The Point-Spread Function (PSF)

The PSF Determines Everything

The point-spread function (PSF) is the single most important quantity for understanding matched filter image quality. It determines the resolution (mainlobe width), the dynamic range (sidelobe level), and the artifacts (sidelobe pattern). Every improvement to the matched filter -- windowing, filtered backpropagation, adaptive beamforming -- is ultimately a manipulation of the PSF.

Definition:

Point-Spread Function (PSF)

The point-spread function of the matched filter is the (q,qβ€²)(q, q') entry of the Gram matrix:

PSF(q,qβ€²)=[AHA]q,qβ€²=AqHAqβ€².\text{PSF}(q, q') = [\mathbf{A}^{H} \mathbf{A}]_{q,q'} = \mathbf{A}_{q}^{H} \mathbf{A}_{q'}.

As a continuous function of target position:

h(p,pβ€²)=A(p)HA(pβ€²).h(\mathbf{p}, \mathbf{p}') = \mathbf{A}(\mathbf{p})^H \mathbf{A}(\mathbf{p}').

Properties:

  • Hermitian: h(p,pβ€²)=h(pβ€²,p)βˆ—h(\mathbf{p}, \mathbf{p}') = h(\mathbf{p}', \mathbf{p})^*.
  • For translation-invariant systems: h(p,pβ€²)=h(pβˆ’pβ€²)h(\mathbf{p}, \mathbf{p}') = h(\mathbf{p} - \mathbf{p}').
  • The mainlobe width determines the resolution limit.
  • The sidelobe level determines the dynamic range.

Theorem: The Diffraction Limit

For a multi-static stepped-frequency radar with NtN_t transmitters, NrN_r receivers, and NfN_f frequencies spanning bandwidth BB around center frequency fcf_c, the matched filter resolution is:

Range resolution:

Ξ”r=c2B.\Delta_r = \frac{c}{2B}.

Cross-range resolution:

Ξ”cr=Ξ»c2Deff,\Delta_{\text{cr}} = \frac{\lambda_c}{2 D_{\text{eff}}},

where Ξ»c=c/fc\lambda_c = c/f_c is the center wavelength and DeffD_{\text{eff}} is the effective aperture size (for a virtual array: Deff=NtNrdD_{\text{eff}} = N_t N_r d, where dd is the element spacing).

These limits are fundamental: no matched filter variant can resolve targets closer than Ξ”r\Delta_r in range or Ξ”cr\Delta_{\text{cr}} in cross-range, regardless of SNR.

Resolution is determined by the extent of k-space coverage. Bandwidth controls the radial extent (range), and aperture controls the angular extent (cross-range). The Fourier uncertainty principle connects k-space coverage to spatial resolution.

Proof
k-space coverage

Each (Tx, Rx, frequency) triple maps to a wavenumber ΞΊs,r=ΞΊ\ntntxpos+ΞΊ\ntnrxpos\kappa_{\mathbf{s},\mathbf{r}} = \kappa_{\ntn{txpos}} + \kappa_{\ntn{rxpos}}. The radial extent is Ξ”kr=4Ο€B/c\Delta k_r = 4\pi B / c and the angular extent is Ξ”kΞΈβ‰ˆ4Ο€Deff/(Ξ»cR)\Delta k_\theta \approx 4\pi D_{\text{eff}} / (\lambda_c R).

Fourier resolution

The spatial resolution is the reciprocal of the k-space extent:

Ξ”r=2πΔkr=c2B,Ξ”cr=2πΔkΞΈ=Ξ»c2Deff.\Delta_r = \frac{2\pi}{\Delta k_r} = \frac{c}{2B}, \qquad \Delta_{\text{cr}} = \frac{2\pi}{\Delta k_\theta} = \frac{\lambda_c}{2 D_{\text{eff}}}.

Fundamental nature

Since the matched filter performs an inverse Fourier transform of the k-space data, it cannot produce features finer than the reciprocal bandwidth -- this is the imaging analogue of the Heisenberg uncertainty principle.

Definition:

Sidelobe Structure

For uniform k-space sampling over bandwidth BB with NfN_f frequencies, the 1D range PSF is a Dirichlet kernel (periodic sinc):

hr(Ξ”r)=sin⁑(Ο€NfΞ”fβ‹…2Ξ”r/c)sin⁑(πΔfβ‹…2Ξ”r/c).h_r(\Delta r) = \frac{\sin(\pi N_f \Delta f \cdot 2\Delta r / c)} {\sin(\pi \Delta f \cdot 2\Delta r / c)}.

Key sidelobe properties:

  • First sidelobe: βˆ’13.3-13.3 dB below the mainlobe peak.
  • Sidelobes decay as 1/(Ο€n)1/(\pi n) for the nn-th sidelobe.
  • For a 2D separable system: h(Ξ”r,Δθ)=hr(Ξ”r)β‹…hΞΈ(Δθ)h(\Delta r, \Delta\theta) = h_r(\Delta r) \cdot h_\theta(\Delta\theta), producing a characteristic cross pattern of sidelobes.

A strong scatterer at 0 dB creates sidelobes at βˆ’13-13 dB, masking any target weaker than 5% of its amplitude.

Definition:

Apodization (Window Weighting)

Apodization applies a weight function to the matched filter sum to reduce sidelobes:

c^DAS(p)=βˆ‘i,j,kwi,j,k yi,j,k ej2Ο€fkΟ„i,ji,j(p),\hat{c}_{\text{DAS}}(\mathbf{p}) = \sum_{i,j,k} w_{i,j,k}\, y_{i,j,k}\, e^{j2\pi f_k {\tau_{i,j}}_{i,j}(\mathbf{p})},

where wi,j,kw_{i,j,k} is the apodization window.

Window Peak sidelobe Mainlobe broadening
Uniform (rectangular) βˆ’13.3-13.3 dB 1.0Γ—1.0\times
Hamming βˆ’42.7-42.7 dB 1.50Γ—1.50\times
Taylor (nΛ‰=5\bar{n}=5, βˆ’35-35 dB) βˆ’35.0-35.0 dB 1.25Γ—1.25\times
Chebyshev (βˆ’40-40 dB) βˆ’40.0-40.0 dB 1.35Γ—1.35\times

Tradeoff: Lower sidelobes imply wider mainlobe (poorer resolution). This is the fundamental resolution--dynamic range tradeoff of matched filter imaging. For RF imaging, Taylor weighting is standard -- it provides nearly constant sidelobes with minimal mainlobe broadening.

PSF and k-Space Coverage

Visualizes the relationship between k-space coverage and the PSF.

Left panel: k-space samples for the selected array/bandwidth configuration.

Right panel: The resulting 2D PSF (magnitude in dB).

  • Increasing the aperture narrows the cross-range mainlobe.
  • Increasing the bandwidth narrows the range mainlobe.
  • Windowing reduces sidelobes at the cost of mainlobe width.
Parameters
8
10

Example: PSF for a MIMO Virtual Array

A MIMO radar has Nt=4N_t = 4 transmitters (spacing 2Ξ»2\lambda) and Nr=8N_r = 8 receivers (spacing Ξ»/2\lambda/2), with Nf=32N_f = 32 frequencies spanning 9--11 GHz (B=2B = 2 GHz, fc=10f_c = 10 GHz).

(a) Compute the virtual array aperture DeffD_{\text{eff}}.

(b) Compute the range and cross-range resolution.

(c) Determine the peak sidelobe level for uniform weighting.

Solution
Virtual aperture

The virtual array has NtΓ—Nr=32N_t \times N_r = 32 elements. With the MIMO waveform design (Tx spacing =NrΓ—dRx= N_r \times d_{\text{Rx}}), the virtual elements fill a ULA with Ξ»/2\lambda/2 spacing and Deff=32Γ—Ξ»/2=16Ξ»=0.48D_{\text{eff}} = 32 \times \lambda/2 = 16\lambda = 0.48 m at 10 GHz.

Resolution
  • Range: Ξ”r=c/(2B)=3Γ—108/(2Γ—2Γ—109)=7.5\Delta_r = c/(2B) = 3 \times 10^8 / (2 \times 2 \times 10^9) = 7.5 cm.
  • Cross-range at 5 m range: Ξ”cr=Ξ»/(2Γ—16Ξ»)=1/(32)β‰ˆ1.8Β°\Delta_{\text{cr}} = \lambda / (2 \times 16\lambda) = 1/(32) \approx 1.8Β°.
Sidelobe level

For uniform weighting with 32 virtual elements and 32 frequencies: the sidelobe level is approximately βˆ’13.3-13.3 dB in both range and cross-range. The 2D peak sidelobe at the cross pattern is also βˆ’13.3-13.3 dB (separable).

Common Mistake: The PSF Is Not Always Shift-Invariant

Mistake:

Assuming the PSF is the same everywhere in the scene: for near-field geometries (target range comparable to aperture size), the PSF varies spatially. Using a single PSF for deconvolution introduces artifacts.

Correction:

The PSF h(p,pβ€²)h(\mathbf{p}, \mathbf{p}') is shift-invariant only in the far field. For near-field imaging (e.g., indoor RF imaging with large arrays), use spatially-varying PSF analysis or the full Gram matrix AHA\mathbf{A}^{H} \mathbf{A}.

Common Mistake: Resolution Does Not Improve with SNR

Mistake:

Expecting that higher SNR will resolve targets closer than the diffraction limit Ξ”r\Delta_r or Ξ”cr\Delta_{\text{cr}}. More SNR reduces the noise floor but does not narrow the mainlobe.

Correction:

The diffraction limit is set by the k-space extent (bandwidth and aperture), not by the noise level. To improve resolution beyond the diffraction limit, one needs either more k-space coverage (larger array, wider bandwidth) or super-resolution methods (Capon, MUSIC, sparse recovery) that exploit scene structure.

Why This Matters: PSF as the Array Beam Pattern

The cross-range PSF of the matched filter is exactly the array beam pattern (or radiation pattern) of the Tx-Rx virtual array, evaluated at the scene range. The concepts of mainlobe, sidelobes, grating lobes, and null steering from antenna theory (Chapter 7) directly apply to the imaging PSF.

This connection runs deep: every improvement to the array beam pattern (larger aperture, non-uniform element spacing, sparse arrays) directly improves the imaging PSF.

Quick Check

A matched filter image of a single point scatterer at 0 dB shows a sidelobe at βˆ’13-13 dB. A second scatterer at βˆ’20-20 dB is located at the sidelobe position. What do you observe in the MF image?

Both scatterers are clearly visible as separate peaks

The weak scatterer is masked by the sidelobe of the strong one

The sidelobe and weak scatterer cancel each other

Correction:
The weak scatterer is masked by the sidelobe of the strong one

Since βˆ’13-13 dB >βˆ’20> -20 dB, the sidelobe artifact dominates the weak target signal. This is the dynamic range limitation of the matched filter.

Key Takeaway

The PSF =AHA= \mathbf{A}^{H} \mathbf{A} determines everything about matched filter image quality. Range resolution is c/(2B)c/(2B), cross-range resolution is Ξ»/(2Deff)\lambda/(2D_{\text{eff}}), and the first sidelobe at βˆ’13-13 dB limits the dynamic range. Windowing trades resolution for sidelobe suppression -- a fundamental tradeoff that only super-resolution or sparse recovery methods can circumvent.

The Matched Filter / Backpropagation EstimatorFiltered Backpropagation