---
title: Switch Loss Across Multiple Domains
url: https://www.emergentmind.com/topics/switch-loss
type: topic
---

# Switch Loss Across Multiple Domains

Searching arXiv for recent and foundational papers on “switch loss” across RF, photonics, packet switching, power electronics, and quantum switch contexts.
“Switch loss” is not a single invariant quantity across the literature. In the cited arXiv corpus, it denotes several distinct but structurally related penalties associated with switching operations: ON-state insertion loss and OFF-state isolation in RF and photonic devices, packet loss rate in buffered packet switches and finite queues, turn-on and turn-off energy dissipation in power converters, and information loss under a quantum switch. A separate but terminologically adjacent usage appears in neural PDE training, where switching the optimization objective induces a “loss jump” rather than a device-level loss process [1701.01763] [1006.0841] [2212.05246] [2307.01964] [2405.03095].

## 1. Domain-specific meanings of switch loss

The literature uses the same phrase for materially different observables. In RF and integrated photonics, the dominant quantity is usually insertion loss, commonly inferred from \(S_{21}\) in the ON state; isolation and return loss are complementary figures. In optical packet switching and queueing, the relevant loss is packet dropping under contention or finite-buffer overflow. In power electronics, switching loss is explicitly separated from conduction loss and tied to transient commutation energy. In the quantum-switch literature, the central object is information loss relative to a fixed point, together with a switch-induced memory term [1701.01763] [2503.02149] [1006.0841] [2212.05246] [2307.01964].

| Context | Loss quantity | Representative papers |
|---|---|---|
| RF and microwave switches | ON-state insertion loss, OFF-state isolation, return loss | [1701.01763], [2406.12105], [2503.02149], [2012.07818] |
| Photonic and optical switches | Insertion loss, extinction ratio, crosstalk, loss-efficiency product | [1807.07114], [2405.07809], [2411.18139], [2507.14729], [1208.4813] |
| Packet switching and queues | Packet loss rate, loss fluctuations, temporal correlations | [1006.0841], [0803.3314] |
| Power electronics | Turn-on/turn-off switching energy and average switching power | [2212.05246], [1906.04428], [2311.07029], [2506.18635] |
| Quantum and algorithmic settings | Information loss, switch-induced memory, loss jump after objective switching | [2307.01964], [2405.03095] |

A recurrent misconception is to equate switch loss exclusively with insertion loss. The surveyed literature does not support that restriction. Another common simplification is to treat loss as a scalar detached from architecture; in nearly every domain here, the reported loss depends on topology, control policy, parasitics, or optimization geometry.

## 2. RF and microwave switch loss as transmission penalty

In RF switch design, switch loss is typically the ON-state transmission penalty. The SOI RF switch for wireless sensor and health-care front-end use operates over \(0\!-\!5\ \mathrm{GHz}\) in a shunt-series topology. The paper identifies \(S_{11}\) as input reflection coefficient, \(S_{21}\) as insertion loss / forward transmission, and \(S_{31}\) as isolation / transmission to the off arm. At \(5\ \mathrm{GHz}\), it reports insertion loss \(=0.906\ \mathrm{dB}\), isolation \(=30.95\ \mathrm{dB}\), input-referred IP3 \(=53.05\ \mathrm{dBm}\), and \(1\ \mathrm{dB}\) compression point \(=50.06\ \mathrm{dBm}\); the design uses a modified series-shunt FET switch with stacked devices in \(0.18\ \mu\mathrm{m}\) Power Jazz SOI technology and Cadence software [1701.01763].

The same basic interpretation appears in mechanically actuated RF switching, but with different loss mechanisms. The pneumatically controlled AeroSwitch paper uses measured \(S_{21}\) from \(100\ \mathrm{MHz}\) to \(500\ \mathrm{MHz}\) as the primary proxy for insertion loss and isolation. The single-switch measurements report AeroSwitch average insertion loss \(-0.15\ \mathrm{dB}\), PIN diode average insertion loss \(-0.25\ \mathrm{dB}\), AeroSwitch average isolation \(-35\ \mathrm{dB}\), and PIN diode average isolation \(-20\ \mathrm{dB}\). The paper explicitly notes that the numerical sign convention in the results section is somewhat inconsistent with the narrative claim of “slightly reduced average insertion loss,” but it attributes the favorable behavior to lower conductive resistance and the absence of RF choke, DC-feed, and DC-block networks in the main RF path. In an MRI-relevant matching network, the AeroSwitch shows an average \(62\%\) improved Q-factor compared to the PIN diode, and under \(500\ \mathrm{MHz}\), \(100\ \mathrm{W}\), \(10\)-minute testing it remains below \(30^\circ\mathrm{C}\) while the PIN diode reaches \(40.5^\circ\mathrm{C}\) in the OFF condition after \(10\) minutes [2503.02149].

Cryogenic RF switching introduces an additional bandwidth and temperature dimension. The compact mm-wave SPST switch in \(22\ \mathrm{nm}\) FDSOI CMOS reports at \(2\ \mathrm{K}\) an insertion loss of \(1.3\ \mathrm{dB}\) at \(1\ \mathrm{GHz}\) degrading gradually to \(2.3\ \mathrm{dB}\) at \(70\ \mathrm{GHz}\), isolation of \(62\ \mathrm{dB}\) at \(1\ \mathrm{GHz}\) and \(25.3\ \mathrm{dB}\) at \(70\ \mathrm{GHz}\), and return loss better than \(-11.5\ \mathrm{dB}\) over DC–\(70\ \mathrm{GHz}\). Its design avoids large passive components and uses a third n-MOSFET to improve isolation, with a reported \(3\ \mathrm{dB}\) improvement in isolation at \(60\ \mathrm{GHz}\) from adding that device in post-layout simulation [2406.12105].

Optically controlled RF switching yields still another loss profile. The silicon plasma switch for \(1\!-\!4\ \mathrm{GHz}\) reports measured ON-state insertion loss of less than \(0.33\ \mathrm{dB}\) and return loss better than \(20\ \mathrm{dB}\) across the band under \(1.5\ \mathrm{W}\), \(915\ \mathrm{nm}\) excitation, with OFF-state isolation from \(27\ \mathrm{dB}\) at \(1\ \mathrm{GHz}\) to \(17\ \mathrm{dB}\) at \(4\ \mathrm{GHz}\). The paper models the ON state mainly as a resistor and the OFF state as resistor plus parallel capacitance, linking lower loss to optically generated electron-hole plasma in the silicon chiplet bridging a \(1.075\ \mathrm{mm}\) microstrip gap [2012.07818].

## 3. Optical and photonic switch loss as insertion loss, crosstalk, and loss-equivalence

In integrated photonics, switch loss is often inseparable from extinction ratio, crosstalk, active volume, and spectral bandwidth. The GST-on-SOI nonvolatile on-off silicon photonic switch explicitly frames a trade-off between ON-state insertion loss and extinction ratio. Its optimized partially etched geometry reports \(43\ \mathrm{dB}\) extinction ratio and \(2.76\ \mathrm{dB}\) insertion loss at \(1550\ \mathrm{nm}\) for a GST volume of \(400\ \mathrm{nm}\times180\ \mathrm{nm}\times450\ \mathrm{nm}\), while a lower-loss operating point reaches \(1.2\ \mathrm{dB}\) insertion loss at \(10\ \mathrm{dB}\) extinction ratio for \(100\ \mathrm{nm}\times150\ \mathrm{nm}\times450\ \mathrm{nm}\). The paper attributes the loss to absorption in crystalline GST, mode alteration due to high index change at Si–GST interfaces, and reflections at those interfaces; it also proposes a static-performance figure of merit, \(\mathrm{FOM}=ER/(IL\times \text{active volume})\), with ER and IL taken in linear form [1807.07114].

A different nonvolatile PCM design, based on slot-waveguide concentration with Sb\(_2\)Se\(_3\), drives the optical-loss figure much lower. The proposed \(2\times2\) switch reports at \(1550\ \mathrm{nm}\) cross-state IL \(=-0.27\ \mathrm{dB}\), bar-state IL \(=-0.11\ \mathrm{dB}\), cross-state CT \(=-23.9\ \mathrm{dB}\), and bar-state CT \(=-27.4\ \mathrm{dB}\). The insertion loss remains less than \(0.5\ \mathrm{dB}\) from \(1525\ \mathrm{nm}\) to \(1575\ \mathrm{nm}\), with a \(0.3\ \mathrm{dB}\) IL bandwidth of at least \(55\ \mathrm{nm}\). The architecture uses a \(10\ \mu\mathrm{m}\) slot fully filled with Sb\(_2\)Se\(_3\) and a single-layer graphene heater, and the paper relates low loss to strong slot-mode overlap with a low-loss PCM rather than weak evanescent interaction [2405.07809].

In silicon electro-optic MZI switching, the key loss issue may be not absolute attenuation but arm imbalance. The cascaded-phase-shifter MZI switch identifies free-carrier absorption loss imbalance between the two interferometer arms as the dominant source of crosstalk. It defines arm loss imbalance as \(IB=\log(T_1/T_2)\) and relates crosstalk through
\[
XT = 20\log\left(1 - \frac{2}{10^{IB/20} + 1}\right).
\]
For the optimized geometry, the best simulated design uses a lightly doped length of \(140\ \mu\mathrm{m}\) and a heavily doped length of \(52\ \mu\mathrm{m}\); at \(V_{\pi/2}\approx1.207\ \mathrm{V}\), the arm loss imbalance is about \(0.007\ \mathrm{dB}\), simulated crosstalk reaches \(-58\ \mathrm{dB}\) for BAR and \(-51.2\ \mathrm{dB}\) for CROSS at \(1310\ \mathrm{nm}\), and the overall simulated switch loss is about \(1.31\ \mathrm{dB}\). Experimentally, the fabricated switch exhibits crosstalk between \(-33\) and \(-44.2\ \mathrm{dB}\) at \(1316\ \mathrm{nm}\), maintains crosstalk below \(-30\ \mathrm{dB}\) across a \(61\ \mathrm{nm}\) bandwidth, and reports measured insertion losses around \(2.1\) to \(3.4\ \mathrm{dB}\), with static measured switch losses below \(2.6\ \mathrm{dB}\) [2411.18139].

The liquid-crystal-cladded silicon directional coupler switch introduces a composite metric, the loss-efficiency product \(aV_zL\), where \(a\) is optical propagation loss in \(\mathrm{dB/mm}\) and \(V_{\pi}L\) is the voltage-length efficiency. The demonstrated \(1.5\ \mathrm{mm}\) device reports \(V_{\pi}L=0.0195\ \mathrm{V\cdot mm}\), switching voltage \(V_z=0.013\ \mathrm{V}\), extinction ratio \(\sim9\ \mathrm{dB}\), on-chip loss \(-4.8\pm0.8\ \mathrm{dB}\), and \(aV_zL=0.0624\ \mathrm{V\cdot dB}\). The paper compares this with an earlier MZI switch at \(-11.0\ \mathrm{dB}\) on-chip loss and \(0.2464\ \mathrm{V\cdot dB}\), describing about a \(4\times\) improvement in the composite metric, while also noting that the present \(1.5\ \mathrm{mm}\) demonstration has too much loss for large cascaded arrays in its current form [2507.14729].

A cavity-based all-optical interpretation of low-loss switching appears in the microdisk Zeno switch using EIT. Rather than attenuating the signal directly, absorption changes whether the resonator field can build up. The paper predicts more than \(35\ \mathrm{dB}\) of switching contrast with less than \(0.1\ \mathrm{dB}\) loss using \(2\ \mu\mathrm{W}\) of control-beam power, and gives two operating points: one with through-port contrast \(50\ \mathrm{dB}\), drop-port contrast \(25\ \mathrm{dB}\), through-port loss \(0.5\ \mathrm{dB}\), drop-port loss \(0.02\ \mathrm{dB}\), and about \(516\ \mathrm{MHz}\) bandwidth for each port, and another with \(38\ \mathrm{dB}\) contrast in both ports and \(0.1\ \mathrm{dB}\) loss in both ports [1208.4813].

## 4. Packet-switched and queueing interpretations of switch loss

In optical packet switching, switch loss is packet loss rate rather than transmission attenuation. The two-stage shared FDL optical packet switch uses a main switch together with Aux. Switch-I and Aux. Switch-II, the latter containing both feed-forward and feedback shared FDLs. Its contention-resolution scheme assigns higher priority to releasing packets from Aux. Switch-II than from Aux. Switch-I. The paper evaluates packet loss rate, average delay, and offered load \(p\), with Poisson arrivals, uniformly distributed traffic, packet length equal to one time unit, and a maximum of five recirculations in the feedback FDLs. In a \(32\times32\) switch, it reports \(PLR=0\) at \(p=0.3\) with \(m=12\) FDLs and \(PLR=0\) at \(p=0.6\) with \(m=40\); at heavy load \(p=0.9\), near-zero packet loss is achieved when \(m=60\). More generally, the paper states that zero packet loss rate is achievable when \(p<0.8\), and that Aux. Switch-II can reduce PLR by up to \(90\%\) at \(p=0.6\) and \(25\%\) at \(p=0.9\) [1006.0841].

The finite-buffer queueing literature adds a critical-statistical perspective. In the single-node packet-switched network model, losses are caused by arrivals when the queue is at full capacity. For the discrete model, the number of dropped packets in a window of length \(N\) is
\[
\mathcal{L}_N(n_0)=\sum_{n=n_0+1}^{n_0+N} \delta_{\ell_n,L}\,\delta_{\ell_{n+1},L}.
\]
The mean loss rate has a sharp transition near \(p=1/2\):
\[
\frac{1}{N}\langle \mathcal{L}_N\rangle = p \frac{q^{L+1}-q^L}{q^{L+1}-1}, \qquad q=\frac{p}{1-p},
\]
with asymptotics exponentially small in \(L\) for \(p<1/2\), order \(1/(L+1)\) at \(p=1/2\), and finite for \(p>1/2\). The paper’s principal result is that finite capacity and packet-dropping boundary conditions generate strong fluctuations and temporal correlations even for Markovian arrivals: at criticality, the loss correlation behaves as
\[
R_2(N,M)\big|_{p=1/2} = c^{-1}\sqrt{\frac{N}{2\pi M}},
\]
so loss bursts decay only as a power law in the separation \(M\) [0803.3314].

These results show that, in switching networks, “loss” is a scheduling and boundary phenomenon rather than a propagation penalty. A plausible implication is that switch architecture and release policy play the same role for packet loss that topology and parasitic engineering play for insertion loss in physical switches.

## 5. Switching loss in power electronics

In power electronics, switching loss is explicitly distinguished from conduction loss. The GaN HEMT-based \(3\)-level ANPC inverter paper defines conduction loss through on-state dissipation and switching loss through turn-on and turn-off energy per commutation. Using fitted device-characterization curves,
\[
E_{\mathrm{ON}} = \left(1.0527\, |I_{\mathrm{DS}}|^{1.6291}\right)\times 10^{-6}\; [\mathrm{J}],
\]
\[
E_{\mathrm{OFF}} = \left(2.542\, |I_{\mathrm{DS}}|^{1.1738}\right)\times 10^{-6}\; [\mathrm{J}],
\]
and
\[
P_{\mathrm{sw}} = \sum_{i=1}^{n}\left(P_{\mathrm{ON}i} + P_{\mathrm{OFF}i}\right).
\]
The model compares DNPC, ANPC-SSCM, ANPC-OSCM, and ANPC-FPCM, and states that ANPC-FPCM provides the lowest total losses. Validation against PSIM reports error below \(3\%\) for the studied operating point \((V_{\mathrm{DC}}=800\ \mathrm{V}, f_{\mathrm{sw}}=50\ \mathrm{kHz}, m=0.7, \cos\varphi=0.9)\) [2212.05246].

A complementary behavioral approach is used for phase-shifted full-bridge inverter modules. There, total switching loss is decomposed as
\[
P_{sw} = P_{bd} + P_{gt} + P_{ov},
\]
and the final compact behavioral model identified by GP and NSGA-II is
\[
P_{sw,bhv} = p_0 f_s V_{in} D(1-p_1 D)\frac{1}{R_T} + p_2 f_s.
\]
The training set spans \(1215\) data vectors over \(f_s\), \(V_{in}\), \(D\), \(R_T\), \(V_{dr}\), and \(R_g\). For the highlighted model, the paper reports mean relative error \(H_{err}=0.4\%\), standard deviation \(T_{err}=3\%\), and maximum error \(err_{max}=12\%\), with errors remaining within \(\pm12\%\) over the studied operating range [1906.04428].

Measurement methodology has become a distinct subtopic because wide-bandgap devices complicate loss separation. The hybrid single-pulse plus Sawyer-Tower method writes
\[
E = E_{\mathrm{on}} + E_{\mathrm{off}} = E_{\mathrm{charge}} + E_{\mathrm{overlap}},
\]
with overlap loss
\[
E_{\mathrm{overlap}} = \int_0^{t_{\mathrm{off}}} V_{\mathrm{ds}}(t)\, I_{\mathrm{ds},f}(t)\, dt,
\]
and under soft-switching conditions
\[
E_{\mathrm{off}} - E_{\mathrm{on}} = E_{\mathrm{hysteresis}} + E_{\mathrm{overlap}}.
\]
For the GaN device tested in a \(400\ \mathrm{V}\), \(250\!-\!350\ \mathrm{kHz}\), \(350\ \mathrm{W}\) LLC converter, the paper reports \(E_{\mathrm{off}}=6.889\ \mu\mathrm{J}\) and \(C_{\mathrm{oss}}\) hysteresis loss \(E_{\mathrm{hysteresis}}=0.189\ \mu\mathrm{J}\), with agreement to the converter-based reference within \(40\ \mathrm{nJ}\) according to the conclusion [2506.18635].

Time-segmented analytical modeling addresses the same problem from a waveform perspective. The SiC MOSFET plus SiC SBD study represents switching loss as the time integral of instantaneous power for MOSFET and diode separately,
\[
E_{\text{loss}(\text{MOSFET})}=\int V_{ds}(t)\, i_a(t)\, dt,\qquad
E_{\text{loss}(\text{SBD})}=\int V_F(t)\, i_f(t)\, dt,
\]
and decomposes turn-on into eight stages and turn-off into five. The model emphasizes parasitic inductance \(L_{\text{stray}}=L_s+L_d+L_p\), piecewise-linearized nonlinear capacitances, and SBD junction-capacitance displacement current rather than silicon-style reverse recovery. With varying \(R_G\) and varying external \(C_{gd}\), the reported total loss error remains below about \(6\%\) in the shown cases [2311.07029].

Across these papers, switching loss is not a lumped constant. It is current-dependent, topology-dependent, gate-drive-dependent, and often measurement-method-dependent.

## 6. Information loss under the quantum switch and objective switching in neural PDEs

The quantum-switch literature uses “loss” in an information-theoretic sense. For two initial states \(\rho_1(0)\) and \(\rho_2(0)\) evolving under \(\Phi_t\), information loss is defined as
\[
\Delta \mathcal{I}(\rho_1(t),\rho_2(t)) \equiv D(\rho_1(0), \rho_2(0)) - D(\Phi_t(\rho_1(0)), \Phi_t(\rho_2(0))).
\]
For an ergodic channel with fixed point \(\tau\), the switched and unswitched losses are compared through a switch-induced memory
\[
\mathcal{Q}_S(t)\equiv D(\Phi_t^S(\rho),\Phi_t^\epsilon(\rho)),
\]
and the paper proves the uncertainty-like relation
\[
\Delta \mathcal{I}_S(\rho(t)) + \mathcal{Q}_S(t) \ge \Delta \mathcal{I}_\epsilon(\rho(t)).
\]
For a depolarizing example, the reduced switched dynamics obeys a Lindblad-type master equation with a time-dependent rate \(\Gamma_S(t)\) that becomes negative after
\[
T_-=\frac{1}{4\gamma}\ln(2\sqrt{3}+3),
\]
which the paper interprets as emergent non-Markovianity. In this setting, the switch reduces effective information loss by creating a memory-like contribution rather than lowering insertion loss or packet loss [2307.01964].

A distinct but related lexical use appears in neural PDE training, where the “switch” is a change of loss function rather than a physical device. The paper separates
\[
L_{\mathrm{data}}(\theta)=\frac{1}{N}\sum_{i=1}^{N}\|u_\theta(x_f^i,t_f^i)-u(x_f^i,t_f^i)\|^2
\]
from model-based objectives such as
\[
L_{\mathrm{model}}(\theta)=\sum_{k=0}^n \frac{\lambda_k}{N_k}\sum_{i=1}^{N_k} \|u_\theta^{(k)}(x_f^i,t_f^i)-u^{(k)}(x_f^i,t_f^i)\|^2.
\]
Its central empirical observation is a stable loss-jump phenomenon: when switching from data loss to model loss, the neural network solution significantly deviates from the exact solution immediately, even under small learning rates. The paper documents this for Poisson, Burgers, heat, diffusion, and wave equations, and attributes the phenomenon to different frequency preferences induced by data loss and model loss [2405.03095].

Taken together, these works show that “switch loss” can refer either to a dissipative cost of routing or commutation, or to a degradation of distinguishability or objective value caused—or altered—by switching structure itself. The precise meaning is therefore inseparable from domain, observable, and mathematical model.

Source: https://www.emergentmind.com/topics/switch-loss