Phase Shift Keying (PSK)

All PSK constellation points reside on a single circle centered at the origin of the IQ plane, meaning the instantaneous signal amplitude remains constant. This is the defining geometric signature that distinguishes PSK from Quadrature Amplitude Modulation (QAM).

• Power Efficiency: Allows the use of non-linear, high-efficiency power amplifiers (like Class C) without causing spectral regrowth or distortion.
• Robustness: Immune to amplitude-based channel impairments and amplifier non-linearities.
• Detection: Enables the use of the Constant Modulus Algorithm (CMA) for blind equalization, as any amplitude variation is purely caused by the channel.

A fundamental challenge in PSK demodulation is phase ambiguity. Blind carrier recovery circuits can lock to the wrong phase offset, causing a fixed rotation of the entire constellation. For example, in QPSK, a 90° rotation maps symbols to themselves, making absolute decoding impossible without a reference.

• Differential PSK (DPSK): Solves this by encoding data in the phase difference between successive symbols, not the absolute phase. DBPSK and DQPSK eliminate the need for a coherent phase reference.
• Unique Words: Non-differential systems insert known pilot symbols to resolve the absolute phase rotation.
• Classification Impact: Modulation classifiers must be invariant to this fixed rotation, often using higher-order cumulants which are naturally phase-blind.

The modulation order M in M-PSK determines the number of bits per symbol (log₂M). Higher orders pack more bits into each transmission, increasing spectral efficiency, but at the cost of reduced noise immunity.

• BPSK (M=2): 1 bit/symbol. Two points at 0° and 180°. Most robust, used in low-SNR links and control channels.
• QPSK (M=4): 2 bits/symbol. Four points at 45°, 135°, 225°, 315°. The workhorse of satellite and LTE uplink.
• 8-PSK (M=8): 3 bits/symbol. Used in EDGE (2G) systems. The phase distance between points shrinks to 45°, requiring higher SNR.
• Beyond 8-PSK: 16-PSK and higher are rarely used in practice because QAM or APSK offer better power and spectral efficiency trade-offs.

PSK constellations almost universally employ Gray coding for bit-to-symbol mapping. In this scheme, adjacent constellation points—those most likely to be confused by noise—differ by only a single bit.

• Error Containment: A symbol error crossing a decision boundary into a neighboring Voronoi region causes only 1 bit error out of k bits, rather than a burst of errors.
• BER Approximation: For high SNR, the Bit Error Rate (BER) ≈ Symbol Error Rate (SER) / log₂M.
• Implementation: The mapping is not unique; any labeling where neighbors have a Hamming distance of 1 is a valid Gray code.

While pure PSK has a 0 dB Peak-to-Average Power Ratio (PAPR), its spectral efficiency is limited. Amplitude Phase Shift Keying (APSK) is a direct evolution that combines PSK's phase modulation with multiple amplitude rings.

• DVB-S2/S2X Standards: These satellite broadcast standards use 16-APSK and 32-APSK with optimized ring ratios to operate close to the non-linear saturation point of the transponder.
• Geometric Shaping: Modern systems further optimize the radius and phase of each ring using geometric shaping to maximize mutual information for the specific non-linear channel model.
• Classification Nuance: An automatic classifier must distinguish between a high-order PSK (like 16-PSK) and a 4+12 APSK constellation, which requires analyzing the amplitude distribution, not just the phase.

A Carrier Frequency Offset (CFO) between the transmitter and receiver local oscillators causes the PSK constellation to rotate continuously at a constant angular velocity. This is a critical impairment that must be corrected before symbol decision.

• Visual Signature: On a scatter plot, the points smear into continuous rings, completely obscuring the discrete phase states.
• Compensation: Requires CFO estimation algorithms (e.g., M-th power loop for M-PSK) to de-rotate the signal in real-time.
• Classification Robustness: A robust modulation classifier must either be trained on CFO-impaired data or use features invariant to rotation, such as the amplitude distribution or spectral correlation density (SCD).