Superluminal and Slow Light in Anomalous Dispersion

A single monochromatic plane wave carries no news: it is the same pure tone forever. To send a signal you must modulate — stack a band of frequencies whose interference builds a localised bump, an envelope, and it is the envelope that moves at the group velocity $v_g = \mathrm{d}\omega/\mathrm{d}k$ . Far from any resonance this envelope glides a little slower than light and behaves as intuition demands. The strange country opens near a resonance, where the refractive index climbs and plunges across a linewidth so steeply that the group index $n_g = n + \omega\,\mathrm{d}n/\mathrm{d}\omega$ stops resembling $n$ at all. In one narrow window it swells to $10^{7}$ and light crawls at bicycle speed; a linewidth away it dips below one, through zero, and goes negative, so the smooth peak of a pulse can leave the far face of a cell before the peak has entered the front. These are not thought experiments: slow light at $17\ \mathrm{m/s}$ was measured in a sodium condensate in 1999, a negative group velocity of $-c/310$ in caesium vapour in 2000. Each looks like a loaded gun pointed at relativity, and each is completely consistent with it, because the group velocity tracks only the reshaped crest of an already-smooth envelope, never a genuine signal — which needs a non-analytic edge whose front Sommerfeld and Brillouin proved always advances at exactly $c$ , whatever $v_g$ does. This article owns the group-index construction and the pulse-reshaping picture, recalling Group Velocity and Group Index, Dispersion: Normal and Anomalous, and The Lorentz Oscillator Model by name.

Key takeaways

The envelope of a modulated pulse moves at $v_g = c/n_g$ with the group index $n_g = n + \omega\,\mathrm{d}n/\mathrm{d}\omega$ ; away from resonance $n_g \approx n$ and $v_g < c$ , but a steep dispersion slope drives $n_g$ to any value at all.
Steep positive $\mathrm{d}n/\mathrm{d}\omega$ inside a transparency window gives $n_g \gg 1$ , $v_g \ll c$ — slow light, measured to $17\ \mathrm{m/s}$ (Hau 1999, $n_g = 1.76\times10^{7}$ ).
Steep negative $\mathrm{d}n/\mathrm{d}\omega$ (anomalous dispersion) drives $n_g$ below one, to zero, or negative: fast light ( $v_g > c$ ), infinite ( $n_g \to 0$ ), or negative group velocity ( $v_g < 0$ , peak runs backward), measured at $-c/310$ (Wang 2000).
None of this signals faster than $c$ : $v_g$ describes only the reshaped crest of a smooth envelope, reconstructed from the pulse's own leading tail.
Causality is enforced by the front velocity, which equals $c$ exactly for every medium (Sommerfeld–Brillouin) because $n(\omega)\to 1$ as $\omega\to\infty$ : the sharp edge always arrives at $t = L/c$ .
Key idea: the group velocity $v_g = c/n_g$ can be made sub-luminal, super-luminal, infinite, or negative by shaping $\mathrm{d}n/\mathrm{d}\omega$ , because it tracks only the reshaped peak of an analytic envelope — while the information-bearing front always travels at exactly $c$ , so relativity is never in danger.

The core insight

The group velocity $v_g = c/n_g$ , with $n_g = n + \omega\,\mathrm{d}n/\mathrm{d}\omega$ , takes any value — sub-luminal, super-luminal, infinite, or negative — because it follows only the reshaped crest of a smooth envelope, not a causal marker. The front of a genuine signal is a non-analytic feature whose highest Fourier components see $n(\infty)=1$ , so it always advances at exactly $c$ . Sculpting the dispersion slope sculpts $v_g$ ; it never touches the front.

Where this lives today

Slow light left the paradox museum and became a component. Adiabatically ramping an electromagnetically-induced-transparency control field to zero maps a photon's quantum state onto a long-lived atomic spin coherence — the pulse is stopped, held, and released on command — now a workhorse for atomic-ensemble quantum memories and quantum repeaters. Slow-light delay lines built on stimulated Brillouin scattering in fibre or on photonic-crystal waveguides buffer optical packets without converting them to electronics, and the steep $\mathrm{d}n/\mathrm{d}\omega$ that slows the light also sharpens the phase response of interferometers and slow-light gyroscopes. Fast light has quieter uses: negative-dispersion (white-light) cavities broaden the response of laser gyroscopes and gravitational-wave detectors. The Boyd–Gauthier reviews treat slow and fast light as one tunable resource — the group index dialled anywhere from $10^{7}$ to negative — bounded always, and only, by the front velocity $c$ .

The physical picture

A wave packet is a chord, not a note — a bundle of frequencies whose crests line up at one place to make the envelope's peak. Each frequency travels at its own phase velocity, so the components drift out of step and back in; the peak reappears wherever they are momentarily in phase, and the speed of that rendezvous is $v_g$ . Near a resonance the phase relations twist so violently that the rendezvous can be dragged backward, or made to run faster than $c$ — but it is still interference bookkeeping, not a thing that carries energy or news. Slow light and fast light are two faces of the same reshaping. In a slow-light window the medium delays each spectral component by an amount that changes steeply across the pulse, and the reconstructed peak inches through the sample. In fast or negative-velocity light the medium amplifies (or less strongly absorbs) the leading edge relative to the trailing edge, so the reconstructed peak is pulled forward in time — sometimes so far forward that it emerges before the input peak arrives. The medium invents nothing: the output is stitched together entirely from the smooth leading tail of the incoming pulse, which an analytic envelope hands over in advance.

Figure

Refractive index n(ω) and group index n_g(ω) through an absorption resonance: steep normal-dispersion flanks give slow light (n_g ≫ 1), the anomalous centre drives n_g negative — fast and negative light

A single absorption line already contains both regimes: on its steep flanks $n_g$ swells and light slows, while at its centre $\mathrm{d}n/\mathrm{d}\omega$ is negative and $n_g$ dives below zero. The trouble is that line centre is also where the medium absorbs most fiercely, so fast light there is buried in loss. The art of the laboratory is to keep the steep slope while removing the absorption — electromagnetically induced transparency does it for slow light, a pair of gain lines for fast light — and the punchline is how far the measured group velocities then range.

Figure

The measured span of group velocities on a logarithmic |v_g| axis: EIT slow light at 17 m/s, ordinary glass near 2×10⁸ m/s, the vacuum line c, and gain-assisted fast light at v_g = −c/310 where the peak runs backward

The core equations

The whole subject grows from the dispersion relation $k = n(\omega)\,\omega/c$ . Differentiating it gives the envelope speed directly, and the medium's own frequency dependence $\mathrm{d}n/\mathrm{d}\omega$ rides along in the derivative and can dominate it — large and of either sign, so the group index roams far from $n$ .

v_g = \frac{\mathrm{d}\omega}{\mathrm{d}k} = \frac{c}{n_g}, \qquad n_g = n(\omega) + \omega\,\frac{\mathrm{d}n}{\mathrm{d}\omega} \qquad (1)

where: $v_g$ — group velocity, the speed of the envelope's peak; $k = n\omega/c$ — wavenumber in the medium; $n(\omega)$ — real part of the refractive index; $n_g$ — group index; $c$ — vacuum speed of light. The physical reading: the phase index $n$ tells you how fast crests move; the group index $n_g$ , fattened or starved by the slope $\mathrm{d}n/\mathrm{d}\omega$ , tells you how fast the bump moves. Spectroscopists tabulate the index against wavelength, so it helps to have the same statement in $\lambda$ . Substituting $\omega = 2\pi c/\lambda$ turns the frequency derivative into a wavelength derivative and flips its sign — which is why an anomalous region shows a positive $\mathrm{d}n/\mathrm{d}\lambda$ .

n_g = n + \omega\,\frac{\mathrm{d}n}{\mathrm{d}\omega} = n - \lambda\,\frac{\mathrm{d}n}{\mathrm{d}\lambda} \qquad (2)

where: $\lambda$ — vacuum wavelength; $\mathrm{d}n/\mathrm{d}\lambda$ — dispersion slope in wavelength, positive in an anomalous (fast-light) band. The two forms are identical arithmetic; use whichever matches the axis your data came on. What an experiment actually clocks is not $v_g$ but a time difference: how much later (or earlier) the pulse peak clears a slab of length $L$ than in vacuum. That delay is the whole diagnostic of slow and fast light, set by how far $n_g$ departs from one.

\Delta t = \frac{L}{v_g} - \frac{L}{c} = \frac{L}{c}\,(n_g - 1) \qquad (3)

where: $L$ — length of the medium; $\Delta t$ — transit time relative to vacuum, positive for a delay and negative for an advance. Read the signs: $n_g \gg 1$ gives a large positive delay (slow light); $0 < n_g < 1$ a small advance (fast light); $n_g < 0$ a large advance in which the peak emerges before it entered. None of these can smuggle information past $c$ , guaranteed by a single fact about every medium: no material polarises at arbitrarily high frequency, so its index tends to the vacuum value in the ultraviolet limit. The sharp edge of a genuine signal is built from exactly those runaway-high frequencies, and they see nothing but vacuum.

v_{\text{front}} = \lim_{\omega\to\infty}\frac{c}{n(\omega)} = c \qquad (4)

where: $v_{\text{front}}$ — velocity of the first, non-analytic wavefront (Sommerfeld's forerunner). The physical reading: whatever slow or backward antics the envelope peak performs, the true front — and with it any bit of information — checks in at $t = L/c$ and not one instant sooner.

Derivation AThe group index and the rigid envelope

The aim is to earn equation (1) honestly and, more importantly, to show why a quantity that can exceed $c$ or turn negative still describes the real motion of a real pulse. We build a band-limited packet as a superposition of plane waves about a carrier, expand the dispersion relation to first order across the narrow band, and watch the envelope emerge as a rigid shape translating at one speed — $c/n_g$ , whatever value that happens to be.

Step 1. Write the pulse as a Fourier sum of plane waves, $E(z,t) = \int \tilde{A}(k)\,e^{i[kz - \omega(k)t]}\,\mathrm{d}k$ , with $\tilde{A}(k)$ sharply peaked about a carrier $k_0$ . Why: a modulated signal is a band of frequencies, and only as such can the medium's frequency-dependent response enter.

Step 2. Expand the dispersion relation about the carrier, $\omega(k) \approx \omega_0 + v_g\,(k - k_0)$ , keeping the linear term and defining $v_g = \mathrm{d}\omega/\mathrm{d}k$ at $k_0$ . Why: for a narrowband pulse the quadratic term only broadens the envelope slowly; the peak's motion is entirely in the slope $\mathrm{d}\omega/\mathrm{d}k$ .

Step 3. Substitute and factor out the carrier: $E = e^{i(k_0 z - \omega_0 t)}\int \tilde{A}(k)\,e^{i(k-k_0)(z - v_g t)}\,\mathrm{d}k = A(z - v_g t)\,e^{i(k_0 z - \omega_0 t)}$ . Why: the integral depends on $z$ and $t$ only through $z - v_g t$ — the signature of a shape that slides without changing form, so the envelope $A$ is rigid and moves at $v_g$ .

Step 4. Convert to the index. From $k = n\omega/c$ , $\mathrm{d}k/\mathrm{d}\omega = (n + \omega\,\mathrm{d}n/\mathrm{d}\omega)/c \equiv n_g/c$ , so $v_g = c/n_g$ . Why: nothing above assumed $v_g$ was below $c$ or positive, so if the dispersion drives $n_g$ negative, the envelope translates backward — exactly what the negative-velocity experiments see. *Verified symbolically ( $v_g = c/n_g$ and $n_g = n + \omega\,\mathrm{d}n/\mathrm{d}\omega = n - \lambda\,\mathrm{d}n/\mathrm{d}\lambda$ ) in verify_aq-optics-superluminal-slow-light.py.*

Go deeper · Derivation B · 2 minSlow, fast, and negative light from the dispersion slope

Derivation BSlow, fast, and negative light from the dispersion slope

The goal is to connect the slope $\mathrm{d}n/\mathrm{d}\omega$ to the four measured regimes — slow, super-luminal, infinite, and negative group velocity — and to show that the caesium value $n_g = -310$ is not exotic but the natural reading of two gain lines. We take the Lorentz index near a resonance, read the sign and steepness of its slope in each band, then assemble a gain doublet whose transparent centre carries the steepest negative slope of all.

Step 1. Near a single Lorentzian resonance the real index rises to a maximum just below $\omega_0$ , falls through $n = 1$ at line centre, and reaches a minimum just above — the dispersion S-curve. On the outer flanks $\mathrm{d}n/\mathrm{d}\omega$ is large and positive; between the extrema it is large and negative. Why: the driven oscillators respond in phase below resonance and out of phase above it, and the swing bends the index up then down.

Step 2. Feed each band into $n_g = n + \omega\,\mathrm{d}n/\mathrm{d}\omega$ . On the flanks the positive slope makes $n_g \gg 1$ , so $v_g \ll c$ : slow light. At line centre the negative slope overpowers $n$ and drives $n_g$ below one, through zero, or negative. Why: the term $\omega\,\mathrm{d}n/\mathrm{d}\omega$ carries a large factor $\omega_0$ , so even a modest slope dominates $n_g$ .

Step 3. The catch at line centre is absorption: the steep negative slope sits atop the absorption peak, so a fast-light pulse is devoured. The fix is a gain doublet — two closely spaced amplifying lines. Midway between them the medium is transparent (even amplifying), yet the real index carries a steep negative slope inherited from the two gain features. Why: superposing two gain lines leaves a steep, negative-sloped, low-loss window in the gap — steep dispersion without the absorption penalty.

Step 4. Evaluate the group index there. The large negative slope gives $n_g = n + \omega_0\,\mathrm{d}n/\mathrm{d}\omega \ll -1$ ; the caesium experiment reached $n_g \approx -310$ , hence $v_g \approx -c/310$ and a peak advanced by tens of nanoseconds. Why: a negative group index is a routine consequence of engineered anomalous dispersion, not a violation of anything, because the output is reconstructed from the input's smooth leading edge. *Verified: the four regime signs, the negative Lorentz slope at line centre, and $v_g = -c/310$ for $n_g = -310$ in verify_aq-optics-superluminal-slow-light.py.*

Go deeper · Derivation C · 1 minThe front velocity is always c

Derivation CThe front velocity is always c

The purpose is to close the causality question: whatever the group velocity does, the fastest detectable feature of a signal moves at exactly $c$ in any physical medium. The argument turns on the difference between an analytic pulse, whose future is encoded in its past, and a real signal, which must contain a non-analytic edge whose Fourier content runs to infinite frequency — and on the one universal property of the index there.

Step 1. Distinguish the two waveforms. A smooth Gaussian is analytic: knowing it on any tiny interval fixes it everywhere, so a linear medium can legitimately reconstruct its peak early. A genuine signal must have a non-analytic point — a switch-on, a corner — carrying no such advance information. Why: only a non-analytic feature is unpredictable from the pulse's past, so only it can carry a new bit.

Step 2. Ask what the medium does to the highest frequencies in that edge. No material polarises instantaneously; its charges have inertia, so above every resonance the response dies and $n(\omega) \to 1$ as $\omega \to \infty$ . Why: to those runaway-high Fourier components the medium is indistinguishable from vacuum.

Step 3. Therefore the front travels at $v_{\text{front}} = \lim_{\omega\to\infty} c/n(\omega) = c$ and reaches the far face at $t = L/c$ , the earliest possible time — ahead of the reshaped peak in slow light, behind it in fast light, but always at $c$ . Why: the forerunner (Sommerfeld's precursor) sets a causal clock no reshaped peak can beat.

Step 4. Conclude that the signal velocity — the speed of a detectable change — is bounded by the front velocity, hence by $c$ . The super-luminal or backward peak is a reshaping artefact, fully compatible with special relativity. Why: information lives on the non-analytic edge, which obeys $c$ ; the peak is free precisely because it is not information. *Verified: the high-frequency limit $n(\omega)\to 1$ for the Lorentz index and the resulting $v_{\text{front}} = c$ in verify_aq-optics-superluminal-slow-light.py.*

Limiting cases

Case	Result	Meaning
$n_g \gg 1$ (steep positive slope)	$v_g = c/n_g \ll c$	slow light; EIT window reached $17\ \mathrm{m/s}$ ( $n_g = 1.76\times10^{7}$ )
$0 < n_g < 1$ (mild anomalous)	$v_g > c$	super-luminal peak; front still at $c$
$n_g \to 0$	$v_g \to \pm\infty$	the whole envelope appears to switch on at once along the sample
$n_g < 0$ (strong anomalous)	$v_g < 0$	negative group velocity; peak exits before it enters ( $-c/310$ , Wang 2000)
$n_g = 1$ (dispersionless)	$v_g = c$	vacuum or a flat-index window; envelope and phase move together
any medium, $\omega\to\infty$	$v_{\text{front}} = c$	causality guaranteed regardless of $v_g$

Common trap“A group velocity greater than c means a signal has outrun light”

It means no such thing, and conflating the two is the commonest error in the subject. The group velocity is the speed of the envelope's peak, and near resonance that peak is a reshaped feature — the medium amplifies the leading edge and attenuates the trailing edge, sliding the crest forward without transporting anything from front to back faster than $c$ . A true signal rides a non-analytic edge whose highest Fourier components see $n(\infty) = 1$ ; that front travels at exactly $c$ (Sommerfeld–Brillouin), arriving at $t = L/c$ in every medium ever measured. Treat " $v_g > c$ " as "the peak was reshaped," never as "a bit arrived early," and the paradox evaporates; the 2003 information-velocity experiments measured this directly and found detectable information pinned to $c$ .

Worked examples

1 · The Hau slow-light cell (confidence builder)

In 1999 a laser pulse crossed a sodium Bose–Einstein condensate at $v_g = 17\ \mathrm{m/s}$ . Take the cloud to be $L = 0.2\ \mathrm{mm}$ long. Find the group index and the delay relative to vacuum.

Step 1. Group index from $v_g = c/n_g$ : $n_g = c/v_g = (2.998\times10^{8})/17 = \mathbf{1.76\times10^{7}}$ . The light is slowed by nearly eighteen million — from $3\times10^{8}\ \mathrm{m/s}$ to a brisk walking pace.

Step 2. Transit delay from equation (3): $\Delta t = (L/c)(n_g - 1) \approx L/v_g = (2.0\times10^{-4})/17 = 1.18\times10^{-5}\ \mathrm{s} = \mathbf{11.8\ \mu s}$ . The vacuum transit time $L/c = 0.7\ \mathrm{ps}$ is utterly negligible beside it, which is why the delay is essentially the in-medium transit time itself.

Step 3. In human terms, an $11.8\ \mu\mathrm{s}$ delay over $0.2\ \mathrm{mm}$ is the lag light picks up crossing $3.5\ \mathrm{km}$ of vacuum, packed into a fleck of cold gas. The pulse is also spatially compressed — at $17\ \mathrm{m/s}$ a $1\ \mu\mathrm{s}$ pulse is only $17\ \mu\mathrm{m}$ long, so the whole thing fits in the condensate at once.

Step 4. The design lever: the delay scales as $L(n_g - 1)/c$ , and $n_g$ scales inversely with the transparency-window width, so a narrower EIT window buys more delay — until residual absorption caps the delay–bandwidth product. $n_g = 1.76\times10^{7}$ , delay $= 11.8\ \mu s$ across $0.2\ \mathrm{mm}$ .

2 · The Wang negative-velocity cell (exam level)

In 2000 a pulse crossed an $L = 6\ \mathrm{cm}$ caesium vapour cell prepared with a gain doublet, exiting with a group index $n_g = -310$ . Find the group velocity, the time by which the peak is advanced relative to vacuum, and explain why causality is unharmed.

Step 1. Group velocity: $v_g = c/n_g = (2.998\times10^{8})/(-310) = \mathbf{-9.67\times10^{5}\ m/s}$ . The sign is the whole story — the envelope peak moves backward through the cell even as the pulse as a whole is transmitted forward.

Step 2. Advance relative to vacuum, from equation (3): $\Delta t = (L/c)(n_g - 1) = (0.06/2.998\times10^{8})(-311) = -6.22\times10^{-8}\ \mathrm{s} = \mathbf{-62.2\ ns}$ . Equivalently $\Delta t = -(L/|v_g| + L/c) = -(62.0 + 0.2)\ \mathrm{ns}$ : the peak clears the far face $62.2\ \mathrm{ns}$ before it would have in vacuum, and in fact before the input peak has even reached the front face.

Step 3. Reconcile with causality. The $62.2\ \mathrm{ns}$ advance is a tiny fraction of the microsecond-scale smooth pulse, and it is reconstructed entirely from the pulse's leading tail, which the gain doublet amplifies while attenuating the trailing tail. Nothing has jumped from back to front: the medium reads the analytic pulse's future off its own past. The information-bearing front — a switch-on edge — still arrives at $t = L/c = 0.2\ \mathrm{ns}$ , exactly on the vacuum schedule.

Step 4. The peak may run backward while the front may not, because the peak is analytic and predictable while the front is non-analytic, its highest Fourier components seeing $n(\infty) = 1$ . Push the gain harder and $n_g$ grows more negative and the advance larger, but never enough to move the front — the advance saturates against the causal wall. $v_g = -c/310 = -9.67\times10^{5}\ \mathrm{m/s}$ , peak advanced $62.2\ \mathrm{ns}$ , front unmoved at $L/c = 0.2\ \mathrm{ns}$ .

Physics of the portrait

The man whose name sits on this article spent his life on its central question: when a wave's phase, group, and front move at different speeds, which — if any — carries the message?

Origins · 2 minLéon Brillouin and how this was discovered

Léon Brillouin

1889–1969 · Paris · Collège de France · New York

Léon Brillouin was born into physics — his father Marcel was a professor at the Collège de France — and trained under Arnold Sommerfeld in Munich just before the First World War and under Paul Langevin in Paris. It was with Sommerfeld, in 1913–1914, that the young Brillouin worked out what happens when a sharply switched-on wave first enters a dispersive medium: ahead of the main signal run two faint precursors, now called the Sommerfeld and Brillouin forerunners, and the very first of them travels at exactly $c$ regardless of the medium. That result — the front velocity is $c$ , full stop — is the causal backbone of this article, and Brillouin returned to it across his whole career, collecting the theory in his 1960 monograph Wave Propagation and Group Velocity, which carefully separates the phase, group, signal, and energy velocities that lesser treatments blur together. His name is scattered across modern physics: Brillouin zones organise the band structure of every crystal, and Brillouin scattering (predicted 1922) turns light off sound waves and now underlies fibre sensors and slow-light delay lines. He emigrated to the United States in 1941 and later turned to information theory. He died in New York in 1969, having spent fifty years insisting that a fast group velocity is not a fast signal.

History

1907

— Sommerfeld argues that even where the group velocity exceeds c, the signal front still advances at exactly c, answering a query of Wilhelm Wien

1914

— Sommerfeld and Brillouin work out the precursors of a step-modulated wave and prove the front velocity equals c in any dispersive medium

1960

— Brillouin collects the theory in Wave Propagation and Group Velocity, separating phase, group, signal, and energy velocities

1966

— Basov and co-workers observe pulse peaks advanced in laser amplifiers, the first laboratory fast light

1982

— Chu and Wong measure super-luminal group velocities in an absorbing semiconductor

1999

— Hau, Harris, Dutton, and Behroozi slow light to 17 m/s in a sodium Bose-Einstein condensate using electromagnetically induced transparency

2000

— Wang, Kuzmich, and Dogariu measure a negative group velocity, v_g = −c/310, in gain-assisted caesium vapour

2003

— Stenner, Gauthier, and Neifeld measure the information velocity directly and find it bounded by c

Going further

Texts. Boyd, Nonlinear Optics, 3rd ed., §3.3 and the slow-/fast-light reviews · Milonni, Fast Light, Slow Light and Left-Handed Light (IOP, 2005) — the standard monograph · Saleh & Teich, Fundamentals of Photonics, 3rd ed., §5.6 (pulse propagation and dispersion) · Loudon, The Quantum Theory of Light, 3rd ed., Ch. 2 (pulse propagation, signal velocity) · Brillouin, Wave Propagation and Group Velocity (Academic Press, 1960).

Provenance. The group-index and rigid-envelope results are standard (Saleh & Teich §5.6; Milonni Ch. 1). The slow-light number is Hau, Harris, Dutton & Behroozi, Nature 397, 594 (1999): $v_g \approx 17\ \mathrm{m/s}$ , giving $n_g = c/v_g = 1.76\times10^{7}$ and, for a $0.2\ \mathrm{mm}$ cloud, an $11.8\ \mu\mathrm{s}$ delay. The negative-velocity number is Wang, Kuzmich & Dogariu, Nature 406, 277 (2000): $n_g \approx -310$ , so $v_g = -c/310$ and, over a $6\ \mathrm{cm}$ cell, a $62.2\ \mathrm{ns}$ advance. The front-velocity theorem is Sommerfeld and Brillouin (Brillouin, 1960); the direct information-velocity measurement is Stenner, Gauthier & Neifeld, Nature 425, 695 (2003). Every figure is recomputed in verify_aq-optics-superluminal-slow-light.py.

Prerequisites. Group Velocity and Group Index · Dispersion: Normal and Anomalous · The Lorentz Oscillator Model · Complex Refractive Index and the Extinction Coefficient · Phase Velocity. Leads to. Electromagnetically Induced Transparency · Kramers–Kronig Relations · Optical Buffers and Quantum Memories.

If the pulse peak leaves before it enters, where does the output pulse come from?

From the leading edge of the input pulse, which the medium reads ahead. A smooth (analytic) pulse encodes its entire shape in any small slice of its early tail, so a linear medium with the right dispersion can amplify that tail and attenuate the trailing part, sliding the reconstructed peak forward, even ahead of the input peak. Nothing travels from the back of the pulse to the front faster than $c$ ; the medium merely re-weights frequency components all present in the incoming light from the start. The tell: the effect works only for pulses smooth enough to be predictable — introduce a genuine non-analytic edge and the reshaping cannot anticipate it.

Why can't slow light just use an absorption line, whose dispersion is already steep?

Because the steep dispersion at an absorption line sits directly on the absorption peak, so the pulse is destroyed before the group delay can accumulate. You want the steepness without the loss. Electromagnetically induced transparency opens a narrow, nearly loss-free window inside an absorbing transition using a second, strong control field; the window carries a very steep positive $\mathrm{d}n/\mathrm{d}\omega$ while the absorption at its centre is driven to almost zero. That combination — enormous $n_g$ , negligible loss — is what made $17\ \mathrm{m/s}$ possible. The narrower the window, the steeper the slope and the slower the light, until the delay–bandwidth product runs into residual dephasing.

Is the group velocity ever simply the wrong quantity to use?

Yes, and knowing when is part of the craft. The identity $v_g = c/n_g$ describes a pulse peak only when the envelope stays roughly intact — when absorption, gain, and higher-order dispersion do not badly reshape it over the length. Precisely in the regimes that give the flashiest numbers, those effects often dominate, the pulse distorts, and "the peak" becomes ambiguous or splits. Then the group velocity is a bookkeeping parameter, not the speed of any identifiable thing, and the honest description propagates the full spectrum. The front velocity, by contrast, is always meaningful, depending only on $n(\infty) = 1$ .

What is the difference between phase, group, signal, and front velocity?

They are four distinct speeds that coincide only in vacuum. The phase velocity $v_p = c/n$ is the speed of a single crest and can exceed $c$ with no consequence, since one crest carries nothing. The group velocity $v_g = c/n_g$ is the speed of the envelope peak and, as this article shows, can be anything at all. The signal velocity is the speed at which a detectable change propagates, and it is bounded by $c$ . The front velocity is the speed of the first non-analytic tremor and equals $c$ exactly. Brillouin's 1960 book exists largely to keep these four separate; most "faster-than-light" claims come from quietly swapping the group velocity for the signal velocity.

How is slow light turned into stopped light — a memory?

By handing the photon's information to the atoms and then taking the light away. In electromagnetically induced transparency the propagating excitation is a hybrid of light and atomic spin coherence — a "dark-state polariton." As you slowly reduce the control-field intensity, the group velocity falls toward zero and the polariton becomes almost purely atomic: the photon's amplitude and phase are now written into a long-lived collective spin state, with essentially no light left in the cell. Restoring the control field runs the process backward and regenerates the original pulse, on demand. This is the physical basis of atomic-ensemble quantum memories, where a single-photon qubit can be stored and retrieved for a controllable time.

Does a negative group velocity mean the pulse physically travels backward?

Inside the medium the envelope peak does move backward, toward the entrance — this has been imaged directly, and it is a real reshaping, not an artefact of plotting. What does not move backward is energy or information. The forward-transmitted pulse is continuous with the incident one, and the backward-running interior peak is merely the meeting point of frequency components all still marching forward at their own phase velocities. It is like a row of stadium spectators standing in sequence: the "wave" can be sent around the stadium faster than anyone can run, or in reverse, because no spectator travels with it. The front — the first person to stand — obeys the ordinary speed limit.