1. Introduction — Why the Power Supply Sets the RF Floor

An RF system’s sensitivity is limited by its noise floor, and the noise floor is limited by the power supply. Supply noise couples into oscillators and produces phase noise; it modulates the bias of amplifiers and mixes with the signal to produce spurs; it appears on the reference of an ADC or DAC and degrades the effective number of bits; and it couples through shared ground and supply impedances between stages that were designed as if they were independent. A receiver specified at −110 dBm sensitivity can lose 10 dB of it to a switching regulator running at 400 kHz with insufficient filtering. The good news is that supply-induced degradation is a predictable, calculable and verifiable engineering problem: once the noise spectrum, the conversion mechanism (PSRR, pull/sensitivity, or the impedance of the supply network) and the allowable contribution to the system budget are quantified, the decoupling and filtering design follows. This guide develops the phase-noise relationship, introduces the power distribution network (PDN) impedance target, explains the frequency-dependent behaviour of decoupling capacitors and the self-resonance that decides the network’s topology, covers the filtering and layout measures that actually reduce coupling, and works a complete example of a synthesizer and receiver front end sharing a supply. It complements the antenna matching, RF attenuation and decoupling capacitor material on this site.

2. How Supply Noise Becomes RF Degradation

Three coupling mechanisms dominate:

Mechanism	Victim	How it manifests	Metric
Oscillator bias/varactor modulation	VCO, TCXO, PLL reference	Phase noise sidebands at the supply noise frequency	L(f) in dBc/Hz
Amplifier bias modulation	LNA, PA, mixer	Gain modulation → AM-to-PM conversion, spurs, spectral regrowth	PSRR, pull figure
Shared impedance coupling	All stages via a common rail or ground	One stage’s currents appear as noise in another’s supply	PDN impedance vs frequency

The oscillator case is the most direct. The sideband phase noise caused by a supply disturbance is:

L(f_m) = 20 · log10( K_v · S_v(f_m)^0.5 / (2 · f_m) ) [dBc/Hz]

where K_v is the oscillator’s supply pushing figure (Hz/V), S_v is the power spectral density of the supply noise (V²/Hz) at the offset frequency f_m, and the factor 1/(2f_m) comes from the FM-to-phase-noise conversion. The engineering consequence is dramatic: an oscillator with a 10 MHz/V pushing figure and 100 nV/√Hz of supply noise at 10 kHz offset produces L(10 kHz) = 20 log(2.5 × 10⁻⁵ × 10⁷ / (2 × 10⁴)) ≈ −72 dBc/Hz from the supply alone — enough to dominate a good oscillator’s own noise floor. Note the two levers: reduce the pushing figure (choose the oscillator topology and its supply rejection) or reduce the supply noise density (filtering, low-noise regulation). Both are needed in practice, and the pushing figure is frequently the easier one to specify in a datasheet. Amplifiers have a similar but less predictable relationship: their PSRR falls with frequency, often from 60–80 dB at DC to 20–30 dB at 1 MHz, so a switching regulator at 2 MHz can inject more than the DC-noise calculation suggests. The most useful system-level discipline is to allocate a noise budget: decide how much of the total phase-noise and spur budget the supply may consume (typically 10–20%, i.e. 10–7 dB below the oscillator’s own contribution), then derive the allowable supply noise density at each offset frequency.

3. The PDN Impedance Target

The power distribution network is analysed as an impedance: the voltage noise seen by a load is the product of the load’s current spectrum and the impedance of the network that supplies it:

V_noise(f) = Z_PDN(f) · I_load(f)

For a digital or mixed-signal load, the design target is usually stated as a maximum impedance over a frequency band: Z_max = (allowable ripple voltage) / (transient current). For an RF load the requirement is stated in the frequency domain: the PDN impedance must be low enough at each offset frequency that the resulting supply noise density meets the phase-noise budget. The PDN is a cascade of the regulator’s output impedance, the bulk capacitance, the network of decoupling capacitors, the plane capacitance and the package/lead inductance. It has resonances: the inductance of the connection to the bulk capacitor resonates with the die and package capacitance; the decoupling capacitors’ self-inductance (ESL) limits their high-frequency effect and creates anti-resonances with each other. The practical design method:

• Start Low to High: the regulator’s control loop sets the impedance up to its crossover (typically tens of kHz); the bulk capacitor takes over from the crossover to a few hundred kHz; the local decoupling network covers hundreds of kHz to tens of MHz; the plane and package cover above that.
• Use Different Capacitor Values Sparingly: mixing decades of capacitance (100 µF, 10 µF, 1 µF, 100 nF, 10 nF) creates multiple anti-resonances where the inductance of one value resonates with the capacitance of another, producing high-impedance peaks between the capacitors’ self-resonances. A better strategy is a few well-chosen values with known ESR/ESL, plus plane capacitance, placed to keep the impedance below the target across the band.
• Placement Beats Value: a 100 nF capacitor 10 mm from the load pin has an inductive path (about 8 nH/mm of trace plus via inductance) that dominates its effect above a few MHz. The capacitor’s job is to supply the current locally, so its loop area — capacitor to the pin, and the return path — must be minimised. Two closely placed 100 nF capacitors can outperform a single 10 µF placed further away.
• Self-resonance: a 0402 100 nF capacitor has roughly 1.2 nH of ESL, giving a self-resonance near 14 MHz; above that frequency it is an inductor. This is why the network’s topology must be designed in the frequency domain rather than by adding capacitance in bulk; the decoupling capacitor calculator is the right tool for setting the value from the load’s transient requirement, and the layout rules above then determine whether that value is actually effective.

4. Decoupling and Filtering for RF Stages

Three measures, in increasing order of effectiveness and cost:

Measure	Attenuation	Notes
Local decoupling at each supply pin	10–20 dB above a few MHz	Mandatory; the first line of defence, effective only with short loops
Series ferrite bead or small resistor per stage	10–20 dB in the bead’s range	Creates isolation between stages; the bead must be chosen for the current and must not resonate with the stage’s decoupling
LC or RC low-pass filter per sensitive stage	20–40 dB with a proper design	Multiple stages (a two-section filter) for synthesizers and VCOs; watch the filter’s insertion loss and the regulator’s loop stability
Low-dropout regulator (LDO) with high PSRR	20–60 dB, frequency dependent	Best used after a switching regulator, cascaded for very sensitive rails; PSRR falls above ~100 kHz
Separate rails per functional block	Removes shared-impedance coupling	Effective at the system level; the highest-value architectural decision

Several practicalities decide whether these measures work. A ferrite bead’s impedance is specified at 100 MHz and can be nearly resistive at the operating frequency; check the bead’s Z(f) curve, its DC current rating (saturation and heating both reduce its impedance) and its tolerance. An RC filter’s resistor drops the rail voltage and its noise contributes directly at the load — a 10 Ω resistor with 10 nV/√Hz of thermal noise is 12.8 nV/√Hz of supply noise, which may be worse than the regulator it is filtering; for sensitive oscillators an LC filter with a low-DCR inductor is usually better. A filter’s resonant peak must be damped (a Q of 10 without damping amplifies the regulator’s own control-loop noise at the resonance). Cascading an LDO after a switching regulator is the standard way to obtain a quiet rail; the LDO’s PSRR must be checked at the switching frequency and its output capacitor chosen above the LDO’s stability limit. Finally, the LDO’s own noise (typically 10–100 µV RMS over 10 Hz–100 kHz) sets a floor that no amount of downstream decoupling can lower, so choosing a low-noise LDO is a design decision, not a detail. Where the rail powers a mixed-signal load, the filter design is the same problem as a general EMI filter; the component sizing follows the method in the EMI filter article, and the filter’s attenuation is verified with the EMI filter calculator. The capacitor values that set the filter’s corner and the decoupling network’s impedance are chosen with the decoupling capacitor calculator.

5. Layout, Grounding and Shielding

At RF, the layout is part of the circuit. The measures that matter most, in the order they usually determine success:

• One uninterrupted ground plane under the RF section: every decoupling capacitor, filter and stage returns to the same plane with the shortest possible loop; slots, splits and long return paths under RF traces turn the ground into a radiator.
• Separate the analogue/RF ground from the digital/switching ground at the layout level, join at one point: the join must be at a location where the digital return currents do not flow under the RF circuitry.
• Keep switching-regulator loops tiny: the input-loop (input capacitor to switch) and the output-loop (switch to inductor to output capacitor) are the primary radiators and the primary source of ground bounce; their area is the single most important layout parameter in a mixed-signal board.
• Place the regulator away from the RF section and orient its loops away from sensitive nodes: distance plus the ground plane’s attenuation is usually enough for the electric-field coupling; magnetic coupling from the inductor is handled with distance and orientation.
• Decouple the reference and bias nodes as carefully as the supply: a PLL’s loop filter and reference input are as sensitive as an oscillator’s supply, and a noisy reference degrades phase noise directly.
• Shielding: for synthesizers and receivers, a shield can (or a shielded module) with a good ground connection addresses the near-field coupling that layout alone cannot fix; the shield’s effect on temperature and on the oscillator’s own stability must then be considered.
• Keep the return current close to the trace: a microstrip over a plane has a return path directly beneath, while a trace with a broken plane has a return path that loops around the break — a loop that radiates and picks up noise.

6. Measurement and Verification

Supply-induced degradation is verifiable, and the verification is what makes the design credible. Techniques, in order of value:

• Measure the supply noise with a low-inductance probe: a coaxial connection with a small series resistor at the point of load, or a purpose-made power-rail probe, avoids the probe’s own loop picking up magnetic fields. Never use a standard 10:1 passive probe with a long ground lead for a rail measurement near a switching converter.
• Measure the phase noise with and without the supply perturbation: the difference isolates the supply’s contribution and validates the budget. For oscillators, add a known tone to the supply (via a bias tee) and observe the resulting sidebands to measure the pushing figure and the conversion gain directly.
• Measure the PDN impedance: a vector network analyser with a two-port shunt-through measurement gives Z(f) from the regulator’s crossover to hundreds of MHz; this is the most direct way to find the anti-resonance peaks that a decoupling change has introduced.
• Inject a serial tone and look for spurs: for a receiver or transceiver, injecting a small tone on the supply and sweeping its frequency while monitoring the output identifies the frequencies where the design is fragile — often the switching regulator’s harmonics and the crystal’s harmonics.
• Test at the worst-case operating point: supply noise effects frequently depend on the load current (an LDO’s PSRR falls at high current; a switching regulator’s ripple rises), on temperature and on the channel/frequency setting.

The two calculations worth automating are the impedance target and the pushing-figure conversion; both are simple enough to run as a script during the design review, and both catch the errors that are otherwise found only in the lab.

import math

# 1. Target impedance for a rail, from the allocated ripple and load
def z_target(v_allow_ripple_v, i_load_a, margin_db=6.0):
    """PDN impedance ceiling over the band of interest."""
    return v_allow_ripple_v / i_load_a / (10 ** (margin_db / 20.0))

# Example: 10 mV of allowed ripple on a 200 mA rail, 6 dB of margin
print("Z_target = %.1f mOhm" % (z_target(10e-3, 0.2) * 1e3))

# 2. Supply-induced phase noise from the pushing figure
def pushing_phase_noise_db(f_push_hz_per_v, v_ripple_v, f0_hz, f_offset_hz):
    """Sideband level (dBc) from a sinusoidal supply ripple, f_offset away."""
    df = f_push_hz_per_v * v_ripple_v          # peak frequency deviation
    beta = df / f_offset_hz                    # modulation index
    # narrowband FM: sideband power relative to carrier ~ (beta/2)^2
    return 20 * math.log10(beta / 2.0 + 1e-30)

# 8 MHz/V pushing figure, 1 mV of ripple at the offset of interest
f0, off = 2.4e9, 10e3
print("sideband = %.1f dBc" % pushing_phase_noise_db(8e6, 1e-3, f0, off))

Three points about using these numbers. First, the impedance target is only meaningful over a stated bandwidth: with the load current stepped at the regulator’s crossover frequency, the rail must hold the ripple down to that frequency, while at higher frequencies the decoupling network takes over and the requirement becomes the capacitance’s impedance, not the regulator’s loop. Second, the pushing-figure formula assumes a sinusoidal perturbation and narrowband FM, so for a broadband noise rail the correct procedure is to integrate the supply noise density weighted by the pushing figure and the loop’s transfer function, which yields a phase-noise density rather than a single sideband. Third, both calculations should be repeated at the worst-case load and temperature, because an LDO’s PSRR and a buck’s ripple both degrade at high current — the design must hold at the corner, not only at the nominal point.

7. Worked Example — Synthesizer and Receiver Front End on a Shared Rail

Target: a 2.4 GHz synthesizer and receiver front end, a 5 V input with a 3.3 V rail derived by a 2 MHz switching regulator, the VCO’s pushing figure 8 MHz/V, the system’s phase-noise budget allowing 3 dB of degradation from the supply at 10 kHz and 100 kHz offsets, the receiver’s sensitivity budget allowing −7 dBm of supply-induced spurs.

• Budget allocation: allow the supply to contribute 10% of the phase-noise power (i.e. 10 dB below the oscillator’s own noise) at each offset. For an oscillator with L(10 kHz) = −95 dBc/Hz, the supply’s share must be below −105 dBc/Hz.
• Required supply noise: from L = 20 log(K_v S_v^0.5/(2 f_m)), solve for S_v: S_v^0.5 = 2 f_m × 10^(L/20) / K_v = 2 × 10⁴ × 10^(−105/20) / 8 × 10⁶ = 2 × 10⁴ × 5.6 × 10⁻⁶ / 8 × 10⁶ ≈ 1.4 × 10⁻⁸ V/√Hz = 14 nV/√Hz. At 100 kHz the limit is 140 nV/√Hz (the conversion falls as 1/f_m). This is a demanding but achievable target with an LDO and filtering.
• Architecture: the 2 MHz switcher produces the 3.3 V rail; the VCO and PLL get their own LDO (low-noise, PSRR ≥ 60 dB at 100 kHz) plus an LC filter (1 µH, 10 µF ceramic plus 100 µF bulk) with a damping resistor across the inductor (a Q of ~1) to avoid a resonant peak; the LNA and mixer get a second LDO; the digital section stays on the switcher’s rail. Separate rails are the architectural decision that removes the shared-impedance coupling before any filtering is considered.
• Decoupling: 100 nF at every supply pin with a loop under 2 mm, 1 µF per stage, 10 µF per rail section; bulk 100 µF at the rail entry. The decoupling capacitor calculator gives the value needed from the load’s transient current and allowable ripple (for the VCO’s 20 mA step, a 10 µF with 5 mΩ ESL delivers a 100 mV transient — filter and LDO response must cover the rest).
• Filter validation: EMI filter calculator gives the attenuation of the two-section LC at 2 MHz and its harmonics; simulate with the LDO’s output impedance as the source to confirm the resonant peak is damped and does not amplify the switcher’s ripple at the crossover.
• Layout: the switcher’s input and output loops under 20 mm²; the LDOs placed with their output capacitors adjacent; a single-point ground join between the digital and analogue sections located away from the RF traces; the VCO’s supply trace with no parallel high-current paths.
• Verification: rail noise measured with a coax probe at the VCO’s pin (target <20 nV/√Hz at 10 kHz when the switcher is running); phase noise measured with the switcher on and off (the difference must be under the allocated 3 dB); a 2 MHz tone injected on the rail with the receiver’s output monitored for spurs; the measurement repeated at the extremes of the temperature range and at maximum load current.

8. Common Mistakes

• Relying on the LDO’s DC PSRR: PSRR falls rapidly above its crossover (often 10–30 dB per decade); the switching frequency’s attenuation must be read from the PSRR-versus-frequency curve.
• Adding more capacitor values without analysing anti-resonances: the resulting high-impedance peaks can make the PDN worse than a simpler network.
• Long ground returns on the decoupling capacitors: the via and trace inductance dominates the capacitor’s effect above a few MHz.
• Sharing one rail between the power amplifier and the low-noise receiver: the PA’s current modulation appears in the receiver’s supply; separate rails or at least a series isolation element with careful filtering are required.
• Ignoring the switcher’s layout loops: an otherwise well-designed filter cannot fix the ground bounce generated by a large input loop.
• Measuring rail noise with a long-ground passive probe: the measurement then shows the probe’s loop pickup, not the rail.
• Designing to the nominal load current only: an LDO’s PSRR and a switcher’s ripple both depend on the load; test at the extremes.

9. FAQ

Q: Can a switching regulator ever be good enough for an RF rail? A: With a low-ripple topology, a high switching frequency, excellent layout, a second-stage filter and a low-noise LDO, yes for many receivers; oscillators and synthesizers usually still benefit from a dedicated LDO and filter chain.

Q: How much decoupling is enough? A: The answer is an impedance target, not a capacitance value: design the PDN so that Z(f) stays below V_allowable/I_load across the frequency range of interest, and verify the impedance from the layout’s parasitics rather than by counting capacitors.

Q: Does a bigger bulk capacitor always help? A: Only below its self-resonance and only if the connection inductance is small; above self-resonance it is an inductor, and a large capacitor far from the load may do nothing at the frequencies that matter.

Q: How do I measure an oscillator’s supply pushing figure? A: Inject a small, known AC voltage on the supply through a bias tee, measure the resulting frequency deviation (or the sideband level) at a low offset, and divide the frequency shift by the voltage — the result in Hz/V is the pushing figure at that frequency.

Q: Should the RF section have its own ground plane? A: A single, unbroken plane is best in almost all cases; splitting the plane creates return-path discontinuities and makes the two halves into an unintentional antenna unless the join is carefully engineered.

Q: Is a linear regulator always the right answer for an RF stage? A: No. An LDO rejects low-frequency ripple well but its rejection falls with frequency — 30 dB at 10 kHz and often under 20 dB above 1 MHz is typical — so VCO and PLL rails generally need an LDO followed by an LC filter. The LDO’s own output noise, its dropout and its dissipation must be checked too; for a large voltage drop at significant current, a switching pre-regulator followed by an LDO is the lower-loss arrangement.

Q: How low does the PDN impedance actually need to be? A: Low enough that the supply ripple at the frequency of interest consumes less than the noise share you allocated to the supply — typically a target in the low tens of milliohms near the loop bandwidth and lower still at the carrier offsets that matter, with the caveat that inductance rather than capacitance dominates above a few megahertz.

Q: Why does the phase noise improve when I touch the board? A: Almost always a grounding or decoupling resonance issue: the finger adds loss at the resonant frequency, which means the network is under-damped. Fix it by narrowing the capacitor spread and adding deliberate damping, not by adding mechanical contact.

10. Conclusion

Supply noise degrades RF performance through three well-understood mechanisms, and each can be predicted and measured: the oscillator’s pushing figure times the supply noise density, the amplifier’s frequency-dependent PSRR, and the shared-impedance coupling described by the PDN. Allocate the supply’s share of the system noise budget first, then design the network from low frequency to high (regulator, bulk, decoupling, planes and package), use separate rails for the sensitive blocks, keep every decoupling and switching loop small, and verify with a low-inductance rail measurement, a phase-noise comparison and a PDN impedance sweep. An RF design that ignores its power supply will work on the bench and fail in the field; one that treats the supply as part of the RF chain has a predictable, repeatable noise floor.