Introduction
At a regular Bragg band edge of a loss-less one-dimensional photonic crystal, counterpropagating waves form a standing wave, and both the group velocity and the energy-transport velocity vanish. Temporal periodicity produces a contrasting band edge. In a spatially uniform photonic time crystal (PTC) [1 , 2 , 3 , 4 ] , modulation conserves the wavenumber and couples positive- and negative-frequency components near parametric resonance. The resulting Floquet modes coalesce at exceptional points (EPs) bounding a momentum gap [5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 ] . Approaching an EP from the passband, the band slope diverges [15 ] , yet the Poynting-flux velocity remains bounded by the background light speed in the Maxwell–Lorentz medium considered here [16 ] . How does this singular mode structure manifest itself in the motion of stored energy within each modulation cycle?
A dispersive medium stores energy in its polarization as well as in the field, and in a Maxwell–Lorentz description the stored energy separates into the electromagnetic energy, the potential energy of the bound-charge displacement and the kinetic energy of the charge motion [17 , 18 , 19 , 20 , 21 , 22 , 23 ] . For a PTC with time-periodic permittivity, the cycle-averaged energy velocity has been shown to stay bounded and the divergent band slope has been traced to a modulation-driven geometric drift set by mismatched electric and magnetic phase connections, and the energy balance and the velocity bound have been extended to a Lorentz resonance with a modulated restoring force [16 ] . In non-Hermitian coupled-resonator waveguides, wave-packet group velocities can differ from velocities inferred from the probability flux [24 ] . In static non-Hermitian bands, Berry-connection terms correct the weight and the velocity of a driven packet [25 ] , and near an EP a divergent quantum metric produces anomalous acceleration and drift [26 ] . In a PTC, however, modulation of the restoring coefficient exchanges energy with the potential-energy sector of the polarization within every period, while damping drains its kinetic-energy sector, so the sector weights of the energy centroid change inside each cycle.
We resolve the resulting centroid motion near the Floquet EP with an exact identity that splits the centroid velocity into a Poynting-flux part bounded by the background light speed and a source part, set by how fast modulation and loss exchange energy with each sector and by how far that sector sits from the rest of the pulse. The picture is familiar from pulse amplification, where gain on the leading part of an optical pulse pulls its energy center forward by raising the weight of that part. For a narrow-band pulse on one Floquet branch, the period average of the source part carries the divergent band slope, as in the cycle-averaged picture [16 ] , and its zero-mean part, which the cycle average removes, makes the energy centroid tremble within each modulation period. This source-driven trembling recalls Zitterbewegung, the coherent oscillation of a wave-packet centroid produced by interference between positive- and negative-energy components [27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 ] . Near the gap, the dominant interfering harmonics of this packet are inherited from the positive- and negative-frequency static branches, while every stored-energy sector stays nonnegative. Modulation and loss turn the resulting sector offsets into motion. As the EP is approached, the trembling velocity grows as the inverse square root of the momentum detuning, while the weighted squared sector separations and the energynorm Petermann factor grow as its inverse. The three laws follow from one Jordan chain, yet loss modulation in phase or in antiphase with the restoring coefficient can change the trembling coefficient while leaving the band-slope and Petermann coefficients essentially unchanged.
A finite spectral width rounds the growth into a finite response. The preparation then sets the later motion: in our eigenvector gauge the trembling of an eigen-packet with a real envelope fades once modes inside the gap dominate the pulse, a quadratic spectral chirp keeps a finite gap-side response over the computed times, and a pulse launched in the fields alone excites both coalescing modes and trembles in bursts near their beat recurrence.
Energy sectors and centroid motion
We consider a one-dimensional Maxwell–Lorentz medium whose Lorentz oscillator has a restoring coefficient modulated in time. Keeping the polarization P P P and its velocity Q ≡ P ˙ Q\equiv\dot P Q ≡ P ˙ as dynamical variables resolves the material energy into a potential and a kinetic sector. With background permittivity ϵ b \epsilon_b ϵ b and permeability μ \mu μ , the source-free dynamics read
∂ z E = − μ ∂ t H , ∂ z H = − ∂ t ( ϵ b E + P ) , \partial_z E=-\mu\partial_t H,\qquad\partial_z H=-\partial_t(\epsilon_b E+P), ∂ z E = − μ ∂ t H , ∂ z H = − ∂ t ( ϵ b E + P ) ,
P ¨ + η P ˙ + κ ( t ) P = α E (1) \ddot P+\eta\dot P+\kappa(t)P=\alpha E \tag*{(1)} P ¨ + η P ˙ + κ ( t ) P = α E (1)
where κ ( t ) = κ 0 [ 1 + m cos ( Ω t ) ] \kappa(t)=\kappa_0[1+m\cos(\Omega t)] κ ( t ) = κ 0 [ 1 + m cos ( Ω t )] is the restoring coefficient of period T = 2 π / Ω T=2\pi/\Omega T = 2 π /Ω , η ≥ 0 \eta\geq0 η ≥ 0 is the damping rate and α > 0 \alpha>0 α > 0 sets the strength of the polarization response to the field. These are oscillators of fixed density with a prescribed stiffness and a local viscous loss, and in the continuum limit the same equations describe a transmission line shunt-loaded by series resonators with modulated capacitance and resistance [35 ] . With complex fields averaged over their real and imaginary parts, which for pulses with spectra confined to k > 0 k>0 k > 0 give the sector energies and centroid of the single real field at every instant [35 ] , the field–matter energy density and its exact local balance read
u t o t = 1 4 ( ϵ b ∣ E ∣ 2 + μ ∣ H ∣ 2 ) + 1 4 α [ κ ( t ) ∣ P ∣ 2 + ∣ Q ∣ 2 ] , u_{\mathrm{tot}}=\frac{1}{4}\left(\epsilon_b|E|^2+\mu|H|^2\right)+\frac{1}{4\alpha}\left[\kappa(t)|P|^2+|Q|^2\right], u tot = 4 1 ( ϵ b ∣ E ∣ 2 + μ ∣ H ∣ 2 ) + 4 α 1 [ κ ( t ) ∣ P ∣ 2 + ∣ Q ∣ 2 ] ,
∂ t u t o t + ∂ z S z − κ ˙ κ u P − 2 η u Q ≡ s (2) \partial_tu_{\mathrm{tot}}+\partial_zS_z-\frac{\dot\kappa}{\kappa}u_P-2\eta u_Q\equiv s \tag*{(2)} ∂ t u tot + ∂ z S z − κ κ ˙ u P − 2 η u Q ≡ s (2)
where the four terms of u t o t u_{\mathrm{tot}} u tot are the electric, magnetic, polarization-potential and polarization-kinetic sectors u E u_E u E , u H u_H u H , u P u_P u P and u Q u_Q u Q , all nonnegative for positive κ 0 \kappa_0 κ 0 , ϵ b \epsilon_b ϵ b , μ \mu μ and ∣ m ∣ < 1 |m|<1 ∣ m ∣ < 1 , and S z = Re ( E H ∗ ) / 2 S_z=\operatorname{Re}(EH^*)/2 S z = Re ( E H ∗ ) /2 is the Poynting flux [35 ] . Modulation adds potential energy at the rate ( κ ˙ / κ ) u P (\dot\kappa/\kappa)u_P ( κ ˙ / κ ) u P while the restoring coefficient rises and removes it while it falls, and damping drains kinetic energy at the rate 2 η u Q 2\eta u_Q 2 η u Q .
We define the total energy U = ∫ u t o t d z U=\int u_{\mathrm{tot}}\,dz U = ∫ u tot d z , its centroid Z = U − 1 ∫ z u t o t d z Z=U^{-1}\int z u_{\mathrm{tot}}\,dz Z = U − 1 ∫ z u tot d z and, for each homogeneous sector a a a with complement a ˉ \bar a a ˉ , the sector energy U a = ∫ u a d z U_a=\int u_a\,dz U a = ∫ u a d z , the energy fraction f a = U a / U f_a=U_a/U f a = U a / U and the sector centroid Z a = U a − 1 ∫ z u a d z Z_a=U_a^{-1}\int z u_a\,dz Z a = U a − 1 ∫ z u a d z . Integrating Eq. (2) over a localized pulse gives the exact centroid equation [35 ]
Z ˙ = v f l u x + v s r c , ∣ v f l u x ∣ ≤ c b = ( ϵ b μ ) − 1 / 2 , \dot Z=v_{\mathrm{flux}}+v_{\mathrm{src}},\qquad|v_{\mathrm{flux}}|\leq c_b=(\epsilon_b\mu)^{-1/2}, Z ˙ = v flux + v src , ∣ v flux ∣ ≤ c b = ( ϵ b μ ) − 1/2 ,
v s r c = κ ˙ κ f P f P ˉ Δ Z P P ˉ − 2 η f Q f Q ˉ Δ Z Q Q ˉ (3) v_{\mathrm{src}}=\frac{\dot\kappa}{\kappa}f_Pf_{\bar P}\Delta Z_{P\bar P}-2\eta f_Qf_{\bar Q}\Delta Z_{Q\bar Q} \tag*{(3)} v src = κ κ ˙ f P f P ˉ Δ Z P P ˉ − 2 η f Q f Q ˉ Δ Z Q Q ˉ (3)
where Δ Z a a ˉ = Z a − Z a ˉ \Delta Z_{a\bar a}=Z_a-Z_{\bar a} Δ Z a a ˉ = Z a − Z a ˉ , v f l u x = U − 1 ∫ S z d z v_{\mathrm{flux}}=U^{-1}\int S_z\,dz v flux = U − 1 ∫ S z d z , and the two terms of v s r c v_{\mathrm{src}} v src define the modulation channel v P v_P v P and the damping channel v Q v_Q v Q . Flux transport stays within c b c_b c b because ∣ S z ∣ ≤ c b ( u E + u H ) |S_z|\leq c_b(u_E+u_H) ∣ S z ∣ ≤ c b ( u E + u H ) at every point and the nonnegative material energy only enlarges U U U . Energy added to a displaced sector pulls the centroid toward that sector and energy removed pushes it away, in proportion to the weighted offset f a f a ˉ Δ Z a a ˉ f_af_{\bar a}\Delta Z_{a\bar a} f a f a ˉ Δ Z a a ˉ . The sign of the modulation contribution therefore depends on both the order of the P P P and P ˉ \bar P P ˉ centroids and the sign of κ ˙ \dot\kappa κ ˙ . For example, when the potential-energy sector lies ahead of its complement, softening gives a backward modulation contribution and stiffening gives a forward contribution Figure 1 (a,b). Nonzero damping gives a backward source contribution whenever the kinetic-energy sector lies ahead of its complement Figure 1 (c).
Figure 1. Reweighting an offset sector moves the energy centroid. Solid and dashed profiles show a sector ahead of its complement before and after reweighting by the indicated source term alone. Dotted lines mark the sector centroids and black markers the total centroid Z Z Z . Arrows give the contribution v a = r a f a ( Z a − Z ) = r a f a Δ Z ˉ a a ˉ v_a = r_a f_a(Z_a-Z) = r_a f_a\bar{\Delta Z}_{a\bar a} v a = r a f a ( Z a − Z ) = r a f a Δ Z ˉ a a ˉ of each channel to the centroid velocity, with r P = κ ˙ / κ r_P=\dot{\kappa}/\kappa r P = κ ˙ / κ and r Q = − 2 η r_Q=-2\eta r Q = − 2 η , in the frame z ′ = z − ∫ v f l u x d t z'=z-\int v_{\mathrm{flux}}\,dt z ′ = z − ∫ v flux d t . (a) Softening removes and (b) stiffening adds potential energy in the P ∣ P ˉ P|\bar P P ∣ P ˉ partition, and (c) damping drains kinetic energy at all phases in the Q ∣ Q ˉ Q|\bar Q Q ∣ Q ˉ partition. Top: κ ( t ) / κ 0 \kappa(t)/\kappa_0 κ ( t ) / κ 0 over one period. A sector behind Z Z Z reverses each arrow. Profiles are illustrative.
Because the inter-sector part of the energy variance is f a f a ˉ ( Δ Z a a ˉ ) 2 f_af_{\bar a}(\Delta Z_{a\bar a})^2 f a f a ˉ ( Δ Z a a ˉ ) 2 , a pulse with finite rms width σ z \sigma_z σ z also obeys
∣ v s r c ∣ ≤ ( 1 2 ∣ κ ˙ κ ∣ + η ) σ z (4) |v_{\mathrm{src}}|\leq\left(\frac{1}{2}\left|\frac{\dot\kappa}{\kappa}\right|+\eta\right)\sigma_z \tag*{(4)} ∣ v src ∣ ≤ ( 2 1 κ κ ˙ + η ) σ z (4)
For wide pulses the bound permits source velocities above c b c_b c b , and those computed below reach several times c b c_b c b . They describe how modulation and damping move the weights of stored energy [16 ] . For a single-branch pulse with nonzero spectral weight at the EP, the square-root cusp of the mode leaves the centroids finite but makes the spatial variance σ z 2 \sigma_z^2 σ z 2 diverge logarithmically. Finite bandwidth rounds the trembling-velocity growth (see below); Eq. (4) , however, also requires finite spatial variance. With a globally regular spectral completion, smoothing the mode profile over an interval ε k ≪ σ k \varepsilon_k \ll\sigma_k ε k ≪ σ k about the EP gives a launch with finite σ z \sigma_z σ z and preserves the leading fixed-window response of Eq. (12) as ε k / σ k → 0 \varepsilon_k/\sigma_k \to0 ε k / σ k → 0 [35 ] .
Trembling near the exceptional point.— We now follow a pulse built from a single Floquet branch β \beta β , with spectral width σ k \sigma_k σ k about k 0 k_0 k 0 , and write its Floquet mode as e − i ω β ( k ) t ξ β ( k , t ) e^{-i\omega_\beta(k)t}\xi_\beta(k,t) e − i ω β ( k ) t ξ β ( k , t ) , with complex quasifrequency ω β \omega_\beta ω β and periodic ξ β ( k , t + T ) = ξ β ( k , t ) \xi_\beta(k,t+T)=\xi_\beta(k,t) ξ β ( k , t + T ) = ξ β ( k , t ) whose a a a component is ξ β , a \xi_{\beta,a} ξ β , a . Each component carries its own phase, which defines a component connection,
A β , a ( k , t ) = − Im ⟨ ξ β , a ∗ ∂ k ξ β , a ⟩ ∣ ξ β , a ∣ 2 , \mathcal{A}_{\beta,a}(k,t)=-\frac{\operatorname{Im}\langle\xi_{\beta,a}^{*}\partial_k\xi_{\beta,a}\rangle}{|\xi_{\beta,a}|^2}, A β , a ( k , t ) = − ∣ ξ β , a ∣ 2 Im ⟨ ξ β , a ∗ ∂ k ξ β , a ⟩ ,
Δ Z β , a a ˉ ( t ) = Δ A β , a a ˉ ( k 0 , t ) + O ( σ k ) , (5) \Delta Z_{\beta,a\bar a}(t)=\Delta\mathcal{A}_{\beta,a\bar a}(k_0,t)+O(\sigma_k), \tag*{(5)} Δ Z β , a a ˉ ( t ) = Δ A β , a a ˉ ( k 0 , t ) + O ( σ k ) , (5)
where Δ A β , a a ˉ = A β , a − A β , a ˉ \Delta\mathcal{A}_{\beta,a\bar a}=\mathcal{A}_{\beta,a}-\mathcal{A}_{\beta,\bar a} Δ A β , a a ˉ = A β , a − A β , a ˉ and A β , a ˉ \mathcal{A}_{\beta,\bar a} A β , a ˉ is the energy-weighted connection of the complement. Just as the slope of the spectral phase sets the delay of an optical pulse, a component phase ϕ a \phi_a ϕ a shifts that component by − ∂ k ϕ a -\partial_k\phi_a − ∂ k ϕ a , so the source velocity depends on differences between the component phase gradients within one Floquet mode. The envelope position and the group motion are common to all components and cancel at leading order from this narrow-band difference for σ k ≪ k E P − k 0 \sigma_k \ll k_{\rm EP}-k_0 σ k ≪ k EP − k 0 over a fixed window; at component zeros we use the weighted moments directly [35 ] . In this limit the source weights and connection differences repeat with period T T T , and with ⟨ ⋅ ⟩ T \langle\cdot\rangle_T ⟨ ⋅ ⟩ T the average over one period we call the zero-mean part of the source velocity the trembling velocity and its centered root-mean-square (rms) over one period the trembling-velocity amplitude,
v t r = v s r c − ⟨ v s r c ⟩ T , A t r = ⟨ v t r 2 ⟩ T 1 / 2 . (6) v_{\rm tr}=v_{\rm src}-\langle v_{\rm src}\rangle_T,\qquad A_{\rm tr}=\langle v_{\rm tr}^2\rangle_T^{1/2}. \tag*{(6)} v tr = v src − ⟨ v src ⟩ T , A tr = ⟨ v tr 2 ⟩ T 1/2 . (6)
At fixed passband detuning and over a fixed window, the sector centroids of the pulse, weighted by their energy fractions, advance on average at the band slope. Averaging Eq. (3) over one period then gives, at leading order in σ k \sigma_k σ k ,
⟨ v f l u x ⟩ T + ⟨ v s r c ⟩ T = ∂ k Re ω β (7) \langle v_{\rm flux}\rangle_T+\langle v_{\rm src}\rangle_T=\partial_k\operatorname{Re}\omega_\beta \tag*{(7)} ⟨ v flux ⟩ T + ⟨ v src ⟩ T = ∂ k Re ω β (7)
[35 ] . Because ∣ v f l u x ∣ ≤ c b |v_{\rm flux}|\le c_b ∣ v flux ∣ ≤ c b at every instant, Eq. (7) implies that the divergence of the band slope near the EP resides in the mean source velocity. Near the lower gap edge the two resonant Floquet modes coalesce at k E P k_{\rm EP} k EP [Fig. 2(a)] [36 , 37 , 38 ] . There the one-period propagation matrix of the full field–matter system has a single eigenvector ξ 0 \xi_0 ξ 0 for a doubly degenerate Floquet multiplier and a generalized partner ξ 1 \xi_1 ξ 1 , which form a Jordan chain. Removing the Floquet factor, and from the partner also its term linear in time, gives the periodic vectors ξ 0 p e r ( t ) \xi_0^{\rm per}(t) ξ 0 per ( t ) and ξ 1 p e r ( t ) \xi_1^{\rm per}(t) ξ 1 per ( t ) [35 ] . With d = k E P − k > 0 d=k_{\rm EP}-k>0 d = k EP − k > 0 on the passband side, the chain yields the square-root expansion
ξ β , a ( k , t ) = ξ 0 , a p e r ( t ) + β d χ a ( t ) + O ( d ) , \xi_{\beta,a}(k,t)=\xi_{0,a}^{\rm per}(t)+\beta\sqrt{d}\,\chi_a(t)+O(d), ξ β , a ( k , t ) = ξ 0 , a per ( t ) + β d χ a ( t ) + O ( d ) ,
A β , a ( k , t ) = β C a ( t ) d + O ( 1 ) , (8) \mathcal{A}_{\beta,a}(k,t)=\beta\frac{C_a(t)}{\sqrt{d}}+O(1), \tag*{(8)} A β , a ( k , t ) = β d C a ( t ) + O ( 1 ) , (8)
for β = ± \beta=\pm β = ± , where χ \chi χ follows from ξ 1 p e r \xi_1^{\rm per} ξ 1 per and the second line holds where ξ 0 , a p e r ( t ) ≠ 0 \xi_{0,a}^{\rm per}(t)\ne0 ξ 0 , a per ( t ) = 0 [35 ] . The mode thus has a finite limit at the EP while its wavenumber derivative diverges, and the sector mismatch inherits that divergence wherever C a ≠ C a ˉ C_a\ne C_{\bar a} C a = C a ˉ . Hence the leading source velocity is odd in β \beta β , and when its centered waveform does not vanish identically the trembling-velocity amplitude on either branch obeys
A t r = C s d + O ( 1 ) , d → 0 + , (9) A_{\rm tr}=\frac{C_s}{\sqrt{d}}+O(1),\qquad d\to0^+, \tag*{(9)} A tr = d C s + O ( 1 ) , d → 0 + , (9)
where C s C_s C s , the centered rms of the leading source waveform, follows from the propagated chain, the energy weights and the two exchange rates. The source channels trace a two-humped cycle [Fig. 2(b)], and their rms amplitudes follow d − 1 / 2 d^{-1/2} d − 1/2 over three decades, with C s = 0.0822 C_s=0.0822 C s = 0.0822 for the total source [Fig. 2(c)]. At d = 10 − 4 d=10^{-4} d = 1 0 − 4 the trembling-velocity amplitude reaches 8.1 c b 8.1c_b 8.1 c b , whereas the flux velocity stays below 0.4 c b 0.4c_b 0.4 c b at every plotted detuning. Of the band slope of 6.3 c b 6.3c_b 6.3 c b there, the mean flux velocity carries 0.10 c b 0.10c_b 0.10 c b and the mean source velocity 6.2 c b 6.2c_b 6.2 c b . In position, the detrended centroid of the packet at d 0 = 10 − 4 d_0=10^{-4} d 0 = 1 0 − 4 , where d 0 = k E P − k 0 d_0=k_{\rm EP}-k_0 d 0 = k EP − k 0 is the detuning of the packet center k 0 k_0 k 0 , swings by about 17 c b / Ω 17c_b/\Omega 17 c b /Ω peak to peak, whereas the flux contribution spans less than 0.06 c b / Ω 0.06c_b/\Omega 0.06 c b /Ω [Fig. 2(d)]. Because it integrates the same leading waveform, the displacement amplitude also grows as d − 1 / 2 d^{-1/2} d − 1/2 [35 ] . Field measurements can reconstruct the field-energy centroid Z E M Z_{\rm EM} Z EM , weighted by u E + u H u_E+u_H u E + u H , whose centered rms displacement is about eight times that of Z Z Z for the packets of Figure 2 (d), and its detrended, centered displacements at the two detunings nearly coincide after multiplication by d 0 \sqrt{d_0} d 0 [35 ] . In the transmission-line realization, time-resolved line voltages and currents together with resonator charge and current determine all four sector energies and first moments and thereby reconstruct Z ( t ) Z(t) Z ( t ) and v s r c ( t ) v_{\rm src}(t) v src ( t ) [35 ] . The damping contribution, carried by the kinetic-energy sector, exceeds that of modulation in centered rms at m = 0.25 m=0.25 m = 0.25 and η = 0.5 \eta=0.5 η = 0.5 [35 ] .
Figure 2. Single-branch trembling grows near the exceptional point while the flux velocity stays bounded. (a) Real quasifrequencies of the two resonant Floquet branches near the lower gap edge; the dashed line marks k E P k_{\mathrm{EP}} k EP , and d ≡ k E P − k d \equiv k_{\mathrm{EP}}-k d ≡ k EP − k . (b) Modulation channel v P v_P v P , damping channel v Q v_Q v Q , their sum v s r c v_{\mathrm{src}} v src and the trembling velocity v t r v_{\mathrm{tr}} v tr on the β = − \beta=- β = − branch at d = 10 − 4 d=10^{-4} d = 1 0 − 4 ; the dotted line marks ⟨ v s r c ⟩ T \langle v_{\mathrm{src}}\rangle_T ⟨ v src ⟩ T . (c) Centered rms amplitudes of v P v_P v P , v Q v_Q v Q and v t r v_{\mathrm{tr}} v tr , normalized by A t r A_{\mathrm{tr}} A tr at d = 10 − 3 d=10^{-3} d = 1 0 − 3 ; dashed guide, d − 1 / 2 d^{-1/2} d − 1/2 . Inset: max t ∣ v f l u x ∣ / c b \max_t|v_{\mathrm{flux}}|/c_b max t ∣ v flux ∣/ c b . (d) Mean-centered centroid displacement δ Z \delta Z δ Z after subtraction of the uniform drift v ˉ t \bar{v}t v ˉ t , with v ˉ \bar{v} v ˉ and the centering taken over 1 ≤ t / T ≤ 4 1\leq t/T\leq4 1 ≤ t / T ≤ 4 , for single-branch packets of spectral width σ k = d 0 / 20 \sigma_k=d_0/20 σ k = d 0 /20 , where d 0 = k E P − k 0 d_0=k_{\mathrm{EP}}-k_0 d 0 = k EP − k 0 is the detuning of the packet center k 0 k_0 k 0 ; dashed, flux contribution at d 0 = 10 − 4 d_0=10^{-4} d 0 = 1 0 − 4 . (e) Energy-norm Petermann factor K β ( u ) K_\beta^{(u)} K β ( u ) and centered rms A s e p A_{\mathrm{sep}} A sep of S s e p S_{\mathrm{sep}} S sep , each normalized at d = 10 − 3 d=10^{-3} d = 1 0 − 3 ; dashed guide, d − 1 d^{-1} d − 1 . A t r A_{\mathrm{tr}} A tr is the centered rms of the source velocity v s r c v_{\mathrm{src}} v src over one period. Parameters: ϵ b = μ = Ω = 1 \epsilon_b=\mu=\Omega=1 ϵ b = μ = Ω = 1 , α = 0.6 \alpha=0.6 α = 0.6 , κ 0 = 2 \kappa_0=2 κ 0 = 2 , m = 0.25 m=0.25 m = 0.25 , η = 0.08 \eta=0.08 η = 0.08 , k E P = 0.5743977 k_{\mathrm{EP}}=0.5743977 k EP = 0.5743977 . Lengths are in c b / Ω c_b/\Omega c b /Ω , wavenumbers in Ω / c b \Omega/c_b Ω/ c b and velocities in c b c_b c b .
Nearly parallel modes.— Near the EP the Jordan structure also makes the two coalescing modes nearly parallel, which the Petermann factor quantifies. We evaluate it with the field–matter energy metric and its dual,
K β ( u ) = ⟨ R β ∣ G u ∣ R β ⟩ ⟨ L β ∣ G u − 1 ∣ L β ⟩ ∣ ⟨ L β ∣ R β ⟩ ∣ 2 = K 0 d + O ( d − 1 / 2 ) , (10) K_\beta^{(u)}=\frac{\langle R_\beta|G_u|R_\beta\rangle\langle L_\beta|G_u^{-1}|L_\beta\rangle}{|\langle L_\beta|R_\beta\rangle|^2}
=\frac{K_0}{d}+O\left(d^{-1/2}\right), \tag*{(10)} K β ( u ) = ∣ ⟨ L β ∣ R β ⟩ ∣ 2 ⟨ R β ∣ G u ∣ R β ⟩ ⟨ L β ∣ G u − 1 ∣ L β ⟩ = d K 0 + O ( d − 1/2 ) , (10)
where ∣ R β ⟩ |R_\beta\rangle ∣ R β ⟩ and ⟨ L β ∣ \langle L_\beta| ⟨ L β ∣ are the right and left Floquet eigenvectors, G u G_u G u is the period-averaged energy Gramian, G u − 1 G_u^{-1} G u − 1 defines the dual norm of the left covector and K 0 > 0 K_0>0 K 0 > 0 [35 ] . Introduced for the excess noise of gain-guided lasers [39 ] , the factor quantifies mode nonorthogonality [40 , 41 ] and diverges at EPs [42 , 43 ] , including the momentum-gap edges of PTCs [44 ] . The same Jordan chain controls the weighted squared sector separations. We add the inter-sector variances of the two overlapping source-selected partitions, S s e p = f P f P ˉ ( Δ Z P P ˉ ) 2 + f Q f Q ˉ ( Δ Z Q Q ˉ ) 2 S_{\mathrm{sep}}=f_Pf_{\bar{P}}(\Delta Z_{P\bar{P}})^2+f_Qf_{\bar{Q}}(\Delta Z_{Q\bar{Q}})^2 S sep = f P f P ˉ ( Δ Z P P ˉ ) 2 + f Q f Q ˉ ( Δ Z Q Q ˉ ) 2 , and call A s e p A_{\mathrm{sep}} A sep its centered rms over one period. Taking the narrow-band limit at fixed d d d and then approaching the EP gives A s e p = C s e p / d + o ( d − 1 ) A_{\mathrm{sep}}=C_{\mathrm{sep}}/d+o(d^{-1}) A sep = C sep / d + o ( d − 1 ) whenever the leading centered coefficient C s e p C_{\mathrm{sep}} C sep is nonzero, with a leading term
even in β \beta β [35 ] Figure 2 (e). Under the stated nonzero-coefficient conditions, the common Jordan chain gives the leading local passband laws
A t r ∝ [ K β ( u ) ] 1 / 2 ∝ d − 1 / 2 , (11) A_{\mathrm{tr}}\propto\left[K_\beta^{(u)}\right]^{1/2}\propto d^{-1/2}, \tag*{(11)} A tr ∝ [ K β ( u ) ] 1/2 ∝ d − 1/2 , (11)
A s e p ∝ K β ( u ) ∝ d − 1 . A_{\mathrm{sep}}\propto K_\beta^{(u)}\propto d^{-1}. A sep ∝ K β ( u ) ∝ d − 1 .
The coefficient C s C_s C s contains the energy-exchange rates, whereas C s e p C_{\mathrm{sep}} C sep is set by the sector fractions and separations. The Petermann factor alone therefore fixes neither.
A common control can change both the restoring coefficient and the damping, and their linear responses are in phase or in antiphase when the relative response lag is negligible. We take η ( t ) = η 0 [ 1 + δ cos ( Ω t + φ ) ] \eta(t)=\eta_0[1+\delta\cos(\Omega t+\varphi)] η ( t ) = η 0 [ 1 + δ cos ( Ω t + φ )] , for which each loss profile has its own EP, and compare C s C_s C s with K 0 K_0 K 0 and with the band-slope coefficient C g = lim d → 0 + d ∣ ∂ k Re ω β ∣ C_g=\lim_{d\to0+}\sqrt{d}\,|\partial_k\operatorname{Re}\omega_\beta| C g = lim d → 0 + d ∣ ∂ k Re ω β ∣ Figure 3 . For the medium of Figure 2 , in-phase (φ = 0 \varphi=0 φ = 0 ) and antiphase (φ = π \varphi=\pi φ = π ) loss modulation change C s C_s C s while C g C_g C g and K 0 K_0 K 0 remain close to their constant-loss values [Fig. 3(a)]. At stronger damping, antiphase loss raises C s C_s C s by about a quarter with little change in the other two coefficients, while at φ = 3 π / 2 \varphi=3\pi/2 φ = 3 π /2 C g C_g C g and K 0 \mathcal{K}_0 K 0 decrease and C s C_s C s remains close to its constant-loss value [Fig. 3(b)]. At η 0 = 0.08 \eta_0=0.08 η 0 = 0.08 , for single-branch packets with d 0 = 10 − 4 d_0=10^{-4} d 0 = 1 0 − 4 , full-depth antiphase loss raises the centered rms displacement of Z Z Z by about a tenth and in-phase loss lowers it by about a twentieth, while that of Z E M Z_{\rm EM} Z EM barely changes [35 ] . The ratios C s / C g C_s/C_g C s / C g and C s / K 0 C_s/\sqrt{\mathcal{K}_0} C s / K 0 therefore depend on the loss profile.
Figure 3. Loss modulation separates the trembling coefficient from the band-slope and Petermann coefficients. Leading coefficients C s C_s C s , C g C_g C g and K 0 \mathcal{K}_0 K 0 at the EP of the loss profile η ( t ) = η 0 [ 1 + δ cos ( Ω t + φ ) ] \eta(t)=\eta_0[1+\delta\cos(\Omega t+\varphi)] η ( t ) = η 0 [ 1 + δ cos ( Ω t + φ )] , divided by their values for constant loss η 0 \eta_0 η 0 . (a) Depth dependence for η 0 = 0.08 \eta_0=0.08 η 0 = 0.08 with loss in phase (φ = 0 \varphi=0 φ = 0 , open symbols) and in antiphase (φ = π \varphi=\pi φ = π , filled symbols) with the restoring coefficient. (b) Phase dependence for η 0 = 0.3 \eta_0=0.3 η 0 = 0.3 and δ = 1 \delta=1 δ = 1 ; dotted lines, φ = π \varphi=\pi φ = π and 3 π / 2 3\pi/2 3 π /2 . Shaded bands mark ± 1 % \pm1\% ± 1% about the constant-loss value. Symbols mark computed points, joined by straight segments. Other parameters as in Figure 2 .
Finite bandwidth, gap migration and fade
We consider a Gaussian pulse of rms spectral width σ k \sigma_k σ k centered at k 0 k_0 k 0 on the passband side, with λ = d 0 / σ k \lambda=d_0/\sigma_k λ = d 0 / σ k . For a regular launch with a real envelope in our eigenvector gauge and over a fixed window, the leading passband source waveform is common to all wavenumbers near the EP,
and averaging it over the spectrum gives, at fixed λ \lambda λ ,
A t r w p σ k = C s F ( λ ) + o ( 1 ) , σ k → 0 , A_{\rm tr}^{\rm wp}\sqrt{\sigma_k}=C_sF(\lambda)+o(1),\qquad\sigma_k\to0, A tr wp σ k = C s F ( λ ) + o ( 1 ) , σ k → 0 ,
F ( λ ) = ∫ 0 ∞ e − ( u − λ ) 2 / 2 2 π u d u (12) F(\lambda)=\int_0^\infty\frac{e^{-(u-\lambda)^2/2}}{\sqrt{2\pi u}}\,\mathrm{d}u \tag*{(12)} F ( λ ) = ∫ 0 ∞ 2 π u e − ( u − λ ) 2 /2 d u (12)
with F ( 0 ) = 0.860 F(0)=0.860 F ( 0 ) = 0.860 and F → λ − 1 / 2 F\to\lambda^{-1/2} F → λ − 1/2 for λ ≫ 1 \lambda\gg1 λ ≫ 1 [35 ] . A finite bandwidth thus rounds the local divergence into a response of order C s σ k − 1 / 2 C_s\sigma_k^{-1/2} C s σ k − 1/2 when the pulse overlaps the EP. We propagate the pulses without a narrow-band expansion and launch each eigenvector with its electric component real and positive, a choice that fixes the phase across the EP [35 ] . For σ k = 10 − 5 \sigma_k=10^{-5} σ k = 1 0 − 5 and 10 − 4 10^{-4} 1 0 − 4 the scaled amplitudes A t r w p σ k A_{\mathrm{tr}}^{\mathrm{wp}}\sqrt{\sigma_k} A tr wp σ k approach C s F ( λ ) C_sF(\lambda) C s F ( λ ) and depart from it progressively as the pulse samples wavenumbers beyond the square-root regime Figure 4 (a), whereas for the broader pulse with σ k = 10 − 3 \sigma_k=10^{-3} σ k = 1 0 − 3 the response already departs from this form at d 0 = 0 d_0=0 d 0 = 0 [35 ] .
Figure 4. Finite bandwidth rounds the trembling, and the spectrum migrates into the gap. (a) A t r w p σ k A_{\rm tr}^{\rm wp}\sqrt{\sigma_k} A tr wp σ k over the first period for single-branch packets with σ k = 10 − 5 \sigma_k=10^{-5} σ k = 1 0 − 5 , 10 − 4 10^{-4} 1 0 − 4 and 10 − 3 10^{-3} 1 0 − 3 , against λ = d 0 / σ k \lambda=d_0/\sigma_k λ = d 0 / σ k . Gray, Gaussian average C s F ( λ ) C_sF(\lambda) C s F ( λ ) ; dashed, λ − 1 / 2 \lambda^{-1/2} λ − 1/2 . Insets: spectra overlapping and resolved from the EP. (b) Normalized spectral energy density log 10 ( σ k W ) \log_{10}(\sigma_kW) log 10 ( σ k W ) , with ∫ W d k = 1 \int W\,\mathrm{d}k=1 ∫ W d k = 1 , of the propagated packet with σ k = 10 − 4 \sigma_k=10^{-4} σ k = 1 0 − 4 and λ = 2 \lambda=2 λ = 2 , against τ = t ρ σ k \tau=t\sqrt{\rho}\sigma_k τ = t ρ σ k . Circles, spectral maxima; dashed, saddle of Eq. (13) ; dotted, k = k E P k=k_{\rm EP} k = k EP . (c) Scaled trembling amplitude A t r w p σ k A_{\rm tr}^{\rm wp}\sqrt{\sigma_k} A tr wp σ k , with A t r w p A_{\rm tr}^{\rm wp} A tr wp evaluated as the centered rms of v s r c v_{\rm src} v src over each modulation period, and energy U / U 0 U/U_0 U / U 0 relative to the launch energy, for the same packet; the dotted line marks the crossing of the spectral maximum into the gap. Parameters as in Figure 2 .
On the gap side, h = k − k E P > 0 h=k-k_{\mathrm{EP}}>0 h = k − k EP > 0 , the larger Floquet growth rate Γ ( k ) = Im ω ( k ) \Gamma(k)=\operatorname{Im}\omega(k) Γ ( k ) = Im ω ( k ) rises as Γ ( k ) − Γ E P ≃ ρ h \Gamma(k)-\Gamma_{\mathrm{EP}}\simeq\sqrt{\rho h} Γ ( k ) − Γ EP ≃ ρ h , with ρ = 0.0158 \rho=0.0158 ρ = 0.0158 for this medium, and with y = h / σ k y=h/\sigma_k y = h / σ k this relative gain competes with the Gaussian penalty away from k 0 k_0 k 0 , so the spectral weight migrates. Once the gap-side contribution dominates, its spectral maximum follows the saddle
y s ( y s + λ ) = τ , τ = t ρ σ k , (13) \sqrt{y_s}(y_s+\lambda)=\tau,\qquad\tau=t\sqrt{\rho\sigma_k}, \tag*{(13)} y s ( y s + λ ) = τ , τ = t ρ σ k , (13)
while the local growth expansion holds, with y s ≃ τ 2 / 3 y_s\simeq\tau^{2/3} y s ≃ τ 2/3 once y s ≫ ∣ λ ∣ y_s\gg|\lambda| y s ≫ ∣ λ ∣ .
For λ = 2 \lambda=2 λ = 2 the maximum of the propagated spectrum stays on the passband until τ ≈ 2.2 \tau\approx2.2 τ ≈ 2.2 , jumps into the gap and then follows Eq. (13) [Fig. 4(b)]. Because Γ E P < 0 \Gamma_{\mathrm{EP}}<0 Γ EP < 0 , the packet first loses energy while its normalized spectrum shifts into the gap, passes through an energy minimum near the stage when the saddle reaches the leading net-growth scale h ≃ Γ E P 2 / ρ ≃ 4 × 10 − 4 h\simeq\Gamma_{\mathrm{EP}}^2/\rho\simeq4\times10^{-4} h ≃ Γ EP 2 / ρ ≃ 4 × 1 0 − 4 [Fig. 4(c)], and then recovers [35 ] .
The trembling fades as the gap side takes over [Fig. 4(c)] because of a symmetry of the gap modes. In the frame ( E , − i H , P , Q ) (E,-iH,P,Q) ( E , − i H , P , Q ) , the equation of motion are real. On a smooth gap branch with a nondegenerate negative real Floquet multiplier, our launch convention therefore makes the propagated state real in that frame, albeit with a phase that changes by π \pi π from period to period. For a field-only launch, this implies that the centroid displacement vanishes except during the brief intervals when the state passes through the exceptional point. so the gap-side contributions to every first-moment integral vanish [35 ] . Hence the trembling decays as the passband contributions to the normalized moments shrink. A spectral phase varying nonlinearly with k k k can shift the sector centroids unequally and lift this gap-side null. A quadratic chirp keeps a finite gap-side trembling over the computed window. The real-envelope response instead falls by nearly eight orders of magnitude by τ = 10 \tau= 10 τ = 10 , and the two smoothed finite-variance launches follow this decrease through τ = 10 \tau= 10 τ = 10 [35 ] .
Launch in the fields alone. —An eigen-packet occupies one Floquet branch, whereas a forward field pulse launched with the polarization at rest, x ( k , 0 ) = ψ ( k ) ( 1 , 1 , 0 , 0 ) T x(k,0)=\psi(k)(1,1,0,0)^{\mathrm{T}} x ( k , 0 ) = ψ ( k ) ( 1 , 1 , 0 , 0 ) T for ϵ b = μ = 1 \epsilon_b=\mu=1 ϵ b = μ = 1 , projects onto both coalescing modes. Because the modes are nearly parallel, their branch-diagonal energies greatly exceed the launch energy and are balanced by destructive interference, and their singular source contributions cancel in the same way, so the first-period trembling remains finite as the EP is approached. The two branches split as Re ω ± ≃ Re ω E P ± 2 C g d \operatorname{Re}\omega_{\pm}\simeq\operatorname{Re}\omega_{\mathrm{EP}}\pm2C_g\sqrt{d} Re ω ± ≃ Re ω EP ± 2 C g d , giving the beat time t b ≃ π / ( 2 C g d ) t_b\simeq\pi/(2C_g\sqrt{d}) t b ≃ π / ( 2 C g d ) . Near this recurrence the energy reaches an interference minimum while the two branch components have separated by about π / d \pi/d π / d , their first-moment contributions no longer cancel, and the trembling bursts Figure 5 (a). In the monochromatic limit of this preparation the computed burst amplitudes follow C b d − 3 / 2 C_b d^{-3/2} C b d − 3/2 with C b ≃ 0.0031 C_b\simeq0.0031 C b ≃ 0.0031 over 10 − 6 ≤ d ≤ 10 − 3 10^{-6}\leq d\leq10^{-3} 1 0 − 6 ≤ d ≤ 1 0 − 3 and exceed the single-branch amplitude Figure 5 (b). A finite spectral width spreads the relative beat phase across the pulse, the large branch-diagonal energies amplify its effect, and the burst approaches its monochromatic value only while the resulting energy excess at the recurrence minimum, Δ R ≃ 0.031 σ k 2 / d 0 3 \Delta R\simeq0.031\sigma_k^2/d_0^3 Δ R ≃ 0.031 σ k 2 / d 0 3 , stays small, a condition that fails at fixed σ k / d 0 \sigma_k/d_0 σ k / d 0 as the EP is approached Figure 5 (b), inset [35 ] . For d 0 = 10 − 3 d_0=10^{-3} d 0 = 1 0 − 3 and σ k = d 0 / 20 \sigma_k=d_0/20 σ k = d 0 /20 the burst peaks about 123 periods after launch, when the pulse retains about 2% of its launch energy Figure 5 (c), and the detrended centroid swings by about 150 c b / Ω 150c_b/\Omega 150 c b /Ω peak to peak, a small fraction of the initial rms width 1 / ( 2 σ k ) 1/(2\sigma_k) 1/ ( 2 σ k ) Figure 5 (d).
Figure 5. A field-only launch trembles in bursts near the beat recurrence. (a) Period-resolved A t r A_{\mathrm{tr}} A tr , the centered rms of v s r c v_{\mathrm{src}} v src over one modulation period, sampled every third period, for a forward field pulse with the polarization at rest (field only) and for a single-branch packet with the same envelope, d 0 = 10 − 3 d_0=10^{-3} d 0 = 1 0 − 3 and σ k = d 0 / 20 \sigma_k=d_0/20 σ k = d 0 /20 ; dotted line, t b / T t_b/T t b / T . (b) Burst amplitude A b A_b A b , the largest one-period centered rms of v s r c v_{\mathrm{src}} v src near the first recurrence, for the monochromatic launch (circles) and packets with σ k = d 0 / 20 \sigma_k=d_0/20 σ k = d 0 /20 (squares); solid line, C b d − 3 / 2 C_b d^{-3/2} C b d − 3/2 ; dashed line, the single-branch law C s d − 1 / 2 C_s d^{-1/2} C s d − 1/2 . Finite packets approach this response when the recurrence-energy excess Δ R ≃ 0.031 σ k 2 / d 0 3 \Delta R \simeq0.031\,\sigma_k^2/d_0^3 Δ R ≃ 0.031 σ k 2 / d 0 3 is small. Inset, Δ R \Delta R Δ R , the excess over one of the launch-normalized energy ratio of the two packets at the recurrence minimum (squares), and the estimate (dashed); gray line, Δ R = 1 \Delta R=1 Δ R = 1 . (c) Energy U / U 0 U/U_0 U / U 0 relative to the launch energy for the two packets of (a). (d) Centroid displacement δ Z \delta Z δ Z of the field-only packet, with a least-squares linear fit over each period removed, within the burst period n = 123 n=123 n = 123 and within n = 1 n=1 n = 1 , the latter multiplied by 100. Parameters as in Figure 2 .
Discussion. —Near a PTC gap edge, growing offsets between energy sectors amplify the source contribution to the energy-centroid velocity while the Poynting-flux velocity stays bounded. In the local single-branch passband limit the source contribution has a period average that carries the divergent part of the band slope and a zero-mean part that produces intracycle trembling. Cycle averaging therefore removes this trembling velocity. Neither the band slope nor the Petermann factor fixes its amplitude: we find that loss modulated in phase or in antiphase with the restoring coefficient can change the trembling coefficient while both leading coefficients stay essentially fixed, and that the preparation decides whether the enhanced response appears at once, fades as gap modes dominate, or recurs in bursts near the beat time. Beyond the Maxwell–Lorentz medium, the centroid identity applies to modulated linear wave systems whose stored energy has a positive, wavenumber-independent metric in a fixed physical basis [35 ] .
Acknowledgments
This work is supported by the National Research Foundation of Korea (NRF) through the government of Korea (RS-2022-NR070636) and the Samsung Science and Technology Foundation (SSTF-BA240202). K.L. also acknowledges support from the KAIST Jang Young Sil Fellow Program.
References
Supplemental Material: Energy-Centroid Trembling near Exceptional Points in Photonic Time Crystals
Kyungmin Lee; Younsung Kim; Minwook Kyung; Seungmin Yeon; Bumki Min
Model and conventions
We take E = E x ^ \mathbf{E}=E\hat{\mathbf{x}} E = E x ^ , H = H y ^ \mathbf{H}=H\hat{\mathbf{y}} H = H y ^ , P = P x ^ \mathbf{P}=P\hat{\mathbf{x}} P = P x ^ , D = ϵ b E + P D=\epsilon_b E+P D = ϵ b E + P and B = μ H B=\mu H B = μ H with ϵ b , μ > 0 \epsilon_b,\mu>0 ϵ b , μ > 0 , and the driven Lorentz oscillator
P ¨ + η P ˙ + κ ( t ) P = α E , α > 0 , η ≥ 0 , κ ( t ) = κ 0 [ 1 + m cos Ω t ] > 0 , Q ≡ P ˙ . (14) \ddot{P}+\eta\dot{P}+\kappa(t)P=\alpha E,\qquad\alpha>0,\quad\eta\geq0,\quad\kappa(t)=\kappa_0[1+m\cos\Omega t]>0,\quad Q\equiv\dot{P}. \tag*{(14)} P ¨ + η P ˙ + κ ( t ) P = α E , α > 0 , η ≥ 0 , κ ( t ) = κ 0 [ 1 + m cos Ω t ] > 0 , Q ≡ P ˙ . (14)
The model describes oscillators of fixed density with a prescribed stiffness κ ( t ) \kappa(t) κ ( t ) and a local viscous loss η ( t ) \eta(t) η ( t ) . For bound charges q q q of mass M M M and density N N N in the potential k ( t ) x 2 / 2 k(t)x^2/2 k ( t ) x 2 /2 ,
P = N q x , α = N q 2 M , κ ( t ) = k ( t ) M , (15) P=Nqx,\qquad\alpha=\frac{Nq^2}{M},\qquad\kappa(t)=\frac{k(t)}{M}, \tag*{(15)} P = N q x , α = M N q 2 , κ ( t ) = M k ( t ) , (15)
and u P + u Q u_P+u_Q u P + u Q represents the instantaneous mechanical energy of the oscillators in the complex-field normalization used here. The centroid Z Z Z therefore locates the energy held by the fields and oscillators, while the pump work and the dissipated heat enter through the source terms of Eq. (2) .
In the continuum limit the same equations describe a transmission line with inductance L ′ L' L ′ and capacitance C ′ C' C ′ per length, shunt-loaded at density ρ r \rho_r ρ r by series resonators with inductance L r L_r L r , resistance R ( t ) R(t) R ( t ) and capacitance C r ( t ) C_r(t) C r ( t ) . With the line voltage and current in place of E E E and H H H and P = ρ r q r P=\rho_rq_r P = ρ r q r for the resonator charge q r q_r q r ,
ϵ b = C ′ , μ = L ′ , α = ρ r L r , κ ( t ) = 1 L r C r ( t ) , η ( t ) = R ( t ) L r , (16) \epsilon_b=C',\qquad\mu=L',\qquad\alpha=\frac{\rho_r}{L_r},\qquad\kappa(t)=\frac{1}{L_rC_r(t)},\qquad\eta(t)=\frac{R(t)}{L_r}, \tag*{(16)} ϵ b = C ′ , μ = L ′ , α = L r ρ r , κ ( t ) = L r C r ( t ) 1 , η ( t ) = L r R ( t ) , (16)
and u t o t u_{\rm tot} u tot becomes the line and resonator energy. A conductance in parallel with C r ( t ) C_r(t) C r ( t ) gives a different equation, so η ( t ) \eta(t) η ( t ) describes series loss. For a varactor, this correspondence applies to linear signals about a prescribed pump trajectory when bias-network loading, pump feedthrough, device memory, parasitic and lattice corrections, and signal-dependent series-resistance terms proportional to the pump current are negligible. Cell-resolved measurements give the sector energies directly. For a cell j j j of length Δ z \Delta z Δ z at position z j z_j z j ,
U E , j = C ′ Δ z 2 V j 2 , U H , j = L ′ Δ z 2 I j 2 , U P , j = q j 2 2 C r ( t ) , U Q , j = L r 2 i r , j 2 , (17) U_{E,j}=\frac{C'\Delta z}{2}V_j^2,\qquad U_{H,j}=\frac{L'\Delta z}{2}I_j^2,\qquad U_{P,j}=\frac{q_j^2}{2C_r(t)},\qquad U_{Q,j}=\frac{L_r}{2}i_{r,j}^2, \tag*{(17)} U E , j = 2 C ′ Δ z V j 2 , U H , j = 2 L ′ Δ z I j 2 , U P , j = 2 C r ( t ) q j 2 , U Q , j = 2 L r i r , j 2 , (17)
with q j = C r ( t ) v C , j q_j=C_r(t)v_{C,j} q j = C r ( t ) v C , j from the capacitor voltage and i r , j = q ˙ j i_{r,j}=\dot{q}_j i r , j = q ˙ j the resonator current, summed over the resonators of a cell. The sums U a = ∑ j U a , j U_a=\sum_j U_{a,j} U a = ∑ j U a , j and J a = ∑ j z j U a , j J_a=\sum_j z_jU_{a,j} J a = ∑ j z j U a , j give Z Z Z and the sector centroids, and Eq. (59) gives v s r c v_{\rm src} v src . The reconstruction uses signals linearized about the pump trajectory, calibrated incremental circuit parameters, and a measured section that contains the pulse. With complex signals the squares become squared moduli and the factors 1 / 2 1/2 1/2 become 1 / 4 1/4 1/4 . For the numerical results we set ϵ b = μ = Ω = 1 \epsilon_b=\mu=\Omega=1 ϵ b = μ = Ω = 1 ; time, frequency, wavenumber (including d d d and σ k \sigma_k σ k ), length and velocity are expressed in units of 1 / Ω 1/\Omega 1/Ω , Ω \Omega Ω , Ω / c b \Omega/c_b Ω/ c b , c b / Ω c_b/\Omega c b /Ω and c b c_b c b , respectively, and the quoted coefficients C s C_s C s , C b C_b C b , C g C_g C g and κ 0 \kappa_0 κ 0 are dimensionless in these units.
Fields are written X ( z , t ) = ( 2 π ) − 1 ∫ X ( k , t ) e i k z d k X(z,t)=(2\pi)^{-1}\int X(k,t)e^{ikz}\,dk X ( z , t ) = ( 2 π ) − 1 ∫ X ( k , t ) e ik z d k , so z ↔ i ∂ k z\leftrightarrow i\partial_k z ↔ i ∂ k , and time harmonics as e − i ω t e^{-i\omega t} e − iω t . With complex amplitudes, u E M = ( ϵ b ∣ E ∣ 2 + μ ∣ H ∣ 2 ) / 4 u_{\rm EM}=(\epsilon_b|E|^2+\mu|H|^2)/4 u EM = ( ϵ b ∣ E ∣ 2 + μ ∣ H ∣ 2 ) /4 and S z = Re ( E H ∗ ) / 2 S_z=\operatorname{Re}(E H^*)/2 S z = Re ( E H ∗ ) /2 are the averages of the values for the real and imaginary parts.
For a pulse whose spectrum lies at k > 0 k>0 k > 0 , these averages also give the energies and first moments of the single real field Re X \operatorname{Re}X Re X at every instant. With Re X = ( X + X ∗ ) / 2 \operatorname{Re}X=(X+X^*)/2 Re X = ( X + X ∗ ) /2 the cross terms obey
∫ X 2 d z = 1 2 π ∫ X ( k ) X ( − k ) d k = 0 , ∫ z X 2 d z = i 2 π ∫ X ( − k ) ∂ k X ( k ) d k = 0 , (18) \int X^2\,dz=\frac{1}{2\pi}\int X(k)X(-k)\,dk=0,\qquad\int zX^2\,dz=\frac{i}{2\pi}\int X(-k)\,\partial_kX(k)\,dk=0, \tag*{(18)} ∫ X 2 d z = 2 π 1 ∫ X ( k ) X ( − k ) d k = 0 , ∫ z X 2 d z = 2 π i ∫ X ( − k ) ∂ k X ( k ) d k = 0 , (18)
since the two factors in each integrand have disjoint supports when the spectrum avoids k ≤ 0 k\leq0 k ≤ 0 . The sector energies, first moments and Z Z Z of the real pulse therefore equal those of the complex description, while the local densities differ by terms that oscillate at about 2 k 0 2k_0 2 k 0 in z z z . For the untruncated Gaussian launches used here this correspondence holds up to negative-wavenumber tails of relative weight of order exp ( − k 0 2 / 2 σ k 2 ) \exp(-k_0^2/2\sigma_k^2) exp ( − k 0 2 /2 σ k 2 ) . Reconstructing Z ( t ) Z(t) Z ( t ) from measurements needs a time resolution finer than the modulation period.
At fixed k k k the state x = ( E , H , P , Q ) T \boldsymbol{x}=(E,H,P,Q)^{\mathsf{T}} x = ( E , H , P , Q ) T obeys
x ˙ = A ( k , t ) x , A = ( 0 − i k / ϵ b 0 − 1 / ϵ b − i k / μ 0 0 0 0 0 0 1 α 0 − κ ( t ) − η ) . (19) \dot{\boldsymbol{x}}=A(k,t)\boldsymbol{x},\qquad
A=\begin{pmatrix}
0 & -ik/\epsilon_b & 0 & -1/\epsilon_b\\
-ik/\mu& 0 & 0 & 0\\
0 & 0 & 0 & 1\\
\alpha& 0 & -\kappa(t) & -\eta
\end{pmatrix}. \tag*{(19)} x ˙ = A ( k , t ) x , A = 0 − ik / μ 0 α − ik / ϵ b 0 0 0 0 0 0 − κ ( t ) − 1/ ϵ b 0 1 − η . (19)
For κ = κ 0 \kappa=\kappa_0 κ = κ 0 the medium has ϵ ( ω ) = ϵ b + α / ( κ 0 − ω 2 − i η ω ) \epsilon(\omega)=\epsilon_b+\alpha/(\kappa_0-\omega^2-i\eta\omega) ϵ ( ω ) = ϵ b + α / ( κ 0 − ω 2 − i η ω ) , and k 2 = μ ϵ ( ω ) ω 2 k^2=\mu\epsilon(\omega)\omega^2 k 2 = μ ϵ ( ω ) ω 2 gives the static Lorentz branches. We call the one-period propagator M ( k ) = U k ( T , 0 ) M(k)=U_k(T,0) M ( k ) = U k ( T , 0 ) the monodromy and write its eigenvalues, the Floquet multipliers, as ζ β = e − i ω β T \zeta_\beta=e^{-i\omega_\beta T} ζ β = e − i ω β T . A Floquet mode is x = e − i ω β t ξ β ( k , t ) \boldsymbol{x}=e^{-i\omega_\beta t}\boldsymbol{\xi}_\beta(k,t) x = e − i ω β t ξ β ( k , t ) with T T T -periodic ξ β \boldsymbol{\xi}_\beta ξ β . Harmonic-space (Sambe) right and left eigenvectors are written ∣ R ⟩ |R\rangle ∣ R ⟩ , ⟨ L ∣ \langle L| ⟨ L ∣ . The Floquet calculations underlying Figure 2 use a Sambe truncation with harmonics n = − 6 , … , 6 n=-6,\ldots,6 n = − 6 , … , 6 , the Petermann factor for Figure 3 uses the time-domain representation of Sec. S5, and the packet modes of Figure 4 are obtained from the time-domain monodromy. Packet dynamics are propagated in the time domain as described in Sec. S7. The two-band illustration of Sec. S4 uses static eigenvectors.
Energy balance and centroid identity
With u m a t = ( ∣ Q ∣ 2 + κ ∣ P ∣ 2 ) / ( 4 α ) u_{\mathrm{mat}}=(|Q|^2+\kappa|P|^2)/(4\alpha) u mat = ( ∣ Q ∣ 2 + κ ∣ P ∣ 2 ) / ( 4 α ) the sector balances are
∂ t u E M + ∂ z S z = − 1 2 Re ( E Q ∗ ) , ∂ t u m a t = 1 2 Re ( E Q ∗ ) − η 2 α ∣ Q ∣ 2 + κ ˙ 4 α ∣ P ∣ 2 . (20) \partial_tu_{\mathrm{EM}}+\partial_zS_z=-\frac{1}{2}\operatorname{Re}(EQ^*),\qquad
\partial_tu_{\mathrm{mat}}=\frac{1}{2}\operatorname{Re}(EQ^*)-\frac{\eta}{2\alpha}|Q|^2+\frac{\dot{\kappa}}{4\alpha}|P|^2. \tag*{(20)} ∂ t u EM + ∂ z S z = − 2 1 Re ( E Q ∗ ) , ∂ t u mat = 2 1 Re ( E Q ∗ ) − 2 α η ∣ Q ∣ 2 + 4 α κ ˙ ∣ P ∣ 2 . (20)
Their sum is the exact local balance
∂ t u t o t + ∂ z S z = s = κ ˙ 4 α u P − 2 η u Q , u P = κ ∣ P ∣ 2 4 α , u Q = ∣ Q ∣ 2 4 α . (21) \partial_tu_{\mathrm{tot}}+\partial_zS_z=s=\frac{\dot{\kappa}}{4\alpha}u_P-2\eta u_Q,\qquad
u_P=\frac{\kappa|P|^2}{4\alpha},\qquad
u_Q=\frac{|Q|^2}{4\alpha}. \tag*{(21)} ∂ t u tot + ∂ z S z = s = 4 α κ ˙ u P − 2 η u Q , u P = 4 α κ ∣ P ∣ 2 , u Q = 4 α ∣ Q ∣ 2 . (21)
with u t o t = u E + u H + u P + u Q ≥ 0 u_{\mathrm{tot}}=u_E+u_H+u_P+u_Q\ge0 u tot = u E + u H + u P + u Q ≥ 0 . Pointwise ∣ S z ∣ ≤ ∣ E ∣ ∣ H ∣ / 2 ≤ c b u E M ≤ c b u t o t |S_z|\le|E||H|/2\le c_bu_{\mathrm{EM}}\le c_bu_{\mathrm{tot}} ∣ S z ∣ ≤ ∣ E ∣∣ H ∣/2 ≤ c b u EM ≤ c b u tot , so ∣ v f l u x ∣ ≤ c b |v_{\mathrm{flux}}|\le c_b ∣ v flux ∣ ≤ c b .
For a localized pulse with S z S_z S z and z S z zS_z z S z vanishing at the boundaries, multiplying Eq. (21) by z z z and integrating give
Z ˙ = 1 U ∫ S z d z + 1 U ∫ ( z − Z ) s d z ≡ v f l u x + v s r c . (22) \dot{Z}=\frac{1}{U}\int S_z\,\mathrm{d}z+\frac{1}{U}\int(z-Z)s\,\mathrm{d}z\equiv v_{\mathrm{flux}}+v_{\mathrm{src}}. \tag*{(22)} Z ˙ = U 1 ∫ S z d z + U 1 ∫ ( z − Z ) s d z ≡ v flux + v src . (22)
With U a = ∫ u a d z U_a=\int u_a\,\mathrm{d}z U a = ∫ u a d z , f a = U a / U f_a=U_a/U f a = U a / U , Z a = U a − 1 ∫ z u a d z Z_a=U_a^{-1}\int zu_a\,\mathrm{d}z Z a = U a − 1 ∫ z u a d z and the complement a ˉ \bar{a} a ˉ , Z = f a Z a + f a ˉ Z a ˉ Z=f_aZ_a+f_{\bar{a}}Z_{\bar{a}} Z = f a Z a + f a ˉ Z a ˉ , so U − 1 ∫ ( z − Z ) u a d z = f a f a ˉ ( Z a − Z a ˉ ) U^{-1}\int(z-Z)u_a\,\mathrm{d}z=f_af_{\bar{a}}(Z_a-Z_{\bar{a}}) U − 1 ∫ ( z − Z ) u a d z = f a f a ˉ ( Z a − Z a ˉ ) , which gives Eq. (3) . For any partition a ∣ a ˉ a|\bar{a} a ∣ a ˉ the energy variance splits as
σ z 2 = f a σ a 2 + f a ˉ σ a ˉ 2 + f a f a ˉ ( Δ Z a a ˉ ) 2 , ∣ f a f a ˉ Δ Z a a ˉ ∣ ≤ f a f a ˉ σ z ≤ σ z / 2 , (23) \sigma_z^2=f_a\sigma_a^2+f_{\bar{a}}\sigma_{\bar{a}}^2+f_af_{\bar{a}}(\Delta Z_{a\bar{a}})^2,\qquad
|f_af_{\bar{a}}\Delta Z_{a\bar{a}}|\le\sqrt{f_af_{\bar{a}}}\,\sigma_z\le\sigma_z/2, \tag*{(23)} σ z 2 = f a σ a 2 + f a ˉ σ a ˉ 2 + f a f a ˉ ( Δ Z a a ˉ ) 2 , ∣ f a f a ˉ Δ Z a a ˉ ∣ ≤ f a f a ˉ σ z ≤ σ z /2 , (23)
where σ a \sigma_a σ a and σ a ˉ \sigma_{\bar{a}} σ a ˉ are the internal widths of the two parts, which gives Eq. (4) .
The identity extends beyond the Maxwell–Lorentz medium to any evolution ∂ t x = − i H ( k , t ) x \partial_t\boldsymbol{x}=-iH(k,t)\boldsymbol{x} ∂ t x = − i H ( k , t ) x in a fixed physical basis with stored energy U = x † G x U=\boldsymbol{x}^{\dagger}G\boldsymbol{x} U = x † G x , where G ( t ) G(t) G ( t ) is positive and independent of k k k , and with centroid Z = Re ( i x † G ∂ k x ) / U Z=\operatorname{Re}(i\boldsymbol{x}^{\dagger}G\partial_k\boldsymbol{x})/U Z = Re ( i x † G ∂ k x ) / U . Then U ˙ = x † Π x \dot{U}=\boldsymbol{x}^{\dagger}\Pi\boldsymbol{x} U ˙ = x † Π x with the power operator Π = i ( H † G − G H ) + G ˙ \Pi=i(H^{\dagger}G-GH)+\dot{G} Π = i ( H † G − G H ) + G ˙ , and
Z ˙ = Re [ x † G ( ∂ k H ) x ] U + Re ( i x † Π ∂ k x ) − Z x † Π x U . (24) \dot{Z}=\frac{\operatorname{Re}[\boldsymbol{x}^{\dagger}G(\partial_kH)\boldsymbol{x}]}{U}
+\frac{\operatorname{Re}(i\boldsymbol{x}^{\dagger}\Pi\partial_k\boldsymbol{x})-Z\boldsymbol{x}^{\dagger}\Pi\boldsymbol{x}}{U}. \tag*{(24)} Z ˙ = U Re [ x † G ( ∂ k H ) x ] + U Re ( i x † Π ∂ k x ) − Z x † Π x . (24)
A common gain Π → Π + c ( t ) G \Pi\to\Pi+c(t)G Π → Π + c ( t ) G with real c ( t ) c(t) c ( t ) leaves the second term unchanged. For a finite packet the numerator and denominator are integrated over k k k , and the moments must exist. If G = ∑ a G a G=\sum_aG_a G = ∑ a G a with nonnegative, k k k -independent sector matrices G a G_a G a that define the sector energies and moments, and Π = ∑ a r a ( t ) G a \Pi=\sum_ar_a(t)G_a Π = ∑ a r a ( t ) G a , the source term splits into the corresponding sector channels. For the Maxwell–Lorentz medium, G = diag ( ϵ b , μ , κ / α , 1 / α ) / 4 G=\operatorname{diag}(\epsilon_b,\mu,\kappa/\alpha,1/\alpha)/4 G = diag ( ϵ b , μ , κ / α , 1/ α ) /4 and Π = diag ( 0 , 0 , κ ˙ / ( 4 α ) , − 2 η / ( 4 α ) ) \Pi=\operatorname{diag}(0,0,\dot{\kappa}/(4\alpha),-2\eta/(4\alpha)) Π = diag ( 0 , 0 , κ ˙ / ( 4 α ) , − 2 η / ( 4 α )) , which recovers Eq. (3) .
Packets, connections and the gap-side null
To express the sector centroids through the spectral amplitudes, write each sector density as u a = c a ( t ) ∣ X a ∣ 2 u_a=c_a(t)|X_a|^2 u a = c a ( t ) ∣ X a ∣ 2 with ( c E , c H , c P , c Q ) = ( ϵ b , μ , κ / α , 1 / α ) / 4 (c_E,c_H,c_P,c_Q)=(\epsilon_b,\mu,\kappa/\alpha,1/\alpha)/4 ( c E , c H , c P , c Q ) = ( ϵ b , μ , κ / α , 1/ α ) /4 . With X a ( k , t ) = F a ( k , t ) X_a(k,t)=\mathcal{F}_a(k,t) X a ( k , t ) = F a ( k , t ) , Parseval’s identity gives
Z a = Re ∫ d k F a ∗ ∂ k F a ∫ d k ∣ F a ∣ 2 . (25) Z_a=\frac{\operatorname{Re}\int\mathrm{d}k\,\mathcal{F}_a^*\partial_k\mathcal{F}_a}{\int\mathrm{d}k\,|\mathcal{F}_a|^2}. \tag*{(25)} Z a = ∫ d k ∣ F a ∣ 2 Re ∫ d k F a ∗ ∂ k F a . (25)
Its denominator times c a / ( 2 π ) c_a/(2\pi) c a / ( 2 π ) is the sector energy U a U_a U a , and J a = U a Z a J_a=U_aZ_a J a = U a Z a is the sector first moment. The identity z ↔ i ∂ k z\leftrightarrow i\partial_k z ↔ i ∂ k thus expresses each sector centroid as an average of − ∂ k arg F a -\partial_k\arg F_a − ∂ k arg F a , weighted by that component’s spectral energy. For a single branch, F β , a = ψ ( k ) ξ β , a ( k , t ) e − i ω β t F_{\beta,a}=\psi(k)\xi_{\beta,a}(k,t)e^{-i\omega_\beta t} F β , a = ψ ( k ) ξ β , a ( k , t ) e − i ω β t with ω β = ω β , R + i Γ β \omega_\beta=\omega_{\beta,\mathrm{R}}+i\Gamma_\beta ω β = ω β , R + i Γ β . With ψ = ∣ ψ ∣ e i ϕ ψ \psi=|\psi|e^{i\phi_\psi} ψ = ∣ ψ ∣ e i ϕ ψ we define the envelope connection, the component connection and the component weight,
A e n v = − ∂ k ϕ ψ , A β , a = − Im [ ξ β , a ∗ ∂ k ξ β , a ] ∣ ξ β , a ∣ 2 , p β , a = ∣ ψ ∣ 2 ∣ ξ β , a ∣ 2 e 2 Γ β t , (26) \mathcal{A}_{\mathrm{env}}=-\partial_k\phi_\psi,\qquad\mathcal{A}_{\beta,a}=-\frac{\operatorname{Im}[\xi_{\beta,a}^{*}\partial_k\xi_{\beta,a}]}{|\xi_{\beta,a}|^2},\qquad p_{\beta,a}=|\psi|^2|\xi_{\beta,a}|^2e^{2\Gamma_\beta t}, \tag*{(26)} A env = − ∂ k ϕ ψ , A β , a = − ∣ ξ β , a ∣ 2 Im [ ξ β , a ∗ ∂ k ξ β , a ] , p β , a = ∣ ψ ∣ 2 ∣ ξ β , a ∣ 2 e 2 Γ β t , (26)
and Eq. (S12) then gives
Z β , a w p ( t ) = ⟨ A e n v + t ∂ k ω β , R + A β , a ⟩ β , a , t , (27) Z_{\beta,a}^{\mathrm{wp}}(t)=\left\langle\mathcal{A}_{\mathrm{env}}+t\partial_k\omega_{\beta,\mathrm{R}}+\mathcal{A}_{\beta,a}\right\rangle_{\beta,a,t}, \tag*{(27)} Z β , a wp ( t ) = ⟨ A env + t ∂ k ω β , R + A β , a ⟩ β , a , t , (27)
where ⟨ ⋅ ⟩ β , a , t \langle\cdot\rangle_{\beta,a,t} ⟨ ⋅ ⟩ β , a , t averages with p β , a p_{\beta,a} p β , a . Differences Z β , a w p − Z β , j w p Z_{\beta,a}^{\mathrm{wp}}-Z_{\beta,j}^{\mathrm{wp}} Z β , a wp − Z β , j wp contain the common terms averaged with different weights, so they do not reduce to differences of local connections at finite bandwidth. In this notation the fractions and sector separations are
f β , a w p = c a ∫ p β , a d k ∑ j c j ∫ p β , j d k , Δ Z β , a a ˉ w p = 1 f β , a w p ∑ j ≠ a f β , j w p ( Z β , a w p − Z β , j w p ) . (28) f_{\beta,a}^{\mathrm{wp}}=\frac{c_a\int p_{\beta,a}\,\mathrm{d}k}{\sum_j c_j\int p_{\beta,j}\,\mathrm{d}k},\qquad\Delta Z_{\beta,a\bar a}^{\mathrm{wp}}=\frac{1}{f_{\beta,a}^{\mathrm{wp}}}\sum_{j\ne a}f_{\beta,j}^{\mathrm{wp}}\left(Z_{\beta,a}^{\mathrm{wp}}-Z_{\beta,j}^{\mathrm{wp}}\right). \tag*{(28)} f β , a wp = ∑ j c j ∫ p β , j d k c a ∫ p β , a d k , Δ Z β , a a ˉ wp = f β , a wp 1 j = a ∑ f β , j wp ( Z β , a wp − Z β , j wp ) . (28)
Under the gauge change ξ → e i γ ( k ) ξ \xi\to e^{i\gamma(k)}\xi ξ → e iγ ( k ) ξ and ψ → e − i γ ( k ) ψ \psi\to e^{-i\gamma(k)}\psi ψ → e − iγ ( k ) ψ the amplitude F F F is unchanged, A e n v → A e n v + ∂ k γ \mathcal{A}_{\mathrm{env}}\to\mathcal{A}_{\mathrm{env}}+\partial_k\gamma A env → A env + ∂ k γ and A β , a → A β , a − ∂ k γ \mathcal{A}_{\beta,a}\to\mathcal{A}_{\beta,a}-\partial_k\gamma A β , a → A β , a − ∂ k γ , so the sum A e n v + A β , a \mathcal{A}_{\mathrm{env}}+\mathcal{A}_{\beta,a} A env + A β , a and all connection differences are gauge invariant.
The narrow-band limit recovers the main-text relations, because at fixed regular k 0 k_0 k 0 on the passband, over a fixed window and for σ k ≪ ∣ k E P − k 0 ∣ \sigma_k\ll|k_{\mathrm{EP}}-k_0| σ k ≪ ∣ k EP − k 0 ∣ , ⟨ F ⟩ β , a , t = F ( k 0 , t ) + O ( σ k ) \langle F\rangle_{\beta,a,t}=F(k_0,t)+O(\sigma_k) ⟨ F ⟩ β , a , t = F ( k 0 , t ) + O ( σ k ) . In the sector differences the common terms cancel at leading order, leaving Δ A β , a a ˉ ( k 0 , t ) \Delta\mathcal{A}_{\beta,a\bar a}(k_0,t) Δ A β , a a ˉ ( k 0 , t ) as the leading sector-to-complement separation, with the complement connection A β , a ˉ = ∑ j ≠ a f β , j A β , j / f β , a ˉ \mathcal{A}_{\beta,\bar a}=\sum_{j\ne a}f_{\beta,j}\mathcal{A}_{\beta,j}/f_{\beta,\bar a} A β , a ˉ = ∑ j = a f β , j A β , j / f β , a ˉ and f β , a = c a ∣ ξ β , a ∣ 2 / ∑ j c j ∣ ξ β , j ∣ 2 f_{\beta,a}=c_a|\xi_{\beta,a}|^2/\sum_jc_j|\xi_{\beta,j}|^2 f β , a = c a ∣ ξ β , a ∣ 2 / ∑ j c j ∣ ξ β , j ∣ 2 at k 0 k_0 k 0 . Summing Eq. (S14) with the fractions gives
Z ( t ) = A e n v ( k 0 ) + t ∂ k ω β , R ( k 0 ) + ∑ a f β , a A β , a ( k 0 , t ) + O ( σ k ) . (29) Z(t)=\mathcal{A}_{\mathrm{env}}(k_0)+t\partial_k\omega_{\beta,\mathrm{R}}(k_0)+\sum_a f_{\beta,a}\mathcal{A}_{\beta,a}(k_0,t)+O(\sigma_k). \tag*{(29)} Z ( t ) = A env ( k 0 ) + t ∂ k ω β , R ( k 0 ) + a ∑ f β , a A β , a ( k 0 , t ) + O ( σ k ) . (29)
Its connection part ∑ a f β , a A β , a \sum_a f_{\beta,a}\mathcal{A}_{\beta,a} ∑ a f β , a A β , a is T T T -periodic, whereas the finite-bandwidth remainder need not be. For the Gaussian preparation, differential decay across the spectrum shifts the spectral center by O ( σ k 2 t ) O(\sigma_k^2t) O ( σ k 2 t ) , and averaging Eq. (S9) over one period then gives Eq. (7) at leading order. The error constants depend on k 0 k_0 k 0 and need not stay bounded as k 0 → k E P k_0\to k_{\mathrm{EP}} k 0 → k EP , where the joint limit of Sec. S6 applies.
At component nodes the connections of Eq. (S13) are undefined, so to avoid dividing by component intensities, define
n β , a = c a ∣ ξ β , a ∣ 2 , N β = ∑ a n β , a > 0 , J β , a = − c a Im ( ξ β , a ∗ ∂ k ξ β , a ) , n_{\beta,a}=c_a|\xi_{\beta,a}|^2,\qquad N_\beta=\sum_a n_{\beta,a}>0,\qquad\mathcal{J}_{\beta,a}=-c_a\operatorname{Im}(\xi_{\beta,a}^{*}\partial_k\xi_{\beta,a}), n β , a = c a ∣ ξ β , a ∣ 2 , N β = a ∑ n β , a > 0 , J β , a = − c a Im ( ξ β , a ∗ ∂ k ξ β , a ) ,
L β , a = J β , a − f β , a ∑ j J β , j N β = f β , a f β , a ˉ Δ A β , a a ˉ . (30) \mathcal{L}_{\beta,a}=\frac{\mathcal{J}_{\beta,a}-f_{\beta,a}\sum_j\mathcal{J}_{\beta,j}}{N_\beta}=f_{\beta,a}f_{\beta,\bar a}\Delta\mathcal{A}_{\beta,a\bar a}. \tag*{(30)} L β , a = N β J β , a − f β , a ∑ j J β , j = f β , a f β , a ˉ Δ A β , a a ˉ . (30)
The first expression for L β , a \mathcal{L}_{\beta,a} L β , a stays defined at component nodes, and the local source velocity is
v s r c , β l o c = κ ˙ κ L β , P − 2 η L β , Q . (31) v_{\mathrm{src},\beta}^{\mathrm{loc}}=\frac{\dot{\kappa}}{\kappa}\mathcal{L}_{\beta,P}-2\eta\mathcal{L}_{\beta,Q}. \tag*{(31)} v src , β loc = κ κ ˙ L β , P − 2 η L β , Q . (31)
For a state that mixes branches, as in Sec. S12, the moments of Eq. (S17) are evaluated from the full propagated state and its wavenumber derivative at fixed time. The amplitudes A t r A_{\mathrm{tr}} A tr and A b A_b A b are centered rms values of v s r c v_{\mathrm{src}} v src , and the total centroid velocity adds v f l u x v_{\mathrm{flux}} v flux , with ∣ v f l u x ∣ ≤ c b |v_{\mathrm{flux}}|\le c_b ∣ v flux ∣ ≤ c b .
The gap-side null follows from an antilinear symmetry at real k k k : with P = diag ( 1 , − 1 , 1 , 1 ) \mathcal{P}=\operatorname{diag}(1,-1,1,1) P = diag ( 1 , − 1 , 1 , 1 ) ,
P A ∗ P = A , P U k ∗ P = U k , P M ∗ P = M , (32) \mathcal{P}A^{*}\mathcal{P}=A,\qquad\mathcal{P}U_k^{*}\mathcal{P}=U_k,\qquad\mathcal{P}M^{*}\mathcal{P}=M, \tag*{(32)} P A ∗ P = A , P U k ∗ P = U k , P M ∗ P = M , (32)
so the equations of motion are real in the frame ( E , − i H , P , Q ) (E,-iH,P,Q) ( E , − i H , P , Q ) . Take a smooth branch with a nondegenerate negative real multiplier ζ β \zeta_\beta ζ β on an open k k k -interval inside the gap, with ω β = Ω / 2 + i Γ β \omega_\beta=\Omega/2+i\Gamma_\beta ω β = Ω/2 + i Γ β . Simplicity gives P ∣ r β ∗ ⟩ = e i θ β ( k ) ∣ r β ⟩ \mathcal{P}|r_\beta^{*}\rangle=e^{i\theta_\beta(k)}|r_\beta\rangle P ∣ r β ∗ ⟩ = e i θ β ( k ) ∣ r β ⟩ with unit-modulus factor, and propagation gives
ξ β , a = e i ( Ω t − θ β ) / 2 s a x β , a , ( s E , s H , s P , s Q ) = ( 1 , i , 1 , 1 ) , x β , a ∈ R . (33) \xi_{\beta,a}=e^{i(\Omega t-\theta_\beta)/2}s_ax_{\beta,a},\qquad(s_E,s_H,s_P,s_Q)=(1,i,1,1),\qquad x_{\beta,a}\in\mathbb{R}. \tag*{(33)} ξ β , a = e i ( Ω t − θ β ) /2 s a x β , a , ( s E , s H , s P , s Q ) = ( 1 , i , 1 , 1 ) , x β , a ∈ R . (33)
Then J β , a = ( ∂ k θ β ) n β , a / 2 \mathcal{J}_{\beta,a}=(\partial_k\theta_\beta)n_{\beta,a}/2 J β , a = ( ∂ k θ β ) n β , a /2 and L β , a = 0 \mathcal{L}_{\beta,a}=0 L β , a = 0 , also at nodes, so the local source velocity and the local squared separation vanish channel by channel for either sign of Γ β \Gamma_\beta Γ β . For a packet on this branch, Eq. (27) gives
Z β , a w p = ⟨ G β ⟩ β , a , t , G β ( k ) = A e n v ( k ) + 1 2 ∂ k θ β ( k ) , (34) Z_{\beta,a}^{\mathrm{wp}}=\langle\mathcal{G}_\beta\rangle_{\beta,a,t},\qquad\mathcal{G}_\beta(k)=A_{\mathrm{env}}(k)+\frac{1}{2}\partial_k\theta_\beta(k), \tag*{(34)} Z β , a wp = ⟨ G β ⟩ β , a , t , G β ( k ) = A env ( k ) + 2 1 ∂ k θ β ( k ) , (34)
with gauge-invariant G β \mathcal{G}_\beta G β . Unequal sector weights can give unequal averages of a nonconstant G β \mathcal{G}_\beta G β .
For the unchirped eigen-packets of the main text we launch each eigenvector with its electric component real and positive wherever E ( k , 0 ) ≠ 0 E(k,0)\ne0 E ( k , 0 ) = 0 , under a real envelope. On the gap branch this sets G β ≡ 0 \mathcal{G}_\beta\equiv0 G β ≡ 0 , so the propagated state is real in the frame ( E , − i H , P , Q ) (E,-iH,P,Q) ( E , − i H , P , Q ) up to a k k k -independent factor, the gap-side integrands Re ( i F a ∗ ∂ k F a ) \operatorname{Re}(iF_a^*\partial_kF_a) Re ( i F a ∗ ∂ k F a ) of Eq. (25) vanish, and the gap side adds energy weight to every sector without adding first moment. Hence the source velocity of the packet comes only from passband contributions to the normalized moments, and their decrease accounts for the fade of Figure 4 (c). We describe the gap side through its contributions to the moment integrals of the full continuous packet. A packet cut sharply at k E P k_{\mathrm{EP}} k EP with nonzero endpoint amplitude has 1 / z 1/z 1/ z tails and divergent first moments.
To discuss finite-second-moment approximations, consider the launch ψ 0 ( k ) x ( k ) \psi_0(k)x(k) ψ 0 ( k ) x ( k ) with the real Gaussian envelope ψ 0 \psi_0 ψ 0 and a mode profile x x x whose spectral completion is regular away from the EP. Near the EP the profile has ∂ k x ∝ ∣ k − k E P ∣ − 1 / 2 \partial_kx\propto|k-k_{\mathrm{EP}}|^{-1/2} ∂ k x ∝ ∣ k − k EP ∣ − 1/2 (Sec. S4). First-moment integrands are then integrable, while ∫ ∣ ∂ k x ∣ 2 d k \int|\partial_kx|^2\,dk ∫ ∣ ∂ k x ∣ 2 d k diverges logarithmically, so the ideal eigen-packet has finite centroids and an infinite spatial variance. Replace the profile by x ϵ x_\epsilon x ϵ , equal to x x x outside an interval of width ϵ k \epsilon_k ϵ k about k E P k_{\mathrm{EP}} k EP , with no added rapid phase and with ∥ x ϵ − x ∥ = O ( ϵ k 1 / 2 ) \lVert x_\epsilon-x\rVert=O(\epsilon_k^{1/2}) ∥ x ϵ − x ∥ = O ( ϵ k 1/2 ) and ∥ ∂ k x ϵ ∥ = O ( ϵ k − 1 / 2 ) \lVert\partial_kx_\epsilon\rVert=O(\epsilon_k^{-1/2}) ∥ ∂ k x ϵ ∥ = O ( ϵ k − 1/2 ) inside it. With continuous matching at the smoothing boundaries and a square-integrable derivative of the full prepared spectrum outside the smoothing interval, the regularized launch has finite variance. The local smoothing changes its unnormalized first moments by O ( ϵ k 1 / 2 ) O(\epsilon_k^{1/2}) O ( ϵ k 1/2 ) and its energy by O ( ϵ k 3 / 2 ) O(\epsilon_k^{3/2}) O ( ϵ k 3/2 ) . At fixed λ \lambda λ , over a fixed window and for small σ k \sigma_k σ k , the packet energy obeys c σ k ≤ U ( t ) ≤ C σ k c\sigma_k\le U(t)\le C\sigma_k c σ k ≤ U ( t ) ≤ C σ k , so the source velocity changes by O ( ϵ k 1 / 2 / σ k ) O(\epsilon_k^{1/2}/\sigma_k) O ( ϵ k 1/2 / σ k ) and σ k A t r w p \sqrt{\sigma_k}A_{\mathrm{tr}}^{\mathrm{wp}} σ k A tr wp by O [ ( ϵ k / σ k ) 1 / 2 ] O[(\epsilon_k/\sigma_k)^{1/2}] O [( ϵ k / σ k ) 1/2 ] . Therefore the scaled response of Eq. (12) is unchanged as ϵ k / σ k → 0 \epsilon_k/\sigma_k\to0 ϵ k / σ k → 0 . Section S9 implements this regularization numerically.
Floquet modes near the exceptional point
Expanding ξ a = ∑ n ξ a , n e − i n Ω t \xi_a=\sum_n\xi_{a,n}e^{-in\Omega t} ξ a = ∑ n ξ a , n e − in Ω t gives the connection in harmonic form,
A a = J a d i a g + J a c o h ( t ) N a d i a g + N a c o h ( t ) , N a c o h = ∑ n ≠ n ′ ξ a , n ∗ ξ a , n ′ e i ( n − n ′ ) Ω t , \mathcal{A}_a=\frac{\mathcal{J}_a^{\mathrm{diag}}+\mathcal{J}_a^{\mathrm{coh}}(t)}{\mathcal{N}_a^{\mathrm{diag}}+\mathcal{N}_a^{\mathrm{coh}}(t)},\qquad
\mathcal{N}_a^{\mathrm{coh}}=\sum_{n\ne n'}\xi_{a,n}^{*}\xi_{a,n'}e^{i(n-n')\Omega t}, A a = N a diag + N a coh ( t ) J a diag + J a coh ( t ) , N a coh = n = n ′ ∑ ξ a , n ∗ ξ a , n ′ e i ( n − n ′ ) Ω t ,
J a c o h = − Im ∑ n ≠ n ′ ξ a , n ∗ ∂ k ξ a , n ′ e i ( n − n ′ ) Ω t . (35) \mathcal{J}_a^{\mathrm{coh}}=-\operatorname{Im}\sum_{n\ne n'}\xi_{a,n}^{*}\partial_k\xi_{a,n'}e^{i(n-n')\Omega t}. \tag*{(35)} J a coh = − Im n = n ′ ∑ ξ a , n ∗ ∂ k ξ a , n ′ e i ( n − n ′ ) Ω t . (35)
with N a d i a g = ∑ n ∣ ξ a , n ∣ 2 \mathcal{N}_a^{\mathrm{diag}}=\sum_n|\xi_{a,n}|^2 N a diag = ∑ n ∣ ξ a , n ∣ 2 and J a d i a g = − Im ∑ n ξ a , n ∗ ∂ k ξ a , n \mathcal{J}_a^{\mathrm{diag}}=-\operatorname{Im}\sum_n\xi_{a,n}^{*}\partial_k\xi_{a,n} J a diag = − Im ∑ n ξ a , n ∗ ∂ k ξ a , n ; the common sector weight c a c_a c a of Eq. (30) cancels between numerator and denominator. All time dependence of the connection comes from coherence between unequal harmonics.
Near the EP the modes follow from a Jordan chain, since at k E P k_{\mathrm{EP}} k EP the monodromy M 0 M_0 M 0 has a defective double multiplier ζ 0 \zeta_0 ζ 0 with
( M 0 − ζ 0 ) ξ 0 = 0 , ( M 0 − ζ 0 ) ξ 1 = ξ 0 , ℓ T ( M 0 − ζ 0 ) = 0 , (M_0-\zeta_0)\xi_0=0,\qquad(M_0-\zeta_0)\xi_1=\xi_0,\qquad\ell^T(M_0-\zeta_0)=0, ( M 0 − ζ 0 ) ξ 0 = 0 , ( M 0 − ζ 0 ) ξ 1 = ξ 0 , ℓ T ( M 0 − ζ 0 ) = 0 ,
ℓ T ξ 0 = 0 , ℓ T ξ 1 = 1 , (36) \ell^T\xi_0=0,\qquad\ell^T\xi_1=1, \tag*{(36)} ℓ T ξ 0 = 0 , ℓ T ξ 1 = 1 , (36)
and ξ 1 ⊥ ξ 0 \xi_1\perp\xi_0 ξ 1 ⊥ ξ 0 . Because M 0 M_0 M 0 is real in the frame ( E , − i H , P , Q ) (E,-iH,P,Q) ( E , − i H , P , Q ) and ζ 0 \zeta_0 ζ 0 is real and negative, ξ 0 \xi_0 ξ 0 and ξ 1 \xi_1 ξ 1 are real there. The propagated vectors a ( t ) = U k E P ( t ) ξ 0 a(t)=U_{k_{\mathrm{EP}}}(t)\xi_0 a ( t ) = U k EP ( t ) ξ 0 and b ( t ) = U k E P ( t ) ξ 1 b(t)=U_{k_{\mathrm{EP}}}(t)\xi_1 b ( t ) = U k EP ( t ) ξ 1 obey a ( t + T ) = ζ 0 a ( t ) a(t+T)=\zeta_0a(t) a ( t + T ) = ζ 0 a ( t ) and b ( t + T ) = ζ 0 b ( t ) + a ( t ) b(t+T)=\zeta_0b(t)+a(t) b ( t + T ) = ζ 0 b ( t ) + a ( t ) , and the periodic chain is
ξ 0 p e r ( t ) = e − Λ 0 t a ( t ) , ξ 1 p e r ( t ) = e − Λ 0 t [ b ( t ) − t T ζ 0 a ( t ) ] , Λ 0 = ln ∣ ζ 0 ∣ − i π T . (37) \xi_0^{\mathrm{per}}(t)=e^{-\Lambda_0t}a(t),\qquad
\xi_1^{\mathrm{per}}(t)=e^{-\Lambda_0t}\left[b(t)-\frac{t}{T\zeta_0}a(t)\right],\qquad
\Lambda_0=\frac{\ln|\zeta_0|-i\pi}{T}. \tag*{(37)} ξ 0 per ( t ) = e − Λ 0 t a ( t ) , ξ 1 per ( t ) = e − Λ 0 t [ b ( t ) − T ζ 0 t a ( t ) ] , Λ 0 = T ln ∣ ζ 0 ∣ − iπ . (37)
An ordinary Floquet eigenmode ξ 0 p e r \xi_0^{\mathrm{per}} ξ 0 per exists at the EP, and the generalized partner carries the factor linear in t t t .
For small nonzero detuning d = k E P − k d=k_{\mathrm{EP}}-k d = k EP − k from the EP, with M ( k ) = M 0 − d M 1 M(k)=M_0-dM_1 M ( k ) = M 0 − d M 1 to first order, the chain yields the square-root expansion: to leading order the half-splitting of the multipliers obeys δ ζ 2 = K M d \delta\zeta^2=K_Md δ ζ 2 = K M d with K M = − ℓ T M 1 ξ 0 K_M=-\ell^TM_1\xi_0 K M = − ℓ T M 1 ξ 0 , and the eigenvectors are v ± = ξ 0 ± δ ζ ξ 1 v_\pm=\xi_0\pm\delta\zeta\,\xi_1 v ± = ξ 0 ± δ ζ ξ 1 with ζ ± = ζ 0 ± δ ζ \zeta_\pm=\zeta_0\pm\delta\zeta ζ ± = ζ 0 ± δ ζ . We fix the branch of δ ζ \delta\zeta δ ζ so that v + v_+ v + has the larger real quasifrequency. For the medium of the main text, ζ 0 = − 0.98379 \zeta_0=-0.98379 ζ 0 = − 0.98379 and K M = − 0.60339 K_M=-0.60339 K M = − 0.60339 , so δ ζ = i ∣ K M ∣ d \delta\zeta=i\sqrt{\lvert K_M\rvert d} δ ζ = i ∣ K M ∣ d on the passband side. On the gap side the same expansion gives the growth-rate splitting Γ − Γ E P ≃ ρ h \Gamma-\Gamma_{\mathrm{EP}}\simeq\sqrt{\rho h} Γ − Γ EP ≃ ρ h with
ρ = ∣ K M ∣ T 2 ζ 0 2 = 0.0158 , Γ E P = ln ∣ ζ 0 ∣ T . (38) \rho=\frac{\lvert K_M\rvert}{T^2\zeta_0^2}=0.0158,\qquad\Gamma_{\mathrm{EP}}=\frac{\ln\lvert\zeta_0\rvert}{T}. \tag*{(38)} ρ = T 2 ζ 0 2 ∣ K M ∣ = 0.0158 , Γ EP = T ln ∣ ζ 0 ∣ . (38)
The periodic modes are ξ ± ( k , t ) = e − Λ ± t U k ( t ) v ± \xi_{\pm}(k,t)=e^{-\Lambda_{\pm}t}U_k(t)v_{\pm} ξ ± ( k , t ) = e − Λ ± t U k ( t ) v ± with Λ ± = Λ 0 ± δ ζ / ( T ζ 0 ) \Lambda_{\pm}=\Lambda_0\pm\delta\zeta/(T\zeta_0) Λ ± = Λ 0 ± δ ζ / ( T ζ 0 ) to first order in δ ζ \delta\zeta δ ζ , and
ξ β = ξ 0 p e r + β d χ + O ( d ) , χ = δ ζ d ξ 1 p e r , ∂ k ξ β , a = − β 2 d χ a + O ( 1 ) , (39) \xi_{\beta}=\xi_0^{\mathrm{per}}+\beta\sqrt{d}\,\chi+O(d),\qquad\chi=\frac{\delta\zeta}{\sqrt{d}}\xi_1^{\mathrm{per}},\qquad\partial_k\xi_{\beta,a}=-\frac{\beta}{2\sqrt{d}}\chi_a+O(1), \tag*{(39)} ξ β = ξ 0 per + β d χ + O ( d ) , χ = d δ ζ ξ 1 per , ∂ k ξ β , a = − 2 d β χ a + O ( 1 ) , (39)
uniformly for 0 ≤ t ≤ T 0\leq t\leq T 0 ≤ t ≤ T . Away from nodes of ξ 0 , a p e r \xi_{0,a}^{\mathrm{per}} ξ 0 , a per ,
A β , a = β C a ( t ) d + O ( 1 ) , C a = 1 2 Im [ ξ 0 , a p e r ∗ χ a ] ∣ ξ 0 , a p e r ∣ 2 . (40) \mathcal{A}_{\beta,a}=\beta\frac{C_a(t)}{\sqrt{d}}+O(1),\qquad C_a=\frac{1}{2}\frac{\operatorname{Im}[\xi_{0,a}^{\mathrm{per}*}\chi_a]}{\lvert\xi_{0,a}^{\mathrm{per}}\rvert^2}. \tag*{(40)} A β , a = β d C a ( t ) + O ( 1 ) , C a = 2 1 ∣ ξ 0 , a per ∣ 2 Im [ ξ 0 , a per ∗ χ a ] . (40)
To extract the trembling coefficient from this expansion, we use the regular EP limits J a E P ( t ) \mathcal{J}_a^{\mathrm{EP}}(t) J a EP ( t ) of the sector fractions and the positive lower bound of N E P = ∑ a c a ∣ ξ 0 , a p e r ∣ 2 N_{\mathrm{EP}}=\sum_a c_a\lvert\xi_{0,a}^{\mathrm{per}}\rvert^2 N EP = ∑ a c a ∣ ξ 0 , a per ∣ 2 over one period. With
J a ( − 1 ) = c a 2 Im [ ξ 0 , a p e r ∗ χ a ] , S a = J a ( − 1 ) − J a E P ∑ j J j ( − 1 ) N E P , L β , a = β S a d + O ( 1 ) , (41) \mathcal{J}_a^{(-1)}=\frac{c_a}{2}\operatorname{Im}[\xi_{0,a}^{\mathrm{per}*}\chi_a],\qquad S_a=\frac{\mathcal{J}_a^{(-1)}-\mathcal{J}_a^{\mathrm{EP}}\sum_j\mathcal{J}_j^{(-1)}}{N_{\mathrm{EP}}},\qquad L_{\beta,a}=\beta\frac{S_a}{\sqrt{d}}+O(1), \tag*{(41)} J a ( − 1 ) = 2 c a Im [ ξ 0 , a per ∗ χ a ] , S a = N EP J a ( − 1 ) − J a EP ∑ j J j ( − 1 ) , L β , a = β d S a + O ( 1 ) , (41)
where the last relation follows from Eq. (30) uniformly, including nodes, we obtain
v s r c , β = β V ( t ) d + O ( 1 ) , V = κ ˙ κ S P − 2 η S Q , C s = ⟨ ( V − ⟨ V ⟩ T ) 2 ⟩ T 1 / 2 , (42) v_{\mathrm{src},\beta}=\beta\frac{\mathcal{V}(t)}{\sqrt{d}}+O(1),\qquad\mathcal{V}=\frac{\dot{\kappa}}{\kappa}S_P-2\eta S_Q,\qquad C_s=\left\langle\left(\mathcal{V}-\langle\mathcal{V}\rangle_T\right)^2\right\rangle_T^{1/2}, \tag*{(42)} v src , β = β d V ( t ) + O ( 1 ) , V = κ κ ˙ S P − 2 η S Q , C s = ⟨ ( V − ⟨ V ⟩ T ) 2 ⟩ T 1/2 , (42)
which gives Eq. (9) whenever C s ≠ 0 C_s\ne0 C s = 0 , with C s = 0.0822 C_s=0.0822 C s = 0.0822 . Over one period the divergent part of ∂ k x \partial_kx ∂ k x is ∓ ( ∂ δ δ ζ ) b ( t ) \mp(\partial_\delta\delta\zeta)b(t) ∓ ( ∂ δ δ ζ ) b ( t ) for x = U k ( t ) v ± x=U_k(t)v_{\pm} x = U k ( t ) v ± , so the sector offsets arise from interference between the propagated chain vectors a ( t ) a(t) a ( t ) and b ( t ) b(t) b ( t ) . Integrating the same expansion over a fixed number of periods in the local narrow-band limit gives the mean-centered centroid displacement of Sec. S7,
δ Z β ( t ) = β d [ Υ ( t ) − ⟨ Υ ⟩ I ] + O ( 1 ) , Υ ( t ) = ∫ 0 t [ V ( t ′ ) − ⟨ V ⟩ T ] d t ′ , (43) \delta Z_\beta(t)=\frac{\beta}{\sqrt{d}}\left[\Upsilon(t)-\langle\Upsilon\rangle_I\right]+O(1),\qquad\Upsilon(t)=\int_0^t\left[\mathcal{V}(t')-\langle\mathcal{V}\rangle_T\right]dt', \tag*{(43)} δ Z β ( t ) = d β [ Υ ( t ) − ⟨ Υ ⟩ I ] + O ( 1 ) , Υ ( t ) = ∫ 0 t [ V ( t ′ ) − ⟨ V ⟩ T ] d t ′ , (43)
with V \mathcal{V} V of Eq. (42) , I I I the detrending interval of Sec. S7, and the bounded flux contribution in the remainder. A nonzero centered source waveform therefore gives the centered rms displacement the same exponent as A t r A_{\mathrm{tr}} A tr , with a coefficient that follows from the centered primitive and is not fixed by C s C_s C s alone.
A two-band reduction illustrates how interference between the resonant harmonics enters C a C_a C a . Keeping the positive-frequency static branch in the n = 0 n=0 n = 0 harmonic and the negative-frequency branch in n = − 1 n=-1 n = − 1 , with biorthogonal static eigenvectors ∣ R s , ± ⟩ \lvert R_{s,\pm}\rangle ∣ R s , ± ⟩ , ⟨ L s , ± ∣ \langle L_{s,\pm}\rvert ⟨ L s , ± ∣ and frequencies ν ± \nu_\pm ν ± , and projecting the modulation H ± 1 = − i ( κ 0 m / 2 ) e ∓ i Ω t P H_{\pm1}=-i(\kappa_0m/2)e^{\mp i\Omega t}P H ± 1 = − i ( κ 0 m /2 ) e ∓ i Ω t P of H = i A H=iA H = i A onto them, g = ⟨ L s , + ∣ H 1 ∣ R s , − ⟩ g=\langle L_{s,+}\rvert H_1\lvert R_{s,-}\rangle g = ⟨ L s , + ∣ H 1 ∣ R s , − ⟩ , gives
K = ( Δ r e d g − g ∗ − Δ r e d ) , ω β = ω 0 + β Δ r e d 2 − ∣ g ∣ 2 , (44) K=\begin{pmatrix}\Delta_{\mathrm{red}} & g\\ -g^* & -\Delta_{\mathrm{red}}\end{pmatrix},\qquad\omega_\beta=\omega_0+\beta\sqrt{\Delta_{\mathrm{red}}^2-\lvert g\rvert^2}, \tag*{(44)} K = ( Δ red − g ∗ g − Δ red ) , ω β = ω 0 + β Δ red 2 − ∣ g ∣ 2 , (44)
with ω 0 = ( ν + + ν − + Ω ) / 2 \omega_0=(\nu_++\nu_-+\Omega)/2 ω 0 = ( ν + + ν − + Ω ) /2 and Δ r e d = ( ν + − ν − − Ω ) / 2 \Delta_{\mathrm{red}}=(\nu_+-\nu_--\Omega)/2 Δ red = ( ν + − ν − − Ω ) /2 . The form K 21 = − K 12 ∗ K_{21}=-K_{12}^* K 21 = − K 12 ∗ holds in the conjugate-paired gauge ∣ R s , − ⟩ = P ∣ R s , + ⟩ \lvert R_{s,-}\rangle=P\lvert R_{s,+}\rangle ∣ R s , − ⟩ = P ∣ R s , + ⟩ , ⟨ L s , − ∣ = ⟨ L s , + ∣ ∗ P \langle L_{s,-}\rvert=\langle L_{s,+}\rvert^*P ⟨ L s , − ∣ = ⟨ L s , + ∣ ∗ P , which also gives ν − = − ν + ∗ \nu_-=-\nu_+^* ν − = − ν + ∗ and real Δ r e d \Delta_{\mathrm{red}} Δ red . In this reduction the EP lies at k E P r e d = 0.5691 k_{\mathrm{EP}}^{\mathrm{red}}=0.5691 k EP red = 0.5691 with ρ r e d = − ∂ k ( Δ r e d 2 − ∣ g ∣ 2 ) = 0.0147 \rho_{\mathrm{red}}=-\partial_k(\Delta_{\mathrm{red}}^2-\lvert g\rvert^2)=0.0147 ρ red = − ∂ k ( Δ red 2 − ∣ g ∣ 2 ) = 0.0147 , distinct from the full values k E P ≃ 0.5744 k_{\mathrm{EP}}\simeq0.5744 k EP ≃ 0.5744 and ρ = 0.0158 \rho=0.0158 ρ = 0.0158 . With all EP quantities evaluated at k E P r e d k_{\mathrm{EP}}^{\mathrm{red}} k EP red and the reduced detuning d r e d = k E P r e d − k d_{\mathrm{red}}=k_{\mathrm{EP}}^{\mathrm{red}}-k d red = k EP red − k , reconstructing the two harmonics gives the reduced coefficient
C a r e d ( t ) = ρ r e d 2 ∣ g E P ∣ Im [ e − i φ E P ( R s , + ; a E P ) ∗ R s , − ; a E P e i Ω t ] ∣ R s , + ; a E P + e − i φ E P R s , − ; a E P e i Ω t ∣ 2 . (45) \mathcal{C}_a^{\mathrm{red}}(t)=\frac{\sqrt{\rho_{\mathrm{red}}}}{2\lvert g_{\mathrm{EP}}\rvert}\frac{\operatorname{Im}\left[e^{-i\varphi_{\mathrm{EP}}}\left(R_{s,+;a}^{\mathrm{EP}}\right)^*R_{s,-;a}^{\mathrm{EP}}e^{i\Omega t}\right]}{\left\lvert R_{s,+;a}^{\mathrm{EP}}+e^{-i\varphi_{\mathrm{EP}}}R_{s,-;a}^{\mathrm{EP}}e^{i\Omega t}\right\rvert^2}. \tag*{(45)} C a red ( t ) = 2 ∣ g EP ∣ ρ red R s , + ; a EP + e − i φ EP R s , − ; a EP e i Ω t 2 Im [ e − i φ EP ( R s , + ; a EP ) ∗ R s , − ; a EP e i Ω t ] . (45)
with g E P = ∣ g E P ∣ e i φ E P g_{\mathrm{EP}}=\lvert g_{\mathrm{EP}}\rvert e^{i\varphi_{\mathrm{EP}}} g EP = ∣ g EP ∣ e i φ EP . Because the same-harmonic term is real, it drops out, so the d − 1 / 2 d^{-1/2} d − 1/2 divergence comes from interference between the harmonics inherited from the positive- and negative-frequency static branches.
Squared separations and the energy-norm Petermann factor
Each two-sector partition A ∣ B A|B A ∣ B separates the spatial variance into internal sector variances σ A 2 \sigma_A^2 σ A 2 and σ B 2 \sigma_B^2 σ B 2 and a term set by the sector-centroid separation, σ z 2 = f A σ A 2 + f B σ B 2 + f A f B ( Δ Z A B ) 2 \sigma_z^2=f_A\sigma_A^2+f_B\sigma_B^2+f_Af_B(\Delta Z_{AB})^2 σ z 2 = f A σ A 2 + f B σ B 2 + f A f B ( Δ Z A B ) 2 . Modulation and damping select the partitions P ∣ P ˉ P|\bar{P} P ∣ P ˉ and Q ∣ Q ˉ Q|\bar{Q} Q ∣ Q ˉ , which overlap, so S s e p = ∑ a = P , Q f a f a ˉ ( Δ Z a a ˉ ) 2 S_{\mathrm{sep}}=\sum_{a=P,Q}f_a f_{\bar{a}}(\Delta Z_{a\bar{a}})^2 S sep = ∑ a = P , Q f a f a ˉ ( Δ Z a a ˉ ) 2 is a source-selected moment distinct from σ z 2 \sigma_z^2 σ z 2 . Locally on the passband, S s e p = ∑ a = P , Q L β , a 2 / ( f β , a f β , a ˉ ) S_{\mathrm{sep}}=\sum_{a=P,Q}\mathcal{L}_{\beta,a}^2/(f_{\beta,a}f_{\beta,\bar{a}}) S sep = ∑ a = P , Q L β , a 2 / ( f β , a f β , a ˉ ) , with zero assigned to an empty sector or complement. With q = β d q=\beta\sqrt{d} q = β d , the Cauchy–Schwarz inequality gives
0 ≤ d L β , a 2 f β , a f β , a ˉ ≤ ∑ j c j ∣ ∂ q ξ β , j ∣ 2 4 N β , (46) 0\leq d\frac{\mathcal{L}_{\beta,a}^2}{f_{\beta,a}f_{\beta,\bar{a}}}\leq\frac{\sum_j c_j|\partial_q\xi_{\beta,j}|^2}{4N_\beta}, \tag*{(46)} 0 ≤ d f β , a f β , a ˉ L β , a 2 ≤ 4 N β ∑ j c j ∣ ∂ q ξ β , j ∣ 2 , (46)
a bound uniform over the period. Together with the almost-everywhere limit
d S s e p → c s e p ( t ) = ∑ a = P , Q S a 2 f a E P f a ˉ E P = ∑ a = P , Q f a E P f a ˉ E P ( C a − C a ˉ ) 2 , (47) dS_{\mathrm{sep}}\to c_{\mathrm{sep}}(t)=\sum_{a=P,Q}\frac{S_a^2}{f_a^{\mathrm{EP}}f_{\bar{a}}^{\mathrm{EP}}}=\sum_{a=P,Q}f_a^{\mathrm{EP}}f_{\bar{a}}^{\mathrm{EP}}(C_a-C_{\bar{a}})^2, \tag*{(47)} d S sep → c sep ( t ) = a = P , Q ∑ f a EP f a ˉ EP S a 2 = a = P , Q ∑ f a EP f a ˉ EP ( C a − C a ˉ ) 2 , (47)
C s e p = ⟨ ( c s e p − ⟨ c s e p ⟩ T ) 2 ⟩ T 1 / 2 , C_{\mathrm{sep}}=\left\langle(c_{\mathrm{sep}}-\langle c_{\mathrm{sep}}\rangle_T)^2\right\rangle_T^{1/2}, C sep = ⟨ ( c sep − ⟨ c sep ⟩ T ) 2 ⟩ T 1/2 ,
where the second form holds away from nodes with C ˉ a = ∑ j ≠ a f j E P C j / f a ˉ E P \bar{C}_a=\sum_{j\neq a}f_j^{\mathrm{EP}}C_j/f_{\bar{a}}^{\mathrm{EP}} C ˉ a = ∑ j = a f j EP C j / f a ˉ EP , it gives convergence in cycle rms, so d A s e p → C s e p dA_{\mathrm{sep}}\to C_{\mathrm{sep}} d A sep → C sep . Its leading term is even in β \beta β , whereas the leading source velocity is odd. For the medium of the main text C s e p = 2.6 C_{\mathrm{sep}}=2.6 C sep = 2.6 .
The energy Gramian defines the norms in the Petermann factor, whose divergence comes from the vanishing overlap of the left and right eigenvectors at the EP. With g u ( t ) = diag ( ϵ b , μ , κ / α , 1 / α ) / 4 g_u(t)=\operatorname{diag}(\epsilon_b,\mu,\kappa/\alpha,1/\alpha)/4 g u ( t ) = diag ( ϵ b , μ , κ / α , 1/ α ) /4 , the period-averaged Gramian and, because the left eigenvector is a covector, the dual energy norm induced by G u − 1 G_u^{-1} G u − 1 act as
⟨ R β ∣ G u ∣ R β ⟩ = 1 T ∫ 0 T ξ β † g u ξ β d t , ⟨ L β ∣ G u − 1 ∣ L β ⟩ = 1 T ∫ 0 T ℓ β g u − 1 ℓ β † d t , (48) \langle R_\beta|G_u|R_\beta\rangle=\frac{1}{T}\int_0^T\xi_\beta^\dagger g_u\xi_\beta\,\mathrm{d}t,\qquad\langle L_\beta|G_u^{-1}|L_\beta\rangle=\frac{1}{T}\int_0^T\ell_\beta g_u^{-1}\ell_\beta^\dagger\,\mathrm{d}t, \tag*{(48)} ⟨ R β ∣ G u ∣ R β ⟩ = T 1 ∫ 0 T ξ β † g u ξ β d t , ⟨ L β ∣ G u − 1 ∣ L β ⟩ = T 1 ∫ 0 T ℓ β g u − 1 ℓ β † d t , (48)
since in the full harmonic space G u G_u G u represents multiplication by g u ( t ) g_u(t) g u ( t ) and G u − 1 G_u^{-1} G u − 1 multiplication by g u ( t ) − 1 g_u(t)^{-1} g u ( t ) − 1 . With M r β = ζ β r β Mr_\beta=\zeta_\beta r_\beta M r β = ζ β r β , l β M = ζ β l β l_\beta M=\zeta_\beta l_\beta l β M = ζ β l β and Λ β = T − 1 ln ζ β \Lambda_\beta=T^{-1}\ln\zeta_\beta Λ β = T − 1 ln ζ β , the periodic right and left modes are ξ β ( t ) = e − Λ β t U k ( t ) r β \xi_\beta(t)=e^{-\Lambda_\beta t}U_k(t)r_\beta ξ β ( t ) = e − Λ β t U k ( t ) r β and ℓ β ( t ) = e Λ β t l β U k ( t ) − 1 \ell_\beta(t)=e^{\Lambda_\beta t}l_\beta U_k(t)^{-1} ℓ β ( t ) = e Λ β t l β U k ( t ) − 1 , so ℓ β ( t ) ξ β ( t ) = l β r β \ell_\beta(t)\xi_\beta(t)=l_\beta r_\beta ℓ β ( t ) ξ β ( t ) = l β r β and the denominator of Eq. (10) is ∣ l β r β ∣ 2 |l_\beta r_\beta|^2 ∣ l β r β ∣ 2 .
Equation (S35) gives the time-domain representation in the full harmonic space. For Figure 2 (e), Eq. (10) is evaluated using the Sambe truncation specified in Sec. S1 and the inverse of the truncated energy Gramian. This inverse generally differs from the truncation of the full inverse. Rescaling either eigenvector leaves the factor K β ( u ) ≥ 1 K_\beta^{(u)}\geq1 K β ( u ) ≥ 1 of Eq. (10) unchanged. Near the EP, ∣ R β ⟩ = ∣ R 0 ⟩ + β d ∣ R 1 ⟩ |R_\beta\rangle=|R_0\rangle+\beta\sqrt{d}|R_1\rangle ∣ R β ⟩ = ∣ R 0 ⟩ + β d ∣ R 1 ⟩ and ⟨ L β ∣ = ⟨ L 0 ∣ + β d ⟨ L 1 ∣ \langle L_\beta|=\langle L_0|+\beta\sqrt{d}\langle L_1| ⟨ L β ∣ = ⟨ L 0 ∣ + β d ⟨ L 1 ∣ to first order in d \sqrt{d} d , with ⟨ L 0 ∣ R 0 ⟩ = 0 \langle L_0|R_0\rangle=0 ⟨ L 0 ∣ R 0 ⟩ = 0 , so ⟨ L β ∣ R β ⟩ = β B d + O ( d ) \langle L_\beta|R_\beta\rangle=\beta\mathcal{B}\sqrt{d}+O(d) ⟨ L β ∣ R β ⟩ = β B d + O ( d ) with B = ⟨ L 1 ∣ R 0 ⟩ + ⟨ L 0 ∣ R 1 ⟩ ≠ 0 \mathcal{B}=\langle L_1|R_0\rangle+\langle L_0|R_1\rangle\neq0 B = ⟨ L 1 ∣ R 0 ⟩ + ⟨ L 0 ∣ R 1 ⟩ = 0 , so
K β ( u ) = K 0 d + O ( d − 1 / 2 ) , K 0 = ⟨ R 0 ∣ G u ∣ R 0 ⟩ ⟨ L 0 ∣ G u − 1 ∣ L 0 ⟩ ∣ B ∣ 2 > 0 , (49) K_\beta^{(u)}=\frac{K_0}{d}+O(d^{-1/2}),\qquad K_0=\frac{\langle R_0|G_u|R_0\rangle\langle L_0|G_u^{-1}|L_0\rangle}{|\mathcal{B}|^2}>0, \tag*{(49)} K β ( u ) = d K 0 + O ( d − 1/2 ) , K 0 = ∣ B ∣ 2 ⟨ R 0 ∣ G u ∣ R 0 ⟩ ⟨ L 0 ∣ G u − 1 ∣ L 0 ⟩ > 0 , (49)
with K 0 = 0.0062 K_0=0.0062 K 0 = 0.0062 here. Consequently A t r / K β ( u ) → C s / K 0 A_{\mathrm{tr}}/\sqrt{K_\beta^{(u)}}\to C_s/\sqrt{K_0} A tr / K β ( u ) → C s / K 0 and A s e p / K β ( u ) → C s e p / K 0 A_{\mathrm{sep}}/K_\beta^{(u)}\to C_{\mathrm{sep}}/K_0 A sep / K β ( u ) → C sep / K 0 .
Finite bandwidth and gap migration
For a finite packet, the source velocity and trembling amplitude are
v s r c , β w p = κ ˙ κ f β , P w p f β , P ˉ w p Δ Z β , P P ˉ w p − 2 η f β , Q w p f β , Q ˉ w p Δ Z β , Q Q ˉ w p , v_{\mathrm{src},\beta}^{\mathrm{wp}}=\frac{\dot{\kappa}}{\kappa}f_{\beta,P}^{\mathrm{wp}}f_{\beta,\bar{P}}^{\mathrm{wp}}\Delta Z_{\beta,P\bar{P}}^{\mathrm{wp}}-2\eta f_{\beta,Q}^{\mathrm{wp}}f_{\beta,\bar{Q}}^{\mathrm{wp}}\Delta Z_{\beta,Q\bar{Q}}^{\mathrm{wp}}, v src , β wp = κ κ ˙ f β , P wp f β , P ˉ wp Δ Z β , P P ˉ wp − 2 η f β , Q wp f β , Q ˉ wp Δ Z β , Q Q ˉ wp ,
A t r w p = ⟨ ( v s r c , β w p − ⟨ v s r c , β w p ⟩ T ) 2 ⟩ T 1 / 2 , (50) A_{\mathrm{tr}}^{\mathrm{wp}}=\left\langle\left(v_{\mathrm{src},\beta}^{\mathrm{wp}}-\langle v_{\mathrm{src},\beta}^{\mathrm{wp}}\rangle_T\right)^2\right\rangle_T^{1/2}, \tag*{(50)} A tr wp = ⟨ ( v src , β wp − ⟨ v src , β wp ⟩ T ) 2 ⟩ T 1/2 , (50)
where the mean and rms are taken over one period, the first period unless stated otherwise. In the joint limit d 0 → 0 + d_0\to0^+ d 0 → 0 + , σ κ / d 0 → 0 \sigma_\kappa/d_0\to0 σ κ / d 0 → 0 over a fixed window and away from nodes, a passband packet has
Δ Z β , a a ˉ w p = β C a − C a ˉ d 0 + o ( d 0 − 1 / 2 ) , (51) \Delta Z_{\beta,a\bar{a}}^{\mathrm{wp}}=\beta\frac{C_a-C_{\bar{a}}}{\sqrt{d_0}}+o(d_0^{-1/2}), \tag*{(51)} Δ Z β , a a ˉ wp = β d 0 C a − C a ˉ + o ( d 0 − 1/2 ) , (51)
an additive error that also applies where C a = C a ˉ \mathcal{C}_a=\mathcal{C}_{\bar a} C a = C a ˉ .
For the Gaussian average, take a regular launch with the phase convention of Sec. S3 and the Gaussian weight w ( k ) ∝ exp [ − ( k − k 0 ) 2 / ( 2 σ k 2 ) ] w(k)\propto\exp[-(k-k_0)^2/(2\sigma_k^2)] w ( k ) ∝ exp [ − ( k − k 0 ) 2 / ( 2 σ k 2 )] , and write the packet moments with the normalized spectral-energy measure d m t ∝ w ( k ) N β ( k , t ) e 2 Γ β ( k ) t d k \mathrm{d}m_t\propto w(k)N_\beta(k,t)e^{2\Gamma_\beta(k)t}\,\mathrm{d}k d m t ∝ w ( k ) N β ( k , t ) e 2 Γ β ( k ) t d k , with N β N_\beta N β of Eq. (30) . At fixed λ \lambda λ , over the first period and within a few σ k \sigma_k σ k of k E P k_{\rm EP} k EP , the relative growth factor e 2 ( Γ β − Γ E P ) t e^{2(\Gamma_\beta-\Gamma_{\rm EP})t} e 2 ( Γ β − Γ EP ) t , the sector weights n β , a n_{\beta,a} n β , a and their sum N β N_\beta N β differ from 1 1 1 , n a E P n_a^{\rm EP} n a EP and N E P > 0 N_{\rm EP}>0 N EP > 0 by O ( σ k ) O(\sqrt{\sigma_k}) O ( σ k ) on both sides of the EP, so the absolute expansions of the sector weights stay valid at nodes. The weighted offset f a w p Δ Z a a ˉ w p f_a^{\rm wp}\Delta Z_{a\bar a}^{\rm wp} f a wp Δ Z a a ˉ wp equals the m t m_t m t -average of the local offset L β , a \mathcal{L}_{\beta,a} L β , a of Eq. (30) plus the m t m_t m t -covariance of f a f_a f a with the connection B z = A e n v + t ∂ k ω β , R + A u B_z=\mathcal{A}_{\rm env}+t\partial_k\omega_{\beta,\mathrm{R}}+\mathcal{A}_u B z = A env + t ∂ k ω β , R + A u , where A u = ∑ j J β , j / N β \mathcal{A}_u=\sum_j\mathcal{J}_{\beta,j}/N_\beta A u = ∑ j J β , j / N β is the energy-weighted connection of the full mode. Although B z B_z B z is singular at the EP, f a − f a E P = O ( ∣ k − k E P ∣ 1 / 2 ) f_a-f_a^{\rm EP}=O(|k-k_{\rm EP}|^{1/2}) f a − f a EP = O ( ∣ k − k EP ∣ 1/2 ) while the singular part of B z B_z B z is O ( ∣ k − k E P ∣ − 1 / 2 ) O(|k-k_{\rm EP}|^{-1/2}) O ( ∣ k − k EP ∣ − 1/2 ) , so the covariance is O ( 1 ) O(1) O ( 1 ) and subleading to the σ k − 1 / 2 \sigma_k^{-1/2} σ k − 1/2 waveform. Gaussian tails outside a fixed k k k -neighborhood of the EP are exponentially small as σ k → 0 \sigma_k\to0 σ k → 0 at fixed λ \lambda λ . With no local source from the gap side (Sec. S3) and the common waveform v t r ( k , t ) = d − 1 / 2 w t r ( t ) v_{\rm tr}(k,t)=d^{-1/2}w_{\rm tr}(t) v tr ( k , t ) = d − 1/2 w tr ( t ) on the passband, the packet waveform is
v t r w p ( t ) = w t r ( t ) ∫ 0 ∞ d d 2 π σ k e − ( d − d 0 ) 2 / ( 2 σ k 2 ) d + o ( σ k − 1 / 2 ) = w t r ( t ) σ k − 1 / 2 F ( λ ) + o ( σ k − 1 / 2 ) (52) \begin{aligned}
v_{\rm tr}^{\rm wp}(t)=w_{\rm tr}(t)\int_0^\infty\frac{\mathrm{d}d}{\sqrt{2\pi}\sigma_k}\frac{e^{-(d-d_0)^2/(2\sigma_k^2)}}{\sqrt{d}}+o(\sigma_k^{-1/2}) \\
&= w_{\rm tr}(t)\sigma_k^{-1/2}F(\lambda)+o(\sigma_k^{-1/2}) \tag*{(52)}
\end{aligned} v tr wp ( t ) = w tr ( t ) ∫ 0 ∞ 2 π σ k d d d e − ( d − d 0 ) 2 / ( 2 σ k 2 ) + o ( σ k − 1/2 ) = w tr ( t ) σ k − 1/2 F ( λ ) + o ( σ k − 1/2 ) (52)
in cycle rms, which gives Eq. (12) . Substituting u = v 2 u=v^2 u = v 2 for the integration variable u = d / σ k u=d/\sigma_k u = d / σ k of Eq. (12) removes the endpoint singularity. F ( 0 ) = 2 − 3 / 4 Γ ( 1 / 4 ) / 2 π = 0.860 F(0)=2^{-3/4}\Gamma(1/4)/\sqrt{2\pi}=0.860 F ( 0 ) = 2 − 3/4 Γ ( 1/4 ) / 2 π = 0.860 , F F F peaks at 1.02 1.02 1.02 for λ = 0.77 \lambda=0.77 λ = 0.77 , F = λ − 1 / 2 [ 1 + 3 / ( 8 λ 2 ) + O ( λ − 4 ) ] F=\lambda^{-1/2}[1+3/(8\lambda^2)+O(\lambda^{-4})] F = λ − 1/2 [ 1 + 3/ ( 8 λ 2 ) + O ( λ − 4 )] for λ ≫ 1 \lambda\gg1 λ ≫ 1 , and F ≃ e − λ 2 / 2 / 2 ∣ λ ∣ F\simeq e^{-\lambda^2/2}/\sqrt{2|\lambda|} F ≃ e − λ 2 /2 / 2∣ λ ∣ for λ → − ∞ \lambda\to-\infty λ → − ∞ . The average holds while d 0 d_0 d 0 and σ k \sigma_k σ k lie inside the square-root region: for σ k = 10 − 5 \sigma_k=10^{-5} σ k = 1 0 − 5 and 10 − 4 10^{-4} 1 0 − 4 in Figure 4 (a) the first-period amplitude departs from C s F ( λ ) / σ k C_sF(\lambda)/\sqrt{\sigma_k} C s F ( λ ) / σ k by about 5 % 5\% 5% at d 0 ≃ 5 × 10 − 4 d_0\simeq5\times10^{-4} d 0 ≃ 5 × 1 0 − 4 and 10 % 10\% 10% at d 0 ≃ 10 − 3 d_0\simeq10^{-3} d 0 ≃ 1 0 − 3 , and for σ k = 10 − 3 \sigma_k=10^{-3} σ k = 1 0 − 3 a finite-bandwidth correction is already present near d 0 = 0 d_0=0 d 0 = 0 .
At later times relative amplification moves spectral weight into the gap, where the branch with the larger growth rate has Γ ( h ) = Γ E P + ρ h + O ( h ) \Gamma(h)=\Gamma_{\rm EP}+\sqrt{\rho h}+O(h) Γ ( h ) = Γ EP + ρ h + O ( h ) with h = k − k E P h=k-k_{\rm EP} h = k − k EP and ρ \rho ρ of Eq. (38) . The common factor e 2 Γ E P t e^{2\Gamma_{\rm EP}t} e 2 Γ EP t cancels from the normalized spectrum, and with y = h / σ k y=h/\sigma_k y = h / σ k and τ = t ρ σ k \tau=t\sqrt{\rho\sigma_k} τ = t ρ σ k the leading exponent is
Φ + ( y ; λ , τ ) = − 1 2 ( y + λ ) 2 + 2 τ y , (53) \Phi_+(y;\lambda,\tau)=-\frac{1}{2}(y+\lambda)^2+2\tau\sqrt{y}, \tag*{(53)} Φ + ( y ; λ , τ ) = − 2 1 ( y + λ ) 2 + 2 τ y , (53)
whose interior stationary point gives Eq. (13) and h s = σ k y s h_s=\sigma_k y_s h s = σ k y s . On the opposite gap branch, however, the square-root term changes sign, so for λ ≥ 0 \lambda\ge0 λ ≥ 0 no interior saddle forms. The saddle describes the spectral maximum once its contribution dominates the passband weight, h s h_s h s lies inside the gap-side square-root region, and the O ( h ) O(h) O ( h ) remainder of Γ \Gamma Γ times t t t stays small, so the relevant spectral distance from the EP is d 0 d_0 d 0 for a resolved passband packet, σ k \sigma_k σ k for a packet overlapping the EP and h s h_s h s once the gap saddle dominates. For this medium Γ E P = − 2.6 × 10 − 3 Ω \Gamma_{\rm EP}=-2.6\times10^{-3}\Omega Γ EP = − 2.6 × 1 0 − 3 Ω , so relative growth reshapes the normalized spectrum while the absolute energy decays. The leading local estimate for the onset of net growth on the gap side is h ≃ Γ E P 2 / ρ ≃ 4 × 10 − 4 h\simeq\Gamma_{\rm EP}^2/\rho\simeq4\times10^{-4} h ≃ Γ EP 2 / ρ ≃ 4 × 1 0 − 4 , and the energy minimum of Figure 4 (c) occurs near the stage when the saddle reaches this scale.
Numerical propagation
We use fourth-order Runge–Kutta integration to propagate x = ( E , H , P , Q ) T \boldsymbol{x}=(E,H,P,Q)^{\mathsf{T}} x = ( E , H , P , Q ) T and its spectral derivative on the transformed grid k = k E P + ∣ ϑ ∣ k=k_{\rm EP}+|\vartheta| k = k EP + ∣ ϑ ∣ . The derivative ∂ ϑ x = ( d k / d ϑ ) ∂ k x \partial_\vartheta\boldsymbol{x}=(\mathrm{d}k/\mathrm{d}\vartheta)\partial_k\boldsymbol{x} ∂ ϑ x = ( d k / d ϑ ) ∂ k x obeys
∂ t ( ∂ ϑ x ) = A ( k , t ) ∂ ϑ x + d k d ϑ ∂ k A ( k , t ) x , \partial_t(\partial_\vartheta\boldsymbol{x})=A(k,t)\,\partial_\vartheta\boldsymbol{x}+\frac{\mathrm{d}k}{\mathrm{d}\vartheta}\partial_kA(k,t)\boldsymbol{x}, ∂ t ( ∂ ϑ x ) = A ( k , t ) ∂ ϑ x + d ϑ d k ∂ k A ( k , t ) x ,
d k d ϑ = 2 ∣ ϑ ∣ . (54) \frac{\mathrm{d}k}{\mathrm{d}\vartheta}=2|\vartheta|. \tag*{(54)} d ϑ d k = 2∣ ϑ ∣. (54)
Sector energies and centroids are evaluated using Eq. (25) , with trapezoidal quadrature in ϑ \vartheta ϑ for Figure 2 (d) and midpoint quadrature for Figure 4 . In the latter, the EP is a cell boundary and is not sampled directly. Initial spectral derivatives are obtained by finite differences for Figure 2 (d) and by differentiating the monodromy eigenproblem for Figure 4 . Long-time evolution is evaluated from the one-period Floquet data.
Each eigen-packet uses a real Gaussian envelope with spectral-intensity width σ k \sigma_k σ k , multiplying ξ β ( k , 0 ) \xi_\beta(k,0) ξ β ( k , 0 ) normalized to unit instantaneous energy with its electric component real and positive. A real envelope contributes only its weight to the energy and first-moment integrals. On the gap side, the β = − \beta=- β = − passband branch is continued into the branch with the larger growth rate.
The medium has ϵ b = μ = Ω = 1 \epsilon_b = \mu= \Omega= 1 ϵ b = μ = Ω = 1 , α = 0.6 \alpha= 0.6 α = 0.6 , κ 0 = 2 \kappa_0 = 2 κ 0 = 2 , m = 0.25 m = 0.25 m = 0.25 , and η = 0.08 \eta= 0.08 η = 0.08 , with k E P = 0.5743977113 k_{\mathrm{EP}} = 0.5743977113 k EP = 0.5743977113 determined from the discriminant of the resonant multiplier pair. The trajectories of Figure 2 (d) use σ k = d 0 / 20 \sigma_k = d_0/20 σ k = d 0 /20 , 801 nodes over k 0 ± 6 σ k k_0 \pm6\sigma_k k 0 ± 6 σ k , and a time step T / 800 T/800 T /800 over four periods. With Z d e t ( t ) = Z ( t ) − v ˉ t Z_{\mathrm{det}}(t) = Z(t) - \bar{v}t Z det ( t ) = Z ( t ) − v ˉ t , v ˉ = [ Z ( 4 T ) − Z ( T ) ] / ( 3 T ) \bar{v} = [Z(4T) - Z(T)]/(3T) v ˉ = [ Z ( 4 T ) − Z ( T )] / ( 3 T ) , and I = [ T , 4 T ] I = [T,4T] I = [ T , 4 T ] , the plotted displacement is δ Z = Z d e t − ⟨ Z d e t ⟩ I \delta Z = Z_{\mathrm{det}} - \langle Z_{\mathrm{det}}\rangle_I δ Z = Z det − ⟨ Z det ⟩ I . The flux contribution ∫ 0 t v f l u x ( t ′ ) d t ′ \int_0^t v_{\mathrm{flux}}(t')\,\mathrm{d}t' ∫ 0 t v flux ( t ′ ) d t ′ is detrended and centered over the same interval. The calculations of Figure 4 use a time step T / 12800 T/12800 T /12800 for the one-period propagation; the long-time packet in panels (b,c) is integrated over k E P − 9 σ k ≤ k ≤ k E P + 16 σ k k_{\mathrm{EP}} - 9\sigma_k \leq k \leq k_{\mathrm{EP}} + 16\sigma_k k EP − 9 σ k ≤ k ≤ k EP + 16 σ k .
The coefficients of Figure 3 come from the local source velocity, the band slope and the time-domain Petermann factor at d = 10 − 4 d = 10^{-4} d = 1 0 − 4 , 10 − 5 10^{-5} 1 0 − 5 and 10 − 6 10^{-6} 1 0 − 6 from the EP of each loss profile, with a time step T / 2000 T/2000 T /2000 , extrapolated to d → 0 d \to0 d → 0 by a quadratic fit in d d d . The field-only packets of Figure 5 use 801 nodes over k 0 ± 6 σ k k_0 \pm6\sigma_k k 0 ± 6 σ k , the one-period monodromy from a time step T / 2000 T/2000 T /2000 , and the centered rms of v s r c v_{\mathrm{src}} v src , sampled within each period at T / 400 T/400 T /400 and displayed every third period in Figure 5 (a); the displacement of Figure 5 (d) is detrended by a least-squares straight line over that period, which leaves a zero-mean residual. The monochromatic calculation uses the monodromy at a single wavenumber from a time step T / 4000 T/4000 T /4000 . For both monochromatic and packet launches, A b A_b A b is the largest one-period centered rms of v s r c v_{\mathrm{src}} v src over periods I n = [ n T , ( n + 1 ) T ] I_n = [nT,(n+1)T] I n = [ n T , ( n + 1 ) T ] whose midpoints lie in [ 0.7 t b , 1.3 t b ] [0.7t_b,1.3t_b] [ 0.7 t b , 1.3 t b ] . The packet calculations use 801 nodes at d 0 = 10 − 3 d_0 = 10^{-3} d 0 = 1 0 − 3 and 1601 nodes at smaller d 0 d_0 d 0 , monodromies from a time step T / 2000 T/2000 T /2000 (T / 4000 T/4000 T /4000 at d 0 = 10 − 6 d_0 = 10^{-6} d 0 = 1 0 − 6 ), and sampling at T / 400 T/400 T /400 or T / 800 T/800 T /800 within the period. The late-time regularized comparison of Sec. S9 uses 6000 nodes on the transformed grid over k E P − 9 σ k ≤ k ≤ k E P + 16 σ k k_{\mathrm{EP}} - 9\sigma_k \leq k \leq k_{\mathrm{EP}} + 16\sigma_k k EP − 9 σ k ≤ k ≤ k EP + 16 σ k with the monodromy from a time step T / 4000 T/4000 T /4000 .
Medium dependence
Modulation and damping determine C s C_s C s through the rms amplitudes and correlation of their leading source waveforms. With w P , Q ( t ) = lim d → 0 + d [ v P , Q − ⟨ v P , Q ⟩ T ] w_{P,Q}(t) = \lim_{d\to0^+} \sqrt{d}\,[v_{P,Q} - \langle v_{P,Q}\rangle_T] w P , Q ( t ) = lim d → 0 + d [ v P , Q − ⟨ v P , Q ⟩ T ] , their rms values C P C_P C P and C Q C_Q C Q over one period and, when both are nonzero, their correlation coefficient ϱ \varrho ϱ , the total coefficient obeys
C s 2 = C P 2 + C Q 2 + 2 ϱ C P C Q . (55) C_s^2 = C_P^2 + C_Q^2 + 2\varrho C_P C_Q. \tag*{(55)} C s 2 = C P 2 + C Q 2 + 2 ϱ C P C Q . (55)
Without damping the modulation channel alone acts and the chain coefficients of Eq. (42) give C P ≃ 0.3 m C_P \simeq0.3\,m C P ≃ 0.3 m for modulation depths between 0.1 and 0.4. Damping acts through the kinetic-energy sector and, in the media we computed, raises C Q C_Q C Q faster than it changes C P C_P C P and adds the two channels with positive correlation, so that at m = 0.25 m = 0.25 m = 0.25 and η = 0.5 \eta= 0.5 η = 0.5 the damping channel (C Q = 0.093 C_Q = 0.093 C Q = 0.093 ) exceeds the modulation channel (C P = 0.058 C_P = 0.058 C P = 0.058 ). For m = 0.25 m = 0.25 m = 0.25 , a modulated loss rate η ( t ) = η 0 [ 1 + δ cos ( Ω t + φ ) ] \eta(t) = \eta_0[1 + \delta\cos(\Omega t + \varphi)] η ( t ) = η 0 [ 1 + δ cos ( Ω t + φ )] with 0 ≤ δ ≤ 1 0 \leq\delta\leq1 0 ≤ δ ≤ 1 enters through the source term − 2 η ( t ) u Q -2\eta(t)u_Q − 2 η ( t ) u Q and the Floquet modes, and each loss profile has its own EP. At each EP we compare the trembling coefficient C s C_s C s with the band-slope coefficient and the Petermann coefficient,
C g = lim d → 0 + d ∣ ∂ k Re ω − ∣ , K 0 = lim d → 0 + d K − ( u ) . (56) C_g = \lim_{d\to0^+} \sqrt{d}\,|\partial_k \operatorname{Re}\omega_-|,\qquad\mathcal{K}_0 = \lim_{d\to0^+} d K_-^{(u)}. \tag*{(56)} C g = d → 0 + lim d ∣ ∂ k Re ω − ∣ , K 0 = d → 0 + lim d K − ( u ) . (56)
with C s C_s C s from the local source velocity of Eq. (31) and K − ( u ) K_-^{(u)} K − ( u ) from the time-domain representation of Eq. (48) . Measured from the restoring coefficient, φ = 0 \varphi= 0 φ = 0 and π \pi π give loss in phase and in antiphase with κ ( t ) \kappa(t) κ ( t ) , as can occur at linear order when a common control modulates both coefficients with negligible relative response lag, and φ = π / 2 \varphi= \pi/2 φ = π /2 and 3 π / 2 3\pi/2 3 π /2 give loss in quadrature. For the medium of Figure 2 , η 0 = 0.08 \eta_0 = 0.08 η 0 = 0.08 , in-phase and antiphase loss change C s C_s C s monotonically with δ \delta δ , lowering it by about 6% and raising it by about 11% at δ = 1 \delta= 1 δ = 1 , while C g C_g C g and K 0 \mathcal{K}_0 K 0 barely change [Figure 3 (a)]. Quadrature loss at this damping moves all three coefficients by comparable fractions. Single-branch packets centered d 0 = 10 − 4 d_0 = 10^{-4} d 0 = 1 0 − 4 below each profile’s EP, with σ k = d 0 / 20 \sigma_k = d_0/20 σ k = d 0 /20 , have first-period trembling amplitudes 11% higher for full-depth antiphase loss and 6% lower for in-phase loss than under constant loss. Their centroid displacements, detrended over [ T , 4 T ] [T,4T] [ T , 4 T ] as in Sec. S7, change in centered rms by + 10 % +10\% + 10% and − 5 % -5\% − 5% for Z Z Z , respectively, and by less than 1% for Z E M Z_{\mathrm{EM}} Z EM . At η 0 = 0.3 \eta_0 = 0.3 η 0 = 0.3 and δ = 1 \delta= 1 δ = 1 the three coefficients follow different curves as the loss phase varies [Figure 3 (b)]. Near φ = π \varphi= \pi φ = π the trembling coefficient exceeds its constant-loss value by about a quarter while C g C_g C g and K 0 \mathcal{K}_0 K 0 return close to their constant-loss values, and near φ = 3 π / 2 \varphi= 3\pi/2 φ = 3 π /2 the trembling coefficient returns close to its constant-loss value while C g C_g C g and K 0 \mathcal{K}_0 K 0 fall by about a fifth and a third.
Over the same media, with η \eta η up to 0.5 at m = 0.25 m = 0.25 m = 0.25 and up to 0.3 at m = 0.1 m = 0.1 m = 0.1 and 0.4, the single-branch amplitude A t r ( d ) A_{\mathrm{tr}}(d) A tr ( d ) follows the inverse-square-root scaling of Eq. (9) between d = 10 − 5 d = 10^{-5} d = 1 0 − 5 and 10 − 4 10^{-4} 1 0 − 4 .
Figure S1. Finite-variance launch with σ k = 10 − 4 \sigma_k=10^{-4} σ k = 1 0 − 4 . (a) Relative change of A t r w p σ k A_{\mathrm{tr}}^{\mathrm{wp}}\sqrt{\sigma_k} A tr wp σ k over the first period against the smoothing width ε k / σ k \varepsilon_k/\sigma_k ε k / σ k , for λ = 0 \lambda=0 λ = 0 , 1 1 1 and 2 2 2 . (b) Spatial rms width σ z \sigma_z σ z at launch against the number N k N_k N k of points on the grid k = k E P + ϑ ∣ ϑ ∣ k=k_{\mathrm{EP}}+\vartheta|\vartheta| k = k EP + ϑ ∣ ϑ ∣ , for the ideal launch and for ε k / σ k = 0.01 \varepsilon_k/\sigma_k=0.01 ε k / σ k = 0.01 and 0.1 0.1 0.1 (λ = 0 \lambda=0 λ = 0 ). Symbols mark computed points.
Finite-variance launch
We implement the regularization of Sec. S3 by replacing the launch profile ξ − ( k , 0 ) \xi_-(k,0) ξ − ( k , 0 ) on ∣ k − k E P ∣ < ε k / 2 |k-k_{\mathrm{EP}}|<\varepsilon_k/2 ∣ k − k EP ∣ < ε k /2 with the linear interpolation between its values at the two ends of that interval, and we propagate the smoothed packets with the full Maxwell–Lorentz equations. This interpolant is continuous and adds no rapid phase. Inside the interval, however, it generally does not keep the gap-side reality of Sec. S3, which holds outside it, and this local change enters the moments at the order estimated there.
For σ k = 10 − 4 \sigma_k=10^{-4} σ k = 1 0 − 4 and ε k ≤ 0.1 σ k \varepsilon_k\leq0.1\sigma_k ε k ≤ 0.1 σ k the scaled amplitude A t r w p σ k A_{\mathrm{tr}}^{\mathrm{wp}}\sqrt{\sigma_k} A tr wp σ k changes by less than 0.15% [Fig. S1(a)], and the smoothed packets keep max t ∣ v f l u x ∣ ≤ 0.17 c b \max_t|v_{\mathrm{flux}}|\leq0.17c_b max t ∣ v flux ∣ ≤ 0.17 c b . For the ideal launch, σ z 2 \sigma_z^2 σ z 2 grows with the logarithm of the number of k k k points as the integral ∫ ∣ ∂ k ξ ∣ 2 d k \int|\partial_k\xi|^2\,\mathrm{d}k ∫ ∣ ∂ k ξ ∣ 2 d k of Sec. S3 is resolved, whereas the smoothed launches have a fixed width σ z ≃ 5.0 × 10 3 c b / Ω \sigma_z\simeq5.0\times10^3c_b/\Omega σ z ≃ 5.0 × 1 0 3 c b /Ω [Fig. S1(b)]. For σ k = 10 − 4 \sigma_k=10^{-4} σ k = 1 0 − 4 and λ = 2 \lambda=2 λ = 2 , the smoothed launches with ε k = 0.01 σ k \varepsilon_k=0.01\sigma_k ε k = 0.01 σ k and 0.1 σ k 0.1\sigma_k 0.1 σ k , propagated with powers of the monodromy as in Sec. S11, follow the decay of the ideal launch through τ = 10 \tau=10 τ = 10 , where all three scaled amplitudes are of order 10 − 9 10^{-9} 1 0 − 9 .
Field-energy and electric-energy centroids
Measurements of E E E and H H H give access to the centroid of the field energy,
Z E M = ∫ z ( u E + u H ) d z ∫ ( u E + u H ) d z , (57) Z_{\mathrm{EM}}=\frac{\int z\left(u_E+u_H\right)\,\mathrm{d}z}{\int\left(u_E+u_H\right)\,\mathrm{d}z}, \tag*{(57)} Z EM = ∫ ( u E + u H ) d z ∫ z ( u E + u H ) d z , (57)
and of the electric energy alone, Z E Z_E Z E . For the packets of Figure 2 (d), detrended and centered as in Sec. S7, Z E M Z_{\mathrm{EM}} Z EM has a centered rms displacement about eight times that of Z Z Z , with a different waveform, and multiplying both displacements by d 0 \sqrt{d_0} d 0 makes the two detunings nearly coincide [Fig. S2(a)]. The flux contribution to the displacement has an rms of at most a few percent of the total. The electric fraction U E / U U_E/U U E / U of the single-branch mode dips within each cycle, and the dip deepens as the EP is approached [Fig. S2(b)]. Near the dip the small denominator makes Z E Z_E Z E sensitive to the electric-sector first moment, and between the two evaluated detunings its displacement increases more strongly than those of Z Z Z and Z E M Z_{\mathrm{EM}} Z EM .
Figure S2. (a) Detrended, centered displacements δ Z \delta Z δ Z of the total-energy centroid Z Z Z and the field-energy centroid Z E M Z_{\mathrm{EM}} Z EM , multiplied by d 0 \sqrt{d_0} d 0 (units ( c b / Ω ) 1 / 2 (c_b/\Omega)^{1/2} ( c b /Ω ) 1/2 ), for single-branch packets with σ k = d 0 / 20 \sigma_k=d_0/20 σ k = d 0 /20 at d 0 = 10 − 3 d_0=10^{-3} d 0 = 1 0 − 3 and 10 − 4 10^{-4} 1 0 − 4 . (b) Electric-energy fraction U E / U U_E/U U E / U of the single-branch mode over one period at d = 10 − 3 d=10^{-3} d = 1 0 − 3 and 10 − 4 10^{-4} 1 0 − 4 . The inset enlarges the dip.
Spectral chirp
Multiplying the launch by e i ϕ c h ( k ) e^{i\phi_{\mathrm{ch}}(k)} e i ϕ ch ( k ) replaces i ∂ k F a i\partial_kF_a i ∂ k F a by e i ϕ c h ( i ∂ k F a − ∂ k ϕ c h F a ) e^{i\phi_{\mathrm{ch}}}(i\partial_kF_a-\partial_k\phi_{\mathrm{ch}}F_a) e i ϕ ch ( i ∂ k F a − ∂ k ϕ ch F a ) in Eq. (25) , so each sector centroid shifts as
Z a → Z a − ⟨ ∂ k ϕ c h ⟩ a , ⟨ ∂ k ϕ c h ⟩ a = ∫ ∂ k ϕ c h ∣ F a ∣ 2 d k ∫ ∣ F a ∣ 2 d k , (58) Z_a \to Z_a-\langle\partial_k\phi_{\mathrm{ch}}\rangle_a,\qquad\langle\partial_k\phi_{\mathrm{ch}}\rangle_a=\frac{\int\partial_k\phi_{\mathrm{ch}}\lvert F_a\rvert^2\,\mathrm{d}k}{\int\lvert F_a\rvert^2\,\mathrm{d}k}, \tag*{(58)} Z a → Z a − ⟨ ∂ k ϕ ch ⟩ a , ⟨ ∂ k ϕ ch ⟩ a = ∫ ∣ F a ∣ 2 d k ∫ ∂ k ϕ ch ∣ F a ∣ 2 d k , (58)
and its offset from the total centroid changes by − ⟨ ∂ k ϕ c h ⟩ a + ⟨ ∂ k ϕ c h ⟩ t o t -\langle\partial_k\phi_{\mathrm{ch}}\rangle_a+\langle\partial_k\phi_{\mathrm{ch}}\rangle_{\mathrm{tot}} − ⟨ ∂ k ϕ ch ⟩ a + ⟨ ∂ k ϕ ch ⟩ tot . A constant phase leaves every centroid unchanged and a linear phase shifts all of them equally. A nonlinear phase, however, can give unequal shifts when the sector weights depend differently on k k k , and can lift the gap-side null of Sec. S3.
At fixed U U U and f a f_a f a the source velocity is linear in the sector first moments J a = U a Z a J_a=U_aZ_a J a = U a Z a ,
v s r c = W t [ J ] ≡ 1 U [ κ ˙ κ ( J P − f P ∑ b J b ) − 2 η ( J Q − f Q ∑ b J b ) ] , (59) v_{\mathrm{src}}=\mathcal{W}_t[J]\equiv\frac{1}{U}\left[\frac{\dot{\kappa}}{\kappa}\left(J_P-f_P\sum_bJ_b\right)-2\eta\left(J_Q-f_Q\sum_bJ_b\right)\right], \tag*{(59)} v src = W t [ J ] ≡ U 1 [ κ κ ˙ ( J P − f P b ∑ J b ) − 2 η ( J Q − f Q b ∑ J b ) ] , (59)
so any split of the J a J_a J a into parts gives source waveforms that add, while their centered rms values do not. Restricting the k k k integrals of J a J_a J a to k > k E P k>k_{\mathrm{EP}} k > k EP in Eq. (59) gives the gap-side part.
For the quadratic chirp ϕ c h = ( k − k 0 ) 2 / ( 4 σ k 2 ) \phi_{\mathrm{ch}}=(k-k_0)^2/(4\sigma_k^2) ϕ ch = ( k − k 0 ) 2 / ( 4 σ k 2 ) with σ k = 10 − 4 \sigma_k=10^{-4} σ k = 1 0 − 4 and λ = 2 \lambda=2 λ = 2 , propagated with powers of the monodromy, the gap-side first-moment contribution approaches the full chirped response at the later computed times, and that response decreases slowly up to τ = 20 \tau=20 τ = 20 [Figure S3 ]. Without the chirp, the same launch fades [Figure S3 ].
Figure S3. Preparation dependence of the trembling. (a) A t r w p σ k A_{\mathrm{tr}}^{\mathrm{wp}}\sqrt{\sigma_k} A tr wp σ k over the period nearest each τ \tau τ for the real-envelope launch and for the quadratic chirp, with the gap-side part of the chirped response, for σ k = 10 − 4 \sigma_k = 10^{-4} σ k = 1 0 − 4 and λ = 2 \lambda= 2 λ = 2 . The plotted range ends at A t r w p σ k = 10 − 4 A_{\mathrm{tr}}^{\mathrm{wp}}\sqrt{\sigma_k} = 10^{-4} A tr wp σ k = 1 0 − 4 , and the real-envelope response continues below it. (b) Energy of the field-only packet divided by that of a single-branch packet with the same envelope, both normalized at launch, for d 0 = 10 − 3 d_0 = 10^{-3} d 0 = 1 0 − 3 and σ k = d 0 / 20 \sigma_k = d_0/20 σ k = d 0 /20 ; vertical dashed line, t b / T t_b/T t b / T of Eq. (61) . (c) Branch-diagonal and inter-branch source waveforms of the field-only packet over the first period at d 0 = 10 − 4 d_0 = 10^{-4} d 0 = 1 0 − 4 (upper axis) and their sum on an expanded scale (lower axis). Parameters as in Figure 2 .
Launch in the fields alone
The launch x 0 = ( 1 , 1 , 0 , 0 ) T x_0=(1,1,0,0)^{\mathrm{T}} x 0 = ( 1 , 1 , 0 , 0 ) T is a forward field pulse with the polarization at rest. Expanding it in the right Floquet eigenvectors r j r_j r j of the monodromy defines the diagonal energies
x 0 = ∑ j c j r j , e j = ∣ c j ∣ 2 r j † g u ( 0 ) r j x 0 † g u ( 0 ) x 0 , (60) x_0=\sum_jc_jr_j,\qquad e_j=\frac{\lvert c_j\rvert^2r_j^\dagger g_u(0)r_j}{x_0^\dagger g_u(0)x_0}, \tag*{(60)} x 0 = j ∑ c j r j , e j = x 0 † g u ( 0 ) x 0 ∣ c j ∣ 2 r j † g u ( 0 ) r j , (60)
and the interference energy 1 − ∑ j e j 1-\sum_je_j 1 − ∑ j e j . The diagonal energies of the resonant pair grow as 1 / d 1/d 1/ d and reach about 30 at d = 10 − 4 d=10^{-4} d = 1 0 − 4 , and the negative interference energy cancels the excess in their sum, returning the normalized total to one. Splitting J a J_a J a into branch-diagonal and inter-branch parts in Eq. (59) gives two source waveforms of nearly equal and opposite shape, which cancel to a small remainder [Figure S3 ].
The two critical branches split as Re ω ± ≃ Re ω E P ± 2 C g d \operatorname{Re}\omega_{\pm} \simeq\operatorname{Re}\omega_{\mathrm{EP}} \pm2C_g\sqrt{d} Re ω ± ≃ Re ω EP ± 2 C g d , so their relative phase returns after
t b = 2 π ∣ Re ω + − Re ω − ∣ ≃ π 2 C g d , (61) t_b = \frac{2\pi}{\left|\operatorname{Re}\omega_{+} - \operatorname{Re}\omega_{-}\right|} \simeq\frac{\pi}{2C_g\sqrt{d}}, \tag*{(61)} t b = ∣ Re ω + − Re ω − ∣ 2 π ≃ 2 C g d π , (61)
while the relative group velocity 2 C g / d 2C_g/\sqrt{d} 2 C g / d separates the two branch components by about π / d \pi/d π / d . Near half a beat the resonant components add, and to leading order the energy relative to a single-branch packet, both normalized at launch, approaches ( e + + e − ) 2 ≃ 0.013 / d (\sqrt{e_{+}}+\sqrt{e_{-}})^2 \simeq0.013/d ( e + + e − ) 2 ≃ 0.013/ d Figure S3 . Near t b t_b t b the energy returns to its interference minimum, the components no longer cancel in the first moments, and the trembling bursts Figure 5 (a).
In the monochromatic limit we propagate x 0 x_0 x 0 at a single wavenumber with powers of the monodromy and evaluate the burst amplitude using the common maximum-rms definition of Sec. S7. The computed burst amplitudes follow
A b ≃ C b d − 3 / 2 , C b ≃ 0.0031 , (62) A_b \simeq C_b d^{-3/2}, \qquad C_b \simeq0.0031, \tag*{(62)} A b ≃ C b d − 3/2 , C b ≃ 0.0031 , (62)
over 10 − 6 ≤ d ≤ 10 − 3 10^{-6} \le d \le10^{-3} 1 0 − 6 ≤ d ≤ 1 0 − 3 Figure 5 (b). We write the state as a sum of the critical pair and a noncritical part. The noncritical multipliers have modulus 0.79 per period against 0.98 for the critical pair, so their contribution is negligible at t b t_b t b . The critical amplitudes, of order d − 1 / 2 d^{-1/2} d − 1/2 and opposite sign, cancel at launch and again near the recurrence, where the relative phase returns, and the decay-removed energy at the recurrence minimum stays between 0.95 and 0.98 of the launch energy. The wavenumber derivative of the accumulated relative phase reaches
t b ∣ ∂ k ( Re ω + − Re ω − ) ∣ ≃ π d (63) t_b \left|\partial_k\left(\operatorname{Re}\omega_{+} - \operatorname{Re}\omega_{-}\right)\right| \simeq\frac{\pi}{d} \tag*{(63)} t b ∣ ∂ k ( Re ω + − Re ω − ) ∣ ≃ d π (63)
at fixed observation time near t b t_b t b . Multiplied by branch amplitudes of order d − 1 / 2 d^{-1/2} d − 1/2 , this phase gradient therefore permits a d − 3 / 2 d^{-3/2} d − 3/2 contribution to the first moments and the source velocity when the decay-removed recurrence state remains finite.
For d 0 = 10 − 3 d_0 = 10^{-3} d 0 = 1 0 − 3 Figure 5 (c,Figure 5 d) the burst peaks about 123 periods after launch and lasts about 20 periods, when the pulse retains about 2% of its launch energy, close to the single-branch packet at the same time, and the detrended centroid swings by about 150 c b / Ω c_b/\Omega c b /Ω peak to peak within the burst period, with a largest excursion of about 90 c b / Ω c_b/\Omega c b /Ω , near 1% of the initial rms width 1 / ( 2 σ k ) = 10 4 c b / Ω 1/(2\sigma_k) = 10^4c_b/\Omega 1/ ( 2 σ k ) = 1 0 4 c b /Ω . For the monochromatic recurrence, whose decay-removed energy stays of order one, the absolute energy at the recurrence is about e 2 Γ E P t b e^{2\Gamma_{\mathrm{EP}}t_b} e 2 Γ EP t b , which falls from 2 × 10 − 2 2 \times10^{-2} 2 × 1 0 − 2 at d = 10 − 3 d = 10^{-3} d = 1 0 − 3 to 2 × 10 − 6 2 \times10^{-6} 2 × 1 0 − 6 at 10 − 4 10^{-4} 1 0 − 4 .
For a packet the beat phase varies across the spectrum by ∣ ∂ k ( Re ω + − Re ω − ) ∣ t b σ k = π σ k / d 0 \left|\partial_k(\operatorname{Re}\omega_{+} - \operatorname{Re}\omega_{-})\right|t_b\sigma_k = \pi\sigma_k/d_0 ∣ ∂ k ( Re ω + − Re ω − ) ∣ t b σ k = π σ k / d 0 at t b t_b t b . To leading order, neglecting the monochromatic baseline offset of the launch-normalized energy ratio from one, this raises the energy ratio at the recurrence minimum above one by
Δ R ≃ ( π σ k 2 d 0 ) 2 ( e + + e − ) 2 ≃ 0.031 σ k 2 d 0 3 , (64) \Delta R \simeq\left(\frac{\pi\sigma_k}{2d_0}\right)^2\left(\sqrt{e_{+}}+\sqrt{e_{-}}\right)^2 \simeq0.031\frac{\sigma_k^2}{d_0^3}, \tag*{(64)} Δ R ≃ ( 2 d 0 π σ k ) 2 ( e + + e − ) 2 ≃ 0.031 d 0 3 σ k 2 , (64)
which the computed minima approach as d 0 d_0 d 0 decreases Figure 5 (b), inset. Approaching the monochromatic burst thus requires Δ R ≪ 1 \Delta R \ll1 Δ R ≪ 1 , and the packet bursts fall increasingly below the monochromatic values as this excess becomes appreciable Figure 5 (b). A launch that is real in the frame ( E , − i H , P , Q ) (E,-iH,P,Q) ( E , − i H , P , Q ) , such as equal forward and backward field pulses with a real envelope and zero relative phase, stays real and gives no source velocity at any time.