Ferkans — Interactive Telecom Tutor

Why Phase Retrieval Matters

In coherent imaging, the measured signal has both amplitude and phase: $E_s = |E_s|\,e^{j\phi}$ . Many RF systems, however, can only measure the intensity $|E_s|^2$ (or amplitude $|E_s|$ ) — the phase $\phi$ is lost. Recovering $\mathbf{x}$ from $\{|\langle \mathbf{a}_i, \mathbf{x} \rangle|^2\}$ is the phase retrieval problem, a fundamentally non-linear inverse problem that is harder than the linear problems of Chapter 11--Chapter 14.

The point is that phase is not a minor detail — it carries the dominant structural information in images. Losing the phase and reconstructing from magnitude alone produces catastrophic artifacts, not graceful degradation. This motivates the dedicated algorithms of this chapter.

Definition:
The Phase Retrieval Problem

Given: Intensity measurements $y_i = |\langle \mathbf{a}_i, \mathbf{x}_0 \rangle|^2$ , $i = 1, \ldots, M$ , where $\mathbf{a}_i \in \mathbb{C}^N$ are known measurement vectors and $\mathbf{x}_0 \in \mathbb{C}^N$ is the unknown signal.

Find: $\hat{\mathbf{x}}$ such that $|\langle \mathbf{a}_i, \hat{\mathbf{x}} \rangle|^2 = y_i$ for all $i$ .

Key properties:

The problem is non-linear (quadratic in $\mathbf{x}$ ).
It has a global phase ambiguity: $e^{j\phi}\mathbf{x}_0$ is also a solution for any $\phi \in [0, 2\pi)$ .
For generic (Gaussian) measurements, $M \geq 4N - 4$ intensity measurements are sufficient for unique recovery (up to global phase) in the complex case.
For Fourier measurements (common in imaging), additional measurements or constraints are needed.

,

Phase Retrieval

The problem of recovering a complex-valued signal $\mathbf{x}$ from magnitude-only (phaseless) measurements $y_i = |\langle \mathbf{a}_i, \mathbf{x}\rangle|^2$ . A non-linear inverse problem with inherent global phase ambiguity.

Global Phase Ambiguity

The fundamental ambiguity in phase retrieval: if $\mathbf{x}_0$ is a solution, then $e^{j\phi}\mathbf{x}_0$ produces identical intensity measurements for any $\phi \in [0, 2\pi)$ . Recovery is therefore defined up to a global phase factor.

Related: The Phase Retrieval Problem

Historical Note: The Oppenheim--Lim Experiment (1981)

1980s

In a landmark 1981 paper, Alan Oppenheim and Jae Lim demonstrated that phase carries far more perceptual information than amplitude in images. They swapped the amplitude and phase spectra of two images and showed that the resulting hybrid always resembles the image whose phase was used, not the one whose amplitude was used.

This experiment — now a staple of signal processing courses — established that edge locations, object positions, and spatial structure are encoded in the phase spectrum, while amplitude controls only overall contrast and energy distribution. The result motivated decades of research into phase retrieval algorithms.

Theorem: Uniqueness of Phase Retrieval from Generic Measurements

Let $\mathbf{a}_1, \ldots, \mathbf{a}_M \in \mathbb{C}^N$ be generic (e.g., i.i.d. Gaussian) measurement vectors. If $M \geq 4N - 4$ , then any signal $\mathbf{x}_0 \in \mathbb{C}^N$ is determined uniquely (up to global phase) by the intensity measurements $y_i = |\langle \mathbf{a}_i, \mathbf{x}_0\rangle|^2$ .

That is, if $|\langle \mathbf{a}_i, \mathbf{x}\rangle|^2 = |\langle \mathbf{a}_i, \mathbf{x}_0\rangle|^2$ for all $i$ , then $\mathbf{x} = e^{j\phi}\mathbf{x}_0$ for some $\phi \in [0, 2\pi)$ .

Each intensity measurement $y_i = |\langle \mathbf{a}_i, \mathbf{x}\rangle|^2$ constrains $\mathbf{x}$ to lie on a circle (manifold of constant inner-product magnitude). With sufficiently many such constraints from generic directions, the intersection shrinks to a single orbit $\{e^{j\phi}\mathbf{x}_0\}$ .

Proof

Lifting to rank-1 matrices

Write $y_i = \text{tr}(\mathbf{A}_i \mathbf{X})$ where $\mathbf{X} = \mathbf{x}\mathbf{x}^H$ and $\mathbf{A}_i = \mathbf{a}_i\mathbf{a}_i^H$ . The intensity measurements are linear in $\mathbf{X}$ .

Dimension counting

The space of rank-1 Hermitian PSD matrices $\mathbf{X} = \mathbf{x}\mathbf{x}^H$ has real dimension $2N - 1$ (modulo global phase). Each measurement provides one real constraint. For unique determination, we need $M \geq 2N - 1$ constraints, but the quadratic nature of the map requires roughly $4N$ measurements for injectivity.

Algebraic geometry argument

Balan, Casazza, and Eddins (2006) showed via algebraic geometry that $4N - 4$ generic measurements make the intensity map injective modulo global phase. The key tool is the analysis of the variety $\{(\mathbf{x}, \mathbf{x}') : |\langle \mathbf{a}_i, \mathbf{x}\rangle|^2 = |\langle \mathbf{a}_i, \mathbf{x}'\rangle|^2 \; \forall i\}$ .

Conclusion

For $M \geq 4N - 4$ generic measurements, the only solutions are $\mathbf{x} = e^{j\phi}\mathbf{x}_0$ . $\blacksquare$

Definition:
Common Measurement Models for Phase Retrieval

1. Gaussian measurements (generic): $\mathbf{a}_i \sim \mathcal{CN}(\mathbf{0}, \mathbf{I})$ . Theoretical gold standard; $M \geq 4N$ suffices for recovery.

2. Fourier measurements (coded diffraction patterns): $\mathbf{a}_i = \mathbf{D}_\ell \mathbf{F} \mathbf{e}_k$ , where $\mathbf{F}$ is the DFT and $\mathbf{D}_\ell$ is a diagonal mask. Multiple masks ( $L \geq 3$ ) with $M = LN$ total measurements ensure recovery.

3. Short-time Fourier transform (STFT): $y_{m,k} = |\sum_n x_n\,w_{n-m}\,e^{-j2\pi kn/N}|^2$ . Natural for time-frequency analysis; redundancy from overlapping windows provides the extra measurements.

4. RF imaging (scattered field magnitude): $y_i = |[\mathbf{A}\mathbf{c}]_i|^2$ where $\mathbf{A}$ is the imaging forward model and $\mathbf{c}$ is the reflectivity vector. The structure of $\mathbf{A}$ (Fourier-like) determines recovery feasibility.

,

Example: Impact of Phase Loss on RF Image Reconstruction

Setup: 2D RF imaging with $N_t = 4$ , $N_r = 8$ , $N_f = 32$ . Scene: 5 point scatterers. SNR = 25 dB.

Reconstruction method	NMSE (dB)
With full complex data (FISTA)	$-25.3$
With magnitude only, no phase retrieval	$-4.1$
With magnitude only + Wirtinger flow	$-19.7$

Explain the massive gap between coherent and magnitude-only reconstruction and why phase retrieval partially closes it.

Solution

Why magnitude-only fails catastrophically

Without phase retrieval, we lose all spatial structure encoded in the phase spectrum. The $-4.1$ dB NMSE shows the reconstruction is dominated by artifacts — essentially random relative to the true scene.

Phase retrieval narrows the gap

Wirtinger flow recovers $-19.7$ dB NMSE, within $\sim$ 5.6 dB of the coherent baseline. This "phase retrieval tax" reflects the information lost by squaring: the measurements $y_i = |z_i|^2$ carry less Fisher information about $\mathbf{x}$ than the full complex measurements $z_i$ .

The gap narrows with more measurements

Increasing the measurement-to-unknown ratio $M/N$ reduces the phase retrieval penalty. At $M/N > 8$ with coded measurements, the gap shrinks below 3 dB.

The Importance of Phase in Imaging

Demonstrates the importance of phase information in imaging.

Top-left: Original scene (point or extended targets). Top-right: Reconstruction from full complex data. Bottom-left: Reconstruction using only amplitude spectrum (phase set to zero). Bottom-right: Reconstruction using only phase spectrum (amplitude set to uniform).

The phase-only reconstruction retains spatial structure (edges, positions), while the amplitude-only reconstruction loses all spatial information — confirming the Oppenheim--Lim finding.

Parameters

Scene type

Grid size

N

32

Common Mistake: Phase Retrieval Has Inherent Ambiguities Beyond Global Phase

Mistake:

A common error is assuming that once the global phase is fixed, the solution is unique. For Fourier measurements, additional ambiguities exist:

Conjugate inversion: $\bar{\mathbf{x}}_{-n}$ (complex conjugate of the time-reversed signal) has the same Fourier magnitude.
Spatial shift: $\mathbf{x}_{n-n_0}$ has the same Fourier magnitude (shift theorem).
These form a group of "trivial ambiguities" specific to Fourier phase retrieval.

For generic (non-Fourier) measurements, only the global phase ambiguity remains — one motivation for using coded or random measurement designs.

Correction:

When evaluating phase retrieval algorithms, always align the recovered signal to the true signal before computing error metrics:

$\text{dist}(\hat{\mathbf{x}}, \mathbf{x}_0) = \min_{\phi \in [0, 2\pi)} \|\hat{\mathbf{x}} - e^{j\phi}\mathbf{x}_0\|_2.$

For Fourier measurements, also check conjugate-inversion and shift ambiguities.

Quick Check

For a complex signal $\mathbf{x}_0 \in \mathbb{C}^{128}$ with generic (Gaussian) measurement vectors, approximately how many intensity measurements $M$ are needed for unique recovery up to global phase?

$M \geq 128$

$M \geq 256$

$M \geq 508$

$M \geq 16384$

Correction:

M \geq 508

The bound is $M \geq 4N - 4 = 508$ . In practice, $M \approx 4N$ to $6N$ measurements are used for robust recovery.

Why This Matters: WiFi RSSI and IoT Sensing

Phase retrieval is not just an optics problem — it arises naturally in RF systems. WiFi RSSI (Received Signal Strength Indicator) reports only the power $|h|^2$ of the channel, not the complex channel coefficient $h$ . Many low-cost IoT sensors similarly output only signal strength.

Indoor localization and imaging systems based on RSSI measurements face exactly the phase retrieval problem: they must reconstruct spatial information from magnitude-only data. The algorithms of Sections 16.2--16.4 apply directly, with the WiFi propagation model replacing the generic measurement matrix.

See full treatment in Phase Retrieval in RF Imaging

Key Takeaway

Phase retrieval recovers a complex signal from magnitude-only measurements — a non-linear inverse problem that is fundamentally harder than linear reconstruction. Phase carries the dominant structural information in images: edges, positions, and layout. For generic measurements, $M \geq 4N$ intensity samples suffice for unique recovery; Fourier measurements require additional constraints. Without phase retrieval, magnitude-only RF imaging produces catastrophic artifacts — a roughly 20 dB NMSE penalty.

Why Phase Retrieval Matters

Why Phase Retrieval Matters

Definition: The Phase Retrieval Problem

Phase Retrieval

Global Phase Ambiguity

Historical Note: The Oppenheim--Lim Experiment (1981)

Theorem: Uniqueness of Phase Retrieval from Generic Measurements

Lifting to rank-1 matrices

Dimension counting

Algebraic geometry argument

Conclusion

Definition: Common Measurement Models for Phase Retrieval

Example: Impact of Phase Loss on RF Image Reconstruction

Why magnitude-only fails catastrophically

Phase retrieval narrows the gap

The gap narrows with more measurements

The Importance of Phase in Imaging

Parameters

Common Mistake: Phase Retrieval Has Inherent Ambiguities Beyond Global Phase

Quick Check

Why This Matters: WiFi RSSI and IoT Sensing

Key Takeaway

Definition:
The Phase Retrieval Problem

Definition:
Common Measurement Models for Phase Retrieval