A negative-group-delay (NGD) circuit makes the peak of a smooth pulse leave the device before the input peak has fully arrived. It does this without breaking causality: it absorbs the trailing part of the pulse spectrum more than the leading part. Engineers use the effect to cancel unwanted delay in antenna feeds and interconnects. The field's central complaint is that more time advance seems to cost more loss. Published devices, however, report that cost in incompatible units. One reports nanoseconds of advance against decibels of insertion loss, another nanoseconds against a 3-dB bandwidth, a third group velocity against decibels per centimetre. As a result, nobody can say whether a given device is near the physical limit or far from it. The same gap prevents testing a striking claim in the literature, that NGD can be obtained "in the absence of absorption resonance."
We propose to make that trade-off measurable. The tool is a classical fact from circuit theory, the gain–phase relation. Any causal, stable, linear device can be split into two parts. The first is a "minimum-phase" part, whose delay is completely fixed by how the transmission magnitude varies with frequency. The second is an "all-pass" part, which changes only phase and can only add positive delay. The most negative delay a device can possibly have is therefore set by its measured magnitude curve alone. The size of that limit depends on the variation of the magnitude across the band, not on its average level. We will compute this limit from vector-network-analyzer measurements of fabricated devices. We will then report each device's "NGD efficiency," meaning how much of the allowed advance it actually delivers. Finally, we will use the split to diagnose disagreements between simulation and measurement. A fabrication error can either reshape the magnitude response or add wasted positive delay, and the split tells these apart.
Core Claim
Write the device's transmission as $T(\omega)$, with magnitude $|T(\omega)|$ and group delay $\tau(\omega) = -d\arg T/d\omega$. For any causal, stable, passive two-port the group delay decomposes as
$$\tau(\omega) = \tau_{\min}(\omega) + \tau_{\mathrm{ap}}(\omega), \qquad \tau_{\mathrm{ap}}(\omega) \ge 0.$$
Here $\tau_{\min}$ is the delay of the minimum-phase network with the same magnitude, obtained by taking the Hilbert transform of $\ln|T(\omega)|$ and differentiating it. The term $\tau_{\mathrm{ap}}$ is the delay from the remaining all-pass factor, which is non-negative for any stable all-pass. Three consequences follow, and each can be tested.
First, the advance a device can deliver is bounded by $\tau_{\min}$, which depends only on the measured magnitude. A frequency-flat loss, such as a uniform resistive pad, contributes nothing to $\tau_{\min}$. Second, consider an attenuation dip of depth $A$ nepers (one neper is about 8.7 dB) and half-width $\Delta\omega$. Dimensional analysis of the Hilbert transform gives a central $\tau_{\min}$ of order $-A/\Delta\omega$. So the delay–bandwidth product is paid for in attenuation contrast, not in absolute insertion loss. Third, real devices will show $\tau_{\mathrm{ap}} > 0$ from feed lines, multipath branches and parasitics. This excess is recoverable design slack, separate from the fundamental cost.
The bound itself is textbook material. What would be new is using it as a measurement standard across NGD mechanisms, together with three uses of it. We would put the resonant, distributed multi-port and non-resonant devices on one common axis. We would test whether the non-resonant mechanism in Negative Group Velocity in the Absence of Absorption Resonance escapes the trade-off or simply shifts the magnitude contrast away from the absorption peak. And we would use the split to tell whether fabrication tolerances degrade NGD by reshaping magnitude or by adding all-pass delay.
Prior Work and How It Positions This
Three experimental NGD papers motivate the proposal, and all three leave the trade-off unmeasured in a common currency.
Negative Group Delay Metamaterials Based on Split-Ring Resonators and Their Application builds tantalum-nitride split-ring resonators on Teflon. A split-ring resonator is a small metal loop with a gap that resonates at microwave frequencies. The paper validates their circuit models against HFSS and an Agilent E8364C network analyzer, reporting up to −0.1 ns at 12–13 GHz for the edge-coupled ring and −0.04 ns with under 2.7 dB loss for the single ring. It reports insertion loss rather than contrast, gives no bandwidth for the single-ring result, and notes that the loss degrades antenna gain.
Fully Microstrip Three-Port Circuit Bandpass NGD Design and Test derives transfer-matrix models for a three-port microstrip network with resistive terminations. It measures about −3.8 ns over 75 MHz near 0.74 GHz and shows robustness to moderate impedance and delay variation. Its branch structure is exactly where all-pass excess delay should arise, and that excess has never been separated out.
Negative Group Velocity in the Absence of Absorption Resonance obtains about −0.3c with −2.18 dB/cm decay in a split-ring slab operated near its plasma frequency at 15° incidence. It frames this as evading the resonance–loss link implied by the Kramers–Kronig relations, which tie a material's absorption to its dispersion. The paper apparently contradicts our core claim, and we address that directly. Our bound does not require a resonance, only magnitude variation across the signal band. A dB/cm figure reports average loss, not contrast. The prediction is that this mechanism obeys the same $\tau_{\min}$ bound, and that its advantage, if real, is a small $\tau_{\mathrm{ap}}$ or a favourable placement of contrast. Low total loss is not the advantage. If measured transmission from that geometry ever showed delay more negative than $\tau_{\min}$, the device would not be behaving as a linear causal system along that path, which would be a significant finding in itself.
The validation literature constrains how the measurements must be done. Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows in FLNG Cryogenic Hoses shows that manufacturing variation, there wall-thickness spread, moves measured failure thresholds away from idealized models. That is why we fabricate deliberately perturbed copies rather than single prototypes. Estimation of vehicle-induced bridge dynamic responses using fiber Bragg grating strain gages shows that a poorly placed sensor can miss a mode entirely while the frequencies it does see match the model to within 3.63%. The analogous risk here is that a finite measurement band misses magnitude structure that the Hilbert transform needs, so we treat band truncation as an explicit uncertainty. For calibrating simulation models against measurements we will start with ordinary least squares. The tree-structured Parzen estimator used for 41-parameter calibration in Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization, and likelihood-free Bayesian optimization from A General Recipe for Likelihood-free Bayesian Optimization (code), are reserved for the case where the uncertain-parameter count grows large.
Research Program
The first component is the estimator. We will write and validate code that takes a measured $|S_{21}|$, the network analyzer's transmission coefficient, over a finite band. The code returns $\tau_{\min}(\omega)$ with an uncertainty interval. The Hilbert transform is global, so out-of-band behaviour must be bracketed. We will do this with physically motivated extrapolations: a constant level, the low-pass roll-off of the substrate and connectors, and a worst-case contrast allowed by passivity. We will also widen the measurement span until $\tau_{\min}$ inside the NGD band stops changing. The estimator is checked on synthetic rational transfer functions with known delay, and on a calibrated through standard, which should yield an excess close to zero.
The second component is a common metric. For every device we report three quantities. The first is the measured advance divided by the allowed advance, $\tau/\tau_{\min}$, which is the NGD efficiency. The second is the delay–bandwidth product per neper of in-band attenuation contrast. The third is absolute insertion loss, listed separately. From the reported numbers alone, the three-port circuit's product is 3.8 ns × 75 MHz. Our scaling estimate predicts that this product implies roughly 7–8 dB of magnitude contrast. Measurement will confirm or refute that before any new design is attempted.
The third component is a device family, measured and calibrated. We will reproduce the three-port microstrip design at 0.74 GHz on the same class of Rogers substrate, along with frequency-scaled versions at several GHz. Gap 2 of the file identifies scaling to higher frequencies as open; it is also where resistor parasitics should push $\tau_{\mathrm{ap}}$ up. We will build planar split-ring loaded lines at 12–13 GHz modelled on the single-ring and edge-coupled cells. Tantalum-nitride laser cutting may be unavailable, so we will use resistive films where possible and report the substitution explicitly. If the slab geometry can be reproduced, we will also measure it at 0° and 15° incidence. Each design is built in a nominal copy and in copies with deliberate gap, width and resistor-value offsets. Simulation models are calibrated to the measurements over the uncertain parameters: permittivity, etch undercut and resistor parasitics. The minimum-phase/all-pass split is then applied separately to simulated and measured responses. This shows which part each discrepancy lives in, and so partly resolves the parameter-degeneracy problem: two parameter sets that match $|S_{21}|$ equally well can still differ in $\tau_{\mathrm{ap}}$.
The fourth component is a design test. Suppose efficiencies come out well below one. Then we redesign the three-port network to minimize $\tau_{\mathrm{ap}}$ at fixed magnitude contrast, for example by shortening the delay branches or rebalancing the terminations. We then test whether measured advance increases without any increase in contrast. That outcome would show the "trade-off" as practiced includes avoidable waste. If efficiencies are already near one, the trade-off is fundamental for these devices, and only reshaping the magnitude response can improve them.
Baselines
The naive reference is the field's current reporting practice: measured peak NGD plotted against insertion loss, as in the split-ring and three-port papers. We will tabulate our devices this way too. The proposal beats it if the new metric orders devices differently from raw insertion loss, and that ordering predicts which redesigns improve advance. One concrete test is a pair of devices with near-equal insertion loss but different contrast, which the naive view treats as equivalent and the bound does not.
The strongest existing device is the reproduced three-port microstrip circuit from Fully Microstrip Three-Port Circuit Bandpass NGD Design and Test, with reported −3.8 ns over a 75 MHz 3-dB NGD bandwidth at about 0.74 GHz. Among split-ring cells, the strongest is the single-ring result from Negative Group Delay Metamaterials Based on Split-Ring Resonators and Their Application, with −0.04 ns at under 2.7 dB loss. For the slab, the reference is the −0.3c at −2.18 dB/cm from Negative Group Velocity in the Absence of Absorption Resonance. Our reproduction must first match the reported three-port numbers within measurement uncertainty before any comparison counts. The redesign beats this baseline if it delivers a larger delay–bandwidth product at the same measured magnitude contrast and the same substrate and component class. Both are measured on one analyzer with one calibration, over a span chosen by the same convergence rule.
The strongest competing method is a purely optimization-driven Pareto search. This means sweeping or Bayesian-optimizing the three-port geometry for raw NGD against insertion loss and bandwidth, as Gap 4 in the file suggests. We will run tree-structured Parzen search and the likelihood-free optimizer on the calibrated circuit model with the same evaluation budget as our redesign. Our approach beats it only if the excess-minimizing design, or the optimizer constrained by the bound, reaches front points the unconstrained search misses, or reaches them with fewer evaluations. If the unconstrained optimizer finds the same designs, the bound's value is purely diagnostic, and we will say so.
The key ablation removes the decomposition. We redesign by matching the target group delay directly, without separating $\tau_{\min}$ from $\tau_{\mathrm{ap}}$, under the same budget. A second ablation computes the bound without out-of-band extrapolation, which shows how much the conclusions depend on the truncation treatment.
Why not just…?
Why not just run Bayesian optimization over the geometry and report the Pareto front? For the three-port circuit the analytic transfer-matrix model is cheap to evaluate, so even a dense grid is affordable. Sample efficiency is therefore not the bottleneck. The real limitation is that a front in raw insertion loss mixes flat loss, which is irrelevant, with contrast, which is the actual cost. Such a front also cannot say whether it lies at a physical limit. This alternative is a genuine threat on the design side, and it is run as a baseline under equal budget. The decomposition earns its place only if the bound-guided design measurably beats it or explains its front.
Why not just add an amplifier to recover the loss? The bound agrees that this works for absolute loss. Flat gain shifts $\ln|T|$ by a constant, which leaves $\tau_{\min}$ unchanged. An amplifier, however, cannot reduce the magnitude contrast that the advance requires. It also adds DC power, noise and stability concerns, which the passive three-port design was explicitly built to avoid. Our metric separates these effects by construction.
Why not just trust full-wave simulation instead of fabricating? The slab paper attributes its simulation–measurement mismatch to fabrication tolerance. The bellows study shows that manufacturing spread shifts measured thresholds. Group delay is a frequency derivative of phase, so it amplifies small errors. The perturbed-copy experiment tests this directly. If calibrated simulation predicts both $\tau_{\min}$ and $\tau_{\mathrm{ap}}$ of the perturbed copies within measurement uncertainty, fabrication adds little and simulation suffices for future designs. That would be a useful negative result.
Why not just cite Kramers–Kronig and skip the measurements? The relation is known, but nobody has measured how close real NGD devices come to it. The non-resonant claim sits in apparent tension with it. Only measured magnitude and phase on the same devices can settle whether that mechanism has small excess delay, favourable contrast placement, or something else.
Proposed Experiments
The first decisive check is consistency. For every fabricated device, measured $\tau(\omega)$ must lie at or above $\tau_{\min}(\omega)$, computed from the same sweep, within the truncation uncertainty interval. A systematic violation would indicate either an estimator flaw, caught first on the synthetic and through-standard checks, or a non-causal measurement artifact. For the slab geometry it would challenge the linear-causal-path picture.
The second is the efficiency survey. We place the reproduced three-port circuits, the frequency-scaled versions, the split-ring lines and, if built, the slab at both incidence angles on the efficiency-versus-contrast axes. The prediction is that efficiency drops with frequency scaling as parasitics add all-pass delay. A second prediction is that the slab's advantage shows up as high efficiency rather than as advance beyond the contrast bound.
The third is the tolerance attribution. Across the perturbed copies, we regress the change in measured advance on the change in $\tau_{\min}$ and the change in $\tau_{\mathrm{ap}}$ separately. This reveals which fabrication errors damage NGD by which route, and whether calibrated simulation captures both.
The fourth tests time-domain validity. We send modulated envelopes through each device, as in the slab experiment, using a carrier at band centre and an envelope narrower than the NGD band. We then confirm that the measured peak advance agrees with the frequency-domain delay. This ensures the efficiency metric describes what a pulse actually experiences.
The final test is the redesign against the optimizer baseline at matched budget and matched measured contrast. The proposal succeeds if the bound-guided three-port circuit measures a larger delay–bandwidth product than both the reproduced original and the unconstrained optimizer's best design at the same contrast. It fails, and becomes a diagnostic tool only, if all three coincide.
References
- Negative Group Velocity in the Absence of Absorption Resonance
- Negative Group Delay Metamaterials Based on Split-Ring Resonators and Their Application
- Fully Microstrip Three-Port Circuit Bandpass NGD Design and Test
- Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows in FLNG Cryogenic Hoses
- Estimation of vehicle-induced bridge dynamic responses using fiber Bragg grating strain gages
- Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization: A Case Study on Pebax Polymer
- A General Recipe for Likelihood-free Bayesian Optimization