Mixers Q&A

A mixer multiplies two signals — RF and LO — to produce outputs at their sum and difference frequencies. This is frequency translation.

In a receiver, 2400 MHz RF mixes with 2300 MHz LO to produce 100 MHz IF — much easier to filter and amplify. Real mixers use a nonlinear element (diode, FET, transistor quad) to generate sum/difference products. A true linear multiplier would work but is impractical at GHz frequencies.

The 3.92 dB limit comes from the Fourier series — the fundamental of a ±1 square wave is 2/π amplitude. Typical real values: single-balanced 5–7 dB, double-balanced ring 6–8 dB, active Gilbert cell −3 to +5 dB (can have conversion gain).

Both f_RF and f_image produce the same IF output — indistinguishable after downconversion. Rejection methods: (1) image reject filter (preselector) before the mixer, (2) higher IF moves image further away making filtering easier, (3) IQ image-reject mixer cancels image using 90° phase relationship.

Always use SSB NF in the Friis cascade for a superheterodyne receiver. If a datasheet specifies DSB NF, add 3 dB before plugging into Friis. Confusing the two causes a systematic 3 dB budget error — extremely common in interviews and real designs.

LO-to-RF isolation: How much the LO signal is attenuated appearing at the RF input port. Typical: 20–40 dB. Poor isolation causes LO re-radiation through the antenna, creating interference and violating spurious emission regulations. Critical in transmitters.

LO-to-IF isolation: How much LO appears at the IF output. Poor isolation puts a strong LO spur directly in the IF passband — especially damaging in direct-conversion where LO falls at DC/baseband. Double-balanced mixers achieve 20–40 dB better isolation than single-balanced designs.

At high RF input levels the diodes spend increasing time in a nonlinear region — the RF signal is large enough to disturb the LO-driven switching threshold. The diodes no longer switch cleanly, conversion efficiency drops, gain compresses. Higher LO drive makes it harder for the RF to perturb the switching threshold → higher P1dB. Every 1 dB more LO power gives ~1 dB more IIP1dB.

The half-IF spur is second-order (m=2, n=2) — requiring 2nd harmonics of both RF and LO. A double-balanced mixer suppresses even harmonics, greatly reducing it. It cannot be moved by choosing a different IF — it is always at f_LO + f_IF/2, making preselector filtering and high IIP2 the only solutions.

Unbalanced (1 diode): No cancellation. All products m×RF ± n×LO present at IF. Poor isolation. Lowest cost.

Single-balanced (2 diodes + 1 balun): Even harmonics of one port cancel at IF. LO-to-IF or RF-to-IF isolation improved by 15–25 dB. Common in balanced detectors.

Double-balanced ring (4 diodes + 2 baluns): Even harmonics of both RF and LO cancel. Only odd×odd products survive: (1,1),(1,3),(3,1)… Excellent isolation (30–60 dB) at all ports. Standard for most RF applications.

Triple-balanced (8 diodes + 3 baluns): Even IF harmonics also cancel. Extremely wideband. Used in test equipment and EW receivers requiring the widest dynamic range.

The Gilbert cell uses a differential transconductance pair (RF port → differential current) and a four-transistor switching quad (steered by LO) that multiplies the current by ±1 at the LO rate. The IF current is converted to voltage by load resistors.

Advantages: Conversion gain (+5 to +15 dB), fully differential (excellent isolation), integrable in CMOS/HBT without transformers, low LO drive power.

Disadvantages vs diode ring: Lower IIP3, higher NF (10–20 dB vs 6–8 dB for DBM), requires DC power (5–20 mW), significant 1/f noise (damaging in direct-conversion), lower maximum operating frequency.

Used at mmWave (60 GHz, 77 GHz, 300 GHz) where generating a high-power clean LO at the full RF frequency is impractical. Tradeoffs: 5–10 dB more conversion loss than fundamental, higher LO drive needed, more complex spur analysis (products at m×f_RF ± n×2×f_LO).

A single mixer loses phase information and cannot distinguish signals above vs below the LO. The IQ mixer's quadrature path provides the complete complex baseband representation, enabling: image rejection (signals above/below LO distinguishable by I/Q phase relationship), direct demodulation of QAM/OFDM, and zero-IF operation.

Amplitude imbalance makes the signal constellation elliptical; phase imbalance rotates points asymmetrically. Both degrade image rejection and EVM. 5G NR 256-QAM requires EVM < 3.5%, demanding phase imbalance <0.3° — achievable only with digital calibration that measures imbalance and applies inverse correction in DSP.

29.3 dB hardware IRR is insufficient for LTE (requires 40+ dB). Digital IQ correction applied at startup typically improves IRR to 50–70 dB by measuring the imbalance coefficients and applying the inverse in DSP.

f_LO = f_RF — signal converts directly to baseband in one IQ mixing step. No IF filter needed.

Advantages: No image frequency problem. No IF SAW/crystal filter (saves cost and PCB area). Fully integrable in CMOS. Simpler architecture. Channel selection done by software-programmable baseband LPF.

Specific failure modes:

• DC offset: LO self-mixing produces DC at baseband, saturating ADC. Solved by adaptive cancellation or AC coupling.
• 1/f (flicker) noise: Transistor flicker noise peaks at DC — exactly where the wanted signal now lives. Solved by PMOS switches or chopping techniques.
• IQ imbalance: Gain/phase mismatch causes image interference and EVM degradation. Solved by digital calibration.
• IIP2: IM2 from two blockers falls at their difference frequency — in-band. Requires IIP2 > +50 dBm.
• LO pulling: TX signal in FDD can injection-lock or frequency-pull the VCO.

Superheterodyne: Best selectivity and sensitivity, needs off-chip IF filter. Best performance. Use for: base stations, radar, test instruments, GPS — anywhere performance beats cost and size.

Direct conversion: Best integration, lowest cost, DC offset/1/f/IIP2 problems requiring digital solutions. Use for: handsets, WiFi, Bluetooth, IoT — mass-market devices where integration and cost dominate.

Low-IF: IQ mixing to a small IF (1–10 MHz). Avoids DC offset and 1/f noise problems. Image is the adjacent channel — rejected by IQ balance (>35 dB required). Use for: Bluetooth classic, ZigBee, DECT — standards where DC offset is a problem but full direct-conversion is difficult.

A strong interferer at f_LO + f_IF/2 has its second harmonic at 2f_LO + f_IF. Mixing with the second harmonic of the LO gives exactly the IF. This cannot be moved by changing IF — it is always at f_LO + f_IF/2.

Suppression: High IIP2 (double-balanced structure cancels even harmonics), preselector filter attenuating f_LO+f_IF/2, higher IF frequency, fully balanced LO drive eliminating even LO harmonics.

Hartley: RF → IQ mixing → 90° phase shift on one IF path → add. Image cancels because the 90° IF shift reverses its relative phase. Limitation: the 90° IF phase shifter (Hilbert filter or polyphase network) must be accurate across the full IF bandwidth — difficult to achieve over >20% bandwidth. Sensitive to amplitude, phase AND IF phase shifter accuracy simultaneously.

Weaver: First IQ stage → first IF → second IQ stage at f_LO2=f_IF1 → subtract. Replaces the wideband 90° IF phase shifter with a second mixing stage. Limitation: introduces a second image at f_LO1±f_LO2±f_IF2 that must also be rejected. Requires two LO frequencies.

Too low LO: Diodes not fully switched — conversion loss rises 3–8 dB, NF degrades equally, IIP3 drops dramatically. The mixer enters a "partial switching" regime generating many extra spurs.

Too high LO: Diodes saturate — no further performance gain. Excess power heats the package. LO harmonics become stronger. Going 3–6 dB above the rated level is generally safe and sometimes slightly improves IIP3.

A strong interferer at frequency offset Δf from the wanted signal mixes with the LO phase noise skirt at that same offset Δf, down-converting noise energy into the IF band. This raises the effective noise floor above thermal, desensitising the receiver even if the interferer is outside the IF filter passband.

In a zero-IF receiver, IM2 from two adjacent-channel blockers falls directly in the baseband signal band, unlike IIP3 products which can be moved by adjusting blocker spacing. LTE requires operation with −15 dBm blockers: IM2 = 2×(−15)−50 = −80 dBm — just at the thermal noise floor for a 20 MHz channel, requiring IIP2 > +50 dBm.

With high-side injection, a signal at a higher RF frequency appears at a lower IF frequency — the IF spectrum is a mirror image. This must be corrected in demodulation (SSB sidebands swap, FM stereo pilot position inverts). Low-side injection is generally preferred to avoid this complication.

An upconverter translates a baseband or IF signal to a higher RF frequency for transmission. The same hardware works — a mixer is reciprocal. Output contains both sum and difference; a BPF selects the wanted upper sideband at f_LO + f_IF.

Key differences in TX use: Much higher signal levels (+0 to +20 dBm). Linearity and spectral purity (spurious emissions) are paramount. Noise figure is irrelevant. LO leakage at the RF port contributes to spurious emissions violating regulatory masks.

Amplitude variations of the RF input cause phase rotations at the IF output due to nonlinear phase response of the switching transistors or diodes. OFDM signals (WiFi, LTE, 5G NR) have 8–12 dB PAPR — large instantaneous amplitude swings. Even small AM-PM coefficients produce significant EVM at peak amplitude moments.

Passive CMOS uses four NMOS switches with no DC bias — either fully ON (low resistance) or OFF. No DC current means zero 1/f noise contribution — critical for direct-conversion where the signal lives near DC. Followed by a TIA (transimpedance amplifier) in "current-mode" architecture dominant since ~2010.

In FDD (TX and RX simultaneously on different frequencies), TX power leaks through the duplexer (−20 to −40 dBm at RX LNA input). This TX leakage mixes with RX LO phase noise at the duplex frequency offset, creating a noise component in the IF band that raises the effective noise floor and reduces sensitivity.

Setup: f1=900.1 MHz, f2=900.2 MHz, LO=800 MHz → IF fundamentals at 100.1 and 100.2 MHz, IM3 at 100.0 and 100.3 MHz. Signal generators must have >30 dB isolation between them (use a hybrid combiner) to prevent mutual intermodulation.

A large interferer pushes the mixer toward its P1dB. In compression, the weak wanted signal is converted with the same reduced gain as the strong blocker — its IF output power drops by ΔG dB, reducing SNR and degrading sensitivity. The receiver appears deaf even though the wanted signal is physically present.

A true linear multiplier produces only sum and difference — ideal mixer behaviour. The problem: building a linear multiplier at GHz frequencies is impractical. Diode/transistor mixers exploit intentional nonlinearity to generate mixing products, accepting harmonic spurs in exchange for high-frequency operation.

Beyond simple power transfer loss, impedance mismatch at mixer ports has three additional effects:

• Conversion loss sensitivity: Passive diode mixer CL is sensitive to RF port VSWR — even 2:1 VSWR can add 1–2 dB excess loss beyond the resistive mismatch alone
• Image-band noise: Poor RF port match at the image frequency reflects image-band thermal noise back into the mixer, adding up to 3 dB to the noise figure
• IF port termination: Passive mixers behave as a current source — the IF load impedance affects conversion efficiency. Improper IF termination degrades specified conversion loss

A polyphase filter is a passive RC ring network that accepts a differential input and produces four outputs with 90° phase separation. Accuracy is excellent at exactly ω=1/RC but degrades away from this frequency. Multi-stage designs extend bandwidth.

Preferred method in modern CMOS: Run VCO at 2×f_LO, divide by 2 with a D flip-flop. Produces inherently 50% duty cycle and exact 90° quadrature. Also reduces VCO phase noise by 6 dB (20·log₁₀(2)).

The mixer contributes only 0.127 noise factor units — less than 0.4 dB. The LNA's 15 dB gain suppresses the mixer's noise by a factor of 31×. Without the LNA: NF = 7 dB. With it: 1.88 dB — a 5.12 dB sensitivity improvement.

SFDR is the input dynamic range over which the mixer operates without IM3 products rising above the noise floor. Below SFDR: signal detectable. At SFDR: IM3 = noise floor. Above SFDR: IM3 dominates. The (2/3) factor comes from the 1:1 fundamental slope vs 3:1 IM3 slope meeting at the IIP3 extrapolation point.

LO: f_LO = 2300 MHz (low-side injection). Image at 2200 MHz — preselector must attenuate by ≥50 dB.

Mixer selection: Active CMOS IQ Gilbert cell for integrated WiFi chip. At −65 dBm input level, well below compression.

This design is mixer-dominated in IIP3. To improve: reduce LNA gain by 3 dB → IIP3 improves to ~−13 dBm, NF degrades to ~3.2 dB. This NF-linearity tradeoff is the fundamental LNA design optimisation problem.

Linearity limited — the IIP3 of −9.8 dBm is poor relative to the excellent NF of 1.35 dB. The mixer's IIP3 input-referred through the 18 dB LNA dominates (−8 dBm contribution vs LNA's −5 dBm). Fix: reduce LNA gain 3 dB → IIP3 improves to ~−7.5 dBm, NF degrades to ~1.8 dB. This gain-NF-linearity tradeoff is the heart of LNA design.