
Table S8. Waist above which Eq. (S20) holds to the stated accuracy, at a 0 ≃ 1 a_0 \simeq1 a 0 ≃ 1 , N = 1 N = 1 N = 1 , circular polarization, from ε < a c / c \varepsilon< \sqrt{ac}/c ε < a c / c and w 0 / λ = 1 / ( 2 π ε ) w_0/\lambda= 1/(2\pi\varepsilon) w 0 / λ = 1/ ( 2 π ε ) . The γ = 10 \gamma= 10 γ = 10 column uses the directly measured c = 10.7 c = 10.7 c = 10.7 of Eq. (S28); the other two are estimated from the measured c 1 c_1 c 1 , giving c = 3.86 × 10 3 c = 3.86 \times10^3 c = 3.86 × 1 0 3 and c = 6.8 c = 6.8 c = 6.8 . Every entry is a lower bound from the field model at c ψ = 0 c_\psi= 0 c ψ = 0 , and Sec. S5.7 shows that c c c carries a systematic factor of about three, which multiplies each waist by 3 \sqrt{3} 3 : the 1% row then reads 17 λ 17\lambda 17 λ , 9.0 λ 9.0\lambda 9.0 λ and 7.2 λ 7.2\lambda 7.2 λ .
Five triads sharing a single orbit, at anomalies a e ∈ { 0 , ± h , ± 2 h } a_e \in\{0, \pm h, \pm2h\} a e ∈ { 0 , ± h , ± 2 h } with h = 0.05 h = 0.05 h = 0.05 , are separated by five-point differences into
r ( a e ) = r 0 + a e r 1 + a e 2 r 2 + O ( a e 3 ) . (39) r(a_e) = r_0 + a_e r_1 + a_e^2 r_2 + O(a_e^3). \tag*{(39)} r ( a e ) = r 0 + a e r 1 + a e 2 r 2 + O ( a e 3 ) . (39)
Both r 0 r_0 r 0 and r 1 r_1 r 1 vanish identically in a plane wave, the first by Eq. (24) and the second by the argument of Sec. S2.2.
The limit ε → 0 \varepsilon\to0 ε → 0
Table S7 gives the scan. All three violations enter at order ε 2 \varepsilon^2 ε 2 , and the range over which that is true has to be stated with the exponents. Over ε = 0.025 \varepsilon= 0.025 ε = 0.025 –0.15 0.15 0.15 the local log-log slopes run 2.71 → 2.02 for ∣ r 0 ∣ |r_0| ∣ r 0 ∣ , 2.71 → 2.14 for ∣ r 1 ∣ |r_1| ∣ r 1 ∣ , 1.92 → 2.00 for the deviation of r 2 r_2 r 2 from its plane-wave value, and 3.11 → 2.06 for the transverse momentum kick Δ u ⊥ \Delta u_\perp Δ u ⊥ . Over the whole scanned range the fitted exponents at γ = 10 \gamma= 10 γ = 10 are 2.87 for ∣ r 0 ∣ |r_0| ∣ r 0 ∣ and 2.89 for ∣ r 1 ∣ |r_1| ∣ r 1 ∣ , and the local slopes on the last two intervals reach 4.0 and 6.3, so the ε 2 \varepsilon^2 ε 2 law is a statement about ε ≲ 0.15 \varepsilon\lesssim0.15 ε ≲ 0.15 and the two widest points of Table S7 are order-of-magnitude entries. The deviation of the holonomic coefficient is
∣ r 2 + 1 2 A n ^ ∣ 1 2 A ≃ 1.05 ε 2 ( γ = 10 ) , (40) \frac{\left|r_2 + \frac{1}{2}\mathcal{A}\hat{\boldsymbol{n}}\right|}{\frac{1}{2}\mathcal{A}} \simeq1.05\varepsilon^2 \qquad(\gamma= 10), \tag*{(40)} 2 1 A r 2 + 2 1 A n ^ ≃ 1.05 ε 2 ( γ = 10 ) , (40)
measured as 1.06, 1.05, 1.05, 1.04, 1.03, 0.995 at ε = 0.025 \varepsilon= 0.025 ε = 0.025 –0.15 0.15 0.15 . At γ = 1 \gamma= 1 γ = 1 the same law holds with coefficient 3.4. The coefficient, unlike the exponent, carries a systematic error of a factor of about three from the field model itself, and Sec. S5.7 propagates it.
For linear polarization r 2 r_2 r 2 stays at ≤ 7.5 × 10 − 11 \le7.5 \times10^{-11} ≤ 7.5 × 1 0 − 11 for every ε \varepsilon ε up to 0.30 (measured). Nothing in Eq. (33) guarantees that outside the plane-wave limit, and we have no argument for why it holds. It is an observation.
A direct run at the physical anomaly a e = 1.15965 × 10 − 3 a_e = 1.15965 \times10^{-3} a e = 1.15965 × 1 0 − 3 , without finite differences, gives the full anomalous observable. The ratio to the plane-wave value is 1.0067, 1.0271 and 1.3033 at ε = 0.025 \varepsilon= 0.025 ε = 0.025 , 0.050 and 0.150, that is
Θ a n o m ( ε ) 1 2 a e 2 A − 1 ≃ 10.7 a 0 ε 2 ( γ = 10 ) , (41) \frac{\Theta_{\mathrm{anom}}(\varepsilon)}{\frac{1}{2}a_e^2\mathcal{A}} - 1 \simeq\frac{10.7}{a_0}\varepsilon^2 \qquad(\gamma= 10), \tag*{(41)} 2 1 a e 2 A Θ anom ( ε ) − 1 ≃ a 0 10.7 ε 2 ( γ = 10 ) , (41)
with the coefficient measured as 10.7, 10.8 and 13.5. Equation (13) of the main text quotes the same numbers. The deviation is accounted for by the recovered term a e r 1 a_e r_1 a e r 1 alone, since c ε 2 / ( a e A / 2 ) = 11.0 ε 2 c\varepsilon^2/(a_e\mathcal{A}/2) = 11.0\varepsilon^2 c ε 2 / ( a e A /2 ) = 11.0 ε 2 from the measured c 1 c_1 c 1 .
The amplitudes depend strongly on energy while the exponents do not. Measured ∣ r 0 ∣ = c 0 ε 2 |r_0| = c_0\varepsilon^2 ∣ r 0 ∣ = c 0 ε 2 and ∣ r 1 ∣ = c 1 ε 2 |r_1| = c_1\varepsilon^2 ∣ r 1 ∣ = c 1 ε 2 give c 0 = 7.22 c_0 = 7.22 c 0 = 7.22 , c 1 = 7.48 c_1 = 7.48 c 1 = 7.48 at γ = 1 \gamma= 1 γ = 1 ; c 0 = 0.0230 c_0 = 0.0230 c 0 = 0.0230 , c 1 = 0.0214 c_1 = 0.0214 c 1 = 0.0214 at γ = 10 \gamma= 10 γ = 10 ; and c 0 = 0.0165 c_0 = 0.0165 c 0 = 0.0165 , c 1 = 0.0165 c_1 = 0.0165 c 1 = 0.0165 at γ = 10 3 \gamma= 10^3 γ = 1 0 3 , a 436-fold range in c 0 c_0 c 0 . Translating Eq. (41) into a waist requirement gives Table S8.
The dependence on intensity runs the other way. The measured ratio ∣ a e r 1 ∣ / ∣ a e 2 r 2 ∣ |a_e r_1|/|a_e^2 r_2| ∣ a e r 1 ∣/∣ a e 2 r 2 ∣ at ε = 0.15 \varepsilon= 0.15 ε = 0.15 and γ = 10 \gamma= 10 γ = 10 is 1.56, 0.79, 0.41, 0.25, 0.28 at a 0 = 0.25 a_0 = 0.25 a 0 = 0.25 , 0.5, 1, 2, 4, falling as 1 / a 0 1/a_0 1/ a 0 up to a 0 ≃ 2 a_0 \simeq2 a 0 ≃ 2 and flattening beyond it. Therefore the holonomy formula gains accuracy as the intensity rises, by 1 / a 0 1/a_0 1/ a 0 over the range where the law holds, and we have not scanned ε \varepsilon ε at a 0 ≳ 10 a_0 \gtrsim10 a 0 ≳ 10 .
What the dominant term is
The largest of the three coefficients is r 0 r_0 r 0 , and it carries no information about g g g , since it is computed from the orbit alone. At γ = 10 3 \gamma= 10^3 γ = 1 0 3 we measured ∣ r 0 ∣ = ∣ Δ u ⊥ ∣ |r_0| = |\Delta u_\perp| ∣ r 0 ∣ = ∣Δ u ⊥ ∣ to four digits with an axis orthogonal to the momentum, ∣ r 0 ∣ / ∣ r ⊥ ∣ < 6 × 10 − 5 |r_0|/|r_\perp| < 6 \times10^{-5} ∣ r 0 ∣/∣ r ⊥ ∣ < 6 × 1 0 − 5 , and that identifies r 0 r_0 r 0 as the Thomas–Wigner rotation [15 , 23 ] accompanying the ponderomotive kick, whose existence in a focused beam follows from the failure of the Lawson–Woodward theorem [40 , 41 ] . The separation of the axes is already much weaker one decade down, 5.7 × 10 − 3 5.7 \times10^{-3} 5.7 × 1 0 − 3 at γ = 10 2 \gamma= 10^2 γ = 1 0 2 , and it fails outright below γ ≃ 3 \gamma\simeq3 γ ≃ 3 , where r 0 r_0 r 0 becomes mostly parallel to the momentum (0.148 against 0.014 transverse at γ = 1 \gamma= 1 γ = 1 , ε = 0.15 \varepsilon= 0.15 ε = 0.15 , a 0 = 1 a_0 = 1 a 0 = 1 ) and exceeds Δ u ⊥ \Delta u_\perp Δ u ⊥ by a factor 10 to 50 because the electric term of Eq. (S6) dominates. That is the same corner of the parameter space in which Sec. S6 places the best classical working point, and the two requirements pull against each other.
The net rotation here is nonzero already at g = 2 g=2 g = 2 , and the axis lies perpendicular to the momentum, so the projection of the spin on the momentum loses the protection it had in a plane wave. A longitudinally polarized beam does depolarize in a focused pulse, by 8.0 × 10 − 4 r a d 8.0 \times10^{-4}\,\mathrm{rad} 8.0 × 1 0 − 4 rad at a 0 = 1 a_0=1 a 0 = 1 , ε = 0.15 \varepsilon=0.15 ε = 0.15 , γ = 10 \gamma=10 γ = 10 and by up to 0.27 r a d 0.27\,\mathrm{rad} 0.27 rad at a 0 = 16 a_0=16 a 0 = 16 (measured). We have not searched the literature on g = 2 g=2 g = 2 spin rotation in focused beams, and this identification should be read as a measurement of ours rather than as a claim of novelty.
We measured ∣ r 1 ∣ / ∣ r 0 ∣ |r_1|/|r_0| ∣ r 1 ∣/∣ r 0 ∣ between 0.79 and 1.07 over the ε ≤ 0.15 \varepsilon\le0.15 ε ≤ 0.15 part of the grid, the ratio reaching 1.21 at γ = 1 \gamma=1 γ = 1 and ε = 0.30 \varepsilon=0.30 ε = 0.30 . The two coefficients are produced by the same mechanism. The focused geometry offers no selectivity for the anomaly.
Recovered carrier-envelope-phase dependence
For circular polarization and an electron on axis the CEP dependence is absent in the focused beam as well: the relative modulation over 16 values of φ 0 \varphi_0 φ 0 is 4.3 × 10 − 12 4.3 \times10^{-12} 4.3 × 1 0 − 12 for r 0 r_0 r 0 , 3.2 × 10 − 10 3.2 \times10^{-10} 3.2 × 1 0 − 10 for r 1 r_1 r 1 and 2.3 × 10 − 12 2.3 \times10^{-12} 2.3 × 1 0 − 12 for r 2 r_2 r 2 (measured). The reason is exact and identical to Sec. S3.6: for α = ( 1 , i ) / 2 \boldsymbol{\alpha}=(1,i)/\sqrt{2} α = ( 1 , i ) / 2 and a cylindrically symmetric ψ \psi ψ , rotating the whole configuration about the beam axis is the same as shifting φ 0 \varphi_0 φ 0 . Off axis the symmetry breaks, and the relative modulation of ∣ r 0 ∣ |r_0| ∣ r 0 ∣ reaches 22% for circular polarization and 35% for linear polarization at an impact parameter b = 1.5 w 0 b=1.5w_0 b = 1.5 w 0 .
For linear polarization the CEP controls the sign and magnitude of the whole rotation. The vector r \mathbf{r} r lies along the normal to the polarization plane, the mean over φ 0 \varphi_0 φ 0 is zero, and the dependence is a pure first harmonic, with A 2 / A 1 ≤ 2 × 10 − 12 A_2/A_1 \le2 \times10^{-12} A 2 / A 1 ≤ 2 × 1 0 − 12 for r 0 r_0 r 0 . The bound is 3.7 × 10 − 12 3.7 \times10^{-12} 3.7 × 1 0 − 12 for r 1 r_1 r 1 and only 1.1 × 10 − 2 1.1 \times10^{-2} 1.1 × 1 0 − 2 for r 2 r_2 r 2 , whose first harmonic is small enough that the second is not negligible beside it. The measured amplitude is A 1 ( r 0 ) ≃ 0.0209 ε 2 r a d A_1(r_0) \simeq0.0209\varepsilon^2\,\mathrm{rad} A 1 ( r 0 ) ≃ 0.0209 ε 2 rad and A 1 ( r 1 ) ≃ 0.0235 ε 2 r a d A_1(r_1) \simeq0.0235\varepsilon^2\,\mathrm{rad} A 1 ( r 1 ) ≃ 0.0235 ε 2 rad at a 0 = 1 a_0=1 a 0 = 1 , γ = 10 \gamma=10 γ = 10 , N = 1 N=1 N = 1 , with local slopes 2.46 → 2.05 2.46 \to2.05 2.46 → 2.05 .
Figure S2. Carrier-envelope-phase amplitude of the spin rotation relative to the momentum, for linear polarization at ε = 0.15 \varepsilon=0.15 ε = 0.15 (w 0 = 1.06 λ w_0=1.06\lambda w 0 = 1.06 λ ), γ = 10 \gamma=10 γ = 10 , N = 1 N=1 N = 1 , electron on axis. Filled symbols: first-harmonic amplitude of the g g g -independent coefficient r 0 r_0 r 0 . Open symbols: the same for the first-order coefficient r 1 r_1 r 1 , which is linear in a 0 a_0 a 0 up to a 0 ≃ 12 a_0 \simeq12 a 0 ≃ 12 (measured local slopes 0.99 down to 0.85) and steepens to 2.0 over the last interval. Open downward triangles on the right-hand axis are the transverse momentum kick ∣ Δ u ⊥ ∣ |\Delta u_\perp| ∣Δ u ⊥ ∣ of the same runs, showing how the ponderomotive ejection grows alongside the signal. The dependence is a pure first harmonic, with the second harmonic below 2 × 10 − 12 2 \times10^{-12} 2 × 1 0 − 12 of the first for r 0 r_0 r 0 . The annotated point near a 0 = 4 a_0=4 a 0 = 4 is a minimum of A 1 ( r 0 ) A_1(r_0) A 1 ( r 0 ) that we cannot explain. At a 0 = 24 a_0=24 a 0 = 24 the rotation swings by 1.56 rad between φ 0 \varphi_0 φ 0 and φ 0 + π \varphi_0+\pi φ 0 + π , while the anomalous part of the same signal is 3.9 × 10 − 5 r a d 3.9 \times10^{-5}\,\mathrm{rad} 3.9 × 1 0 − 5 rad .
Figure S2 shows the growth with intensity: A 1 ( r 0 ) A_1(r_0) A 1 ( r 0 ) is linear in a 0 a_0 a 0 up to a 0 ≃ 2 a_0 \simeq2 a 0 ≃ 2 , passes through a minimum near a 0 ≃ 4 a_0 \simeq4 a 0 ≃ 4 that we cannot explain, and then grows as a 0 2.6 a_0^{2.6} a 0 2.6 –a 0 3.2 a_0^{3.2} a 0 3.2 , reaching 0.781 rad at a 0 = 24 a_0=24 a 0 = 24 and ε = 0.15 \varepsilon=0.15 ε = 0.15 . The first-order amplitude A 1 ( r 1 ) A_1(r_1) A 1 ( r 1 ) is linear in a 0 a_0 a 0 only up to a 0 ≃ 12 a_0 \simeq12 a 0 ≃ 12 : the measured local slopes fall from 0.99 to 0.85 over a 0 = 1 a_0=1 a 0 = 1 –12 and then rise to 1.06 and 2.04 over the last two intervals.
The anomalous part of the signal at the record point a 0 = 24 a_0=24 a 0 = 24 is 3.9 × 10 − 5 r a d 3.9 \times10^{-5}\,\mathrm{rad} 3.9 × 1 0 − 5 rad , that is 5 × 10 − 5 5 \times10^{-5} 5 × 1 0 − 5 of the total. At a 0 = 1 a_0=1 a 0 = 1 , where the rest of this section works, the same fraction is 1.3 × 10 − 3 1.3 \times10^{-3} 1.3 × 1 0 − 3 , twenty-five times larger. Either way that part shares the CEP harmonic, the polarization signature and the γ \gamma γ dependence of the g g g -independent one, so it cannot be separated from it. The amplitude also collapses with pulse length: measured A 1 ( r 0 ) A_1(r_0) A 1 ( r 0 ) at ε = 0.15 \varepsilon=0.15 ε = 0.15 , a 0 = 1 a_0=1 a 0 = 1 , γ = 10 \gamma=10 γ = 10 is 6.29 × 10 − 2 6.29 \times10^{-2} 6.29 × 1 0 − 2 , 9.84 × 10 − 4 9.84 \times10^{-4} 9.84 × 1 0 − 4 , 2.63 × 10 − 7 2.63 \times10^{-7} 2.63 × 1 0 − 7 and 1.36 × 10 − 11 r a d 1.36 \times10^{-11}\,\mathrm{rad} 1.36 × 1 0 − 11 rad at N = 0.5 , 1 , 2 , 4 N=0.5,1,2,4 N = 0.5 , 1 , 2 , 4 , an exponential collapse whose local rate is not constant, steepening to about e − 2.2 σ e^{-2.2\sigma} e − 2.2 σ in the middle and flattening to e − 1.3 σ e^{-1.3\sigma} e − 1.3 σ over the last interval. The last two values lie below the Maxwell residuals of Sec. S5.2 and should be read as zero within the model.
Systematic error of the field model
Three variants of the field were run through identical dynamics: order 3 with c ψ = 0 c_\psi=0 c ψ = 0 , order 3 with c ψ = 1 c_\psi=1 c ψ = 1 , and order 1. The spread in ∣ r 0 ∣ |r_0| ∣ r 0 ∣ is 18% at ε = 0.20 \varepsilon=0.20 ε = 0.20 , 3.2% at ε = 0.10 \varepsilon=0.10 ε = 0.10 and 0.8% at ε = 0.05 \varepsilon=0.05 ε = 0.05 (measured at γ = 10 \gamma=10 γ = 10 , a 0 = 1 a_0=1 a 0 = 1 , N = 1 N=1 N = 1 , φ 0 = 0.3 \varphi_0=0.3 φ 0 = 0.3 ). That spread itself scales as ε 2 \varepsilon^2 ε 2 , so it enters the ε 4 \varepsilon^4 ε 4 correction and leaves the exponent alone, while the amplitudes c 0 c_0 c 0 and c 1 c_1 c 1 carry it in full. Numbers at ε ≥ 0.2 \varepsilon\ge0.2 ε ≥ 0.2 are order-of-magnitude statements. The quantitative range is ε ≤ 0.15 \varepsilon\le0.15 ε ≤ 0.15 .
On the deviation of r 2 r_2 r 2 , the quantity that Eq. (S27) and Eq. (S28) are built from, the systematic behaves quite differently and worse. All three variants give an exponent
and 1.36 × 10 − 11 r a d 1.36 \times10^{-11}\,\mathrm{rad} 1.36 × 1 0 − 11 rad at N = 0.5 , 1 , 2 , 4 N=0.5,1,2,4 N = 0.5 , 1 , 2 , 4 , an exponential collapse whose local rate is not constant, steepening to about e − 2.2 σ e^{-2.2\sigma} e − 2.2 σ in the middle and flattening to e − 1.3 σ e^{-1.3\sigma} e − 1.3 σ over the last interval. The last two values lie below the Maxwell residuals of Sec. S5.2 and should be read as zero within the model. of exactly 2, so that part is firm, but the coefficient does not converge as ε → 0 \varepsilon\to0 ε → 0 . Measured values of the deviation divided by ε 2 \varepsilon^2 ε 2 are 0.950, 1.030 and 1.051 at ε = 0.20 \varepsilon= 0.20 ε = 0.20 , 0.10 and 0.05 for c ψ = 0 c_\psi= 0 c ψ = 0 ; 3.130, 3.072 and 3.062 for c ψ = 1 c_\psi= 1 c ψ = 1 ; and 1.158, 1.074 and 1.062 at order 1. The ratio between the extremes is 3.3, 3.0 and 2.9, a factor of about three that stays put as ε \varepsilon ε shrinks. The coefficient 1.05 of Eq. (40) and the coefficient 10.7 / a 0 10.7/a_0 10.7/ a 0 of Eq. (41) are therefore quoted for one convention of the field model and are uncertain by that factor, which is why Table S8 carries the 3 \sqrt{3} 3 correction to every waist. We work at c ψ = 0 c_\psi= 0 c ψ = 0 because the correction then vanishes on axis and a 0 a_0 a 0 keeps its meaning as the peak potential at focus, and we know of no argument that makes that choice physical.
What was not computed
Exact nonparaxial fields, such as the complex-source or Sewell–Barnett solutions, were not used and the LLM model was never checked against them [60 ] , only the Gaussian envelope was run in the focused geometry, and neither cos 2 \cos^2 cos 2 nor an isodiffracting envelope, which is the one construction that would remove the ε p / σ q \varepsilon^p/\sigma^q ε p / σ q terms of § 3.5.2 , was tried at all. Radiation reaction was not integrated in the focused geometry any more than in the plane wave. Averaging over the focal volume and over the beam spectrum was not done: every number above is for a single electron at a fixed impact parameter with exact synchronization, and the measured spread over b b b within one waist is a factor of 20 in ∣ r 0 ∣ |r_0| ∣ r 0 ∣ . Ellipticity at φ 0 ≠ 0 \varphi_0 \ne0 φ 0 = 0 , a transverse offset along y y y , oblique geometries, chirp, polarization gating, and higher transverse modes were not scanned.
Regime of validity
Everything below is estimated from published scalings. We integrated no radiative process, classical or quantum. The quantum nonlinearity parameter is
χ = ℏ ω m c 2 κ a 0 , κ = k ⋅ u , (42) \chi= \frac{\hbar\omega}{mc^2}\,\kappa a_0,\qquad\kappa= k \cdot u, \tag*{(42)} χ = m c 2 ℏ ω κ a 0 , κ = k ⋅ u , (42)
with k k k normalized to unit frequency, so that κ ≃ 2 γ \kappa\simeq2\gamma κ ≃ 2 γ head-on and χ ≃ 6.07 × 10 − 6 γ a 0 \chi\simeq6.07 \times10^{-6}\gamma a_0 χ ≃ 6.07 × 1 0 − 6 γ a 0 at ℏ ω = 1.55 \hbar\omega= 1.55 ℏ ω = 1.55 eV. The symbol λ \lambda λ is the wavelength here and in the tables, never k ⋅ u k \cdot u k ⋅ u . Since κ \kappa κ is conserved in a plane wave, χ \chi χ is fixed by the initial energy and does not drift during the pulse. In the focused geometry κ \kappa κ is not conserved, and the estimate is low by roughly Δ u ⊥ / γ \Delta u_\perp/\gamma Δ u ⊥ / γ .
The three limits on the classical treatment are the classical radiated fraction R C ≃ 2 3 α a 0 χ N R_C \simeq\frac{2}{3}\alpha a_0\chi N R C ≃ 3 2 α a 0 χ N , the number of emitted photons N γ ∼ α a 0 N N_\gamma\sim\alpha a_0N N γ ∼ α a 0 N , and the radiative spin-flip probability P s f ∼ α a 0 χ 2 N P_{\rm sf} \sim\alpha a_0\chi^2N P sf ∼ α a 0 χ 2 N at χ ≪ 1 \chi\ll1 χ ≪ 1 . The first two are from Ref. [46 ] and the third from Refs. [2 , 3 ] , with the reviews [5 , 6 ] for context. The classical description dominates while χ ≪ 1 \chi\ll1 χ ≪ 1 , R C ≪ 1 R_C \ll1 R C ≪ 1 , and P s f P_{\rm sf} P sf stays below the geometric signal sin 2 ( Θ n e t / 2 ) \sin^2(\Theta_{\rm net}/2) sin 2 ( Θ net /2 ) .
Table S9 evaluates these on the plane-wave grid. Since Θ n e t ∝ a 0 2 N \Theta_{\rm net} \propto a_0^2N Θ net ∝ a 0 2 N and does not depend on γ \gamma γ at all, while all three limits tighten with γ \gamma γ , the optimum runs against intuition, and the practical form is this: keep the electron slow, pay in intensity and duration, and accept the best classical point on the grid, which is circular polarization at γ ≃ 1 \gamma\simeq1 γ ≃ 1 –10, a 0 ≃ 50 a_0 \simeq50 a 0 ≃ 50 –75 and N ≃ 16 N \simeq16 N ≃ 16 –32, where Θ n e t \Theta_{\rm net} Θ net reaches 0.1–0.4 rad with χ ≃ 2 × 10 − 4 \chi\simeq2 \times10^{-4} χ ≃ 2 × 1 0 − 4 –5 × 10 − 3 5 \times10^{-3} 5 × 1 0 − 3 and R C ≲ 1.3 × 10 − 2 R_C \lesssim1.3 \times10^{-2} R C ≲ 1.3 × 1 0 − 2 .
The working point is not free of qualifications, and none of them is numerical. At a 0 = 75 a_0 = 75 a 0 = 75 and N = 32 N = 32 N = 32 the estimated photon number is N γ ≃ 17 N_\gamma\simeq17 N γ ≃ 17 , so the spin state is not strictly pure. Each photon carries ≲ 10 − 3 \lesssim10^{-3} ≲ 1 0 − 3 of the energy and the decoherence is small, but the purity assumption has to be stated. An intensity of 1.2 × 10 22 W c m − 2 1.2 \times10^{22}\ {\rm W\ cm}^{-2} 1.2 × 1 0 22 W cm − 2 at 800 nm with 32 cycles is also a demanding pulse, and whether the required synchronization is realistic is an experimental question we have not addressed.
The third caveat is the one whose data we already have. Every number in Secs. S4 and § 3.5 is for a single electron at a fixed impact parameter, and no average over the transverse profile of the beam, over the focal volume or over the spectrum was performed. That is not a small omission in a focused geometry, because the measured spread of ∣ r 0 ∣ |r_0| ∣ r 0 ∣ over impact parameters within one waist is a factor of 20, and the CEP modulation of ∣ r 0 ∣ |r_0| ∣ r 0 ∣ off axis reaches 35% at b = 1.5 w 0 b = 1.5w_0 b = 1.5 w 0 for linear polarization (Sec. S5.6). A realized measurement therefore samples a distribution of rotations, and converting the numbers below into an expected polarization requires that average, which we did not compute.
The focused geometry is more forgiving on radiation and less forgiving on everything else. At a 0 = 24 a_0 = 24 a 0 = 24 , γ = 10 \gamma= 10 γ = 10 , N = 1 N = 1 N = 1 and ε = 0.15 \varepsilon= 0.15 ε = 0.15 the same scalings give χ = 1.5 × 10 − 3 \chi= 1.5 \times10^{-3} χ = 1.5 × 1 0 − 3 , R C = 1.7 × 10 − 4 R_C = 1.7 \times10^{-4} R C = 1.7 × 1 0 − 4 , N γ = 0.18 N_\gamma= 0.18 N γ = 0.18 and P s f = 3.7 × 10 − 7 P_{\rm sf} = 3.7 \times10^{-7} P sf = 3.7 × 1 0 − 7 (estimated); that combination is not among the deposited runs, so radiation reaction changes the result by ≲ 0.02 % \lesssim0.02\% ≲ 0.02% . The binding constraint there is ponderomotive ejection instead: Δ u ⊥ = 0.52 \Delta u_\perp= 0.52 Δ u ⊥ = 0.52 at that point, a deflection of 0.05 rad, which is tolerable, while Δ u ⊥ = 4.0 \Delta u_\perp= 4.0 Δ u ⊥ = 4.0 at a 0 = 8 a_0 = 8 a 0 = 8 and γ = 1 \gamma= 1 γ = 1 , which is not.
What breaks the theorem
Table S10 separates the mechanisms that leave Eq. (33) intact from those that destroy it. The first group was tested directly. The second was tested only for focusing and for a pulse with nonzero electric area. Radiation reaction, photon emission, a standing wave, a medium and a static field were not integrated, and their entries are estimates built from the scalings of Refs. [5 , 46 ] .
The pattern behind the table is simple. Anything that preserves both A = A ( η ) A = A(\eta) A = A ( η ) and u ( + ∞ ) = u 0 u(+\infty) = u_0 u ( + ∞ ) = u 0 leaves the holonomy alone, because those two facts are the entire input to Secs. S2.1 and S2.2. Anything that supplies a second null direction, gives k k k a mass, or makes the electron lose energy breaks one of them.
Table S9. Applicability of the classical treatment on the plane-wave grid, circular polarization, Gaussian envelope, 800 nm. Θ n e t \Theta_{\mathrm{net}} Θ net is measured; χ \chi χ , R C R_C R C , N γ N_\gamma N γ and P s f P_{\mathrm{sf}} P sf are estimated from the scalings cited in the text and were not integrated.
Of the entries in the table, radiation reaction deserves the longest remark, because it does more than shift a number. The Landau–Lifshitz force carries the classical electron time τ 0 = 2 e 2 / 3 m c 3 \tau_0 = 2e^2/3mc^3 τ 0 = 2 e 2 /3 m c 3 , and once a dimensionful constant sits in the equation of motion the argument of Sec. S3.2 collapses, since τ 0 ω \tau_0\omega τ 0 ω and γ \gamma γ can then be formed and the answer is free to depend on both. Its weight can be estimated from the radiated fraction, a relative distortion 2 R C 2R_C 2 R C of the curve traced in the rest frame against a holonomy of relative size a e a 0 a_ea_0 a e a 0 , so that Θ R R / Θ n e t ∼ 2 R C / ( a e a 0 ) \Theta_{\mathrm{RR}}/\Theta_{\mathrm{net}} \sim2R_C/(a_ea_0) Θ RR / Θ net ∼ 2 R C / ( a e a 0 ) . Evaluated on the rows of Table S9 that estimator is 2 × 10 − 3 2 \times10^{-3} 2 × 1 0 − 3 at γ = 1 \gamma= 1 γ = 1 , a 0 = 10 a_0 = 10 a 0 = 10 , N = 8 N = 8 N = 8 and about 4 at γ = 10 3 \gamma= 10^3 γ = 1 0 3 , where the correction has overtaken the quantity it corrects. It is also 6 × 10 − 2 6 \times10^{-2} 6 × 1 0 − 2 at the best classical point of that table and 0.30 one row below it, at γ = 10 \gamma= 10 γ = 10 , a 0 = 75 a_0 = 75 a 0 = 75 , N = 8 N = 8 N = 8 . A correction of six to thirty percent sits badly with a law quoted to a percent, and only an integration of the Landau–Lifshitz equation would settle the matter. We did not perform one.
The published mechanism closest to this one is the rotation of the spin toward the propagation axis by the anomalous moment in a circularly polarized field [50 ] . That effect is linear in the anomaly, and the reason is instructive here: repeated emission re-prepares the transverse spin with a phase locked to the field, so the cancellation of Sec. S2.2 never gets the closed loop it needs. Switch the emission off and the linear term disappears again.
The first order also returns with no radiation involved at all. A counterpropagating wave introduces a second null direction k 2 k_2 k 2 , and the two little algebras i s o ( 2 ) k 1 \mathfrak{iso}(2)_{k_1} iso ( 2 ) k 1 and i s o ( 2 ) k 2 \mathfrak{iso}(2)_{k_2} iso ( 2 ) k 2 together generate the whole of s o ( 1 , 3 ) \mathfrak{so}(1,3) so ( 1 , 3 ) , so the ordering in Eq. (20) survives even at g = 2 g = 2 g = 2 [52 , 53 ] . A medium does it more cheaply. The dispersion relation gives k k k a mass, k 2 = ω p 2 k^2 = \omega_p^2 k 2 = ω p 2 , the little group stops being I S O ( 2 ) ISO(2) I S O ( 2 ) , and the nilpotency on which Eq. (24) rests is gone. This is the effect Tikhomirov identified [37 ] , and its weight relative to the holonomy is ( n e / n c ) / ( a e a 0 ) (n_e/n_c)/(a_ea_0) ( n e / n c ) / ( a e a 0 ) , already of order 0.5 at n e / n c = 6 × 10 − 3 n_e/n_c = 6 \times10^{-3} n e / n c = 6 × 1 0 − 3 and a 0 = 10 a_0 = 10 a 0 = 10 . Residual gas at that density competes with the effect on equal terms.
A dressed anomaly a e ( χ ) a_e(\chi) a e ( χ ) leaves the structure intact, and it turns the result into a measurement. Since Θ n e t \Theta_{\mathrm{net}} Θ net is quadratic in a e a_e a e and no other plane-wave mechanism produces a net rotation, the transverse polarization angle measures the square of the field-dressed anomalous moment [34 , 55 ] with no g g g -independent background to subtract. One qualification is internal to the derivation. The cancellation of Sec. S2.2 takes a e a_e a e outside the integral, so a phase-dependent a e ( χ ( η ) ) a_e(\chi(\eta)) a e ( χ ( η )) would leave ∫ a e ( η ) n ^ d λ ≠ 0 \int a_e(\eta)\,\hat{\mathbf{n}}\,d\lambda\ne0 ∫ a e ( η ) n ^ d λ = 0 and produce a first-order term. At the working point χ ≃ 2 × 10 − 4 \chi\simeq2 \times10^{-4} χ ≃ 2 × 1 0 − 4 the dressing is far too weak for that to matter, but the statement is exact only for constant a e a_e a e . How the area law is modified when χ \chi χ approaches unity, where a e ( χ ) a_e(\chi) a e ( χ ) is not defined outside the locally constant field approximation, we did not analyze.
References
Table S10. Mechanisms that do and do not break Eq. (33) . Rows marked measured were computed here; rows marked estimated rest on published scalings and were not integrated. Θ p w \Theta_{\rm pw} Θ pw denotes the plane-wave holonomy − 1 2 a e 2 A -\frac{1}{2}a_e^2\mathcal{A} − 2 1 a e 2 A .