Double-Blind Holography Recovers FEL Pulses and Attosecond Delays

Self-amplified spontaneous-emission free-electron lasers can produce ultrashort XUV and X-ray pulses, but each pulse carries stochastic spectral and temporal structure. That variability complicates experiments whose signal depends on pulse duration, phase, or pump–probe arrival time: sorting shots after acquisition requires a measurement made on the same shot. Azzolin et al. present an all-optical route intended to supply both pieces of information at once. Their Double-Blind Holography (DBH) measurement interferes an FEL pulse with a high-harmonic-generation (HHG) pulse and reconstructs the FEL temporal field while tagging the relative delay.

Core Contribution

The central distinction is that neither input pulse is assumed to be fully known. Rather than treating HHG as a calibrated reference pulse, DBH uses the redundancy in a two-dimensional spectral-spatial interferogram to recover the two fields jointly. The authors position this as a way to combine single-shot waveform reconstruction and arrival-time tagging, a pairing that conventional pulse diagnostics and beamline timing monitors do not provide in one all-optical measurement.

The experimental layout is quasi-collinear: independent HHG and FEL beams are spectrally dispersed by an XUV grating, and the FEL overlaps spatially, at least in part, with harmonic 21 at the detector. The resulting fringes change slope with the sign of the HHG–FEL delay. Figure 1 makes the operational point clear: the same detector record contains the individual spectra, their interference, and a Fourier-domain cross-correlation whose displaced lobes encode the delay. This converts pulse timing from a separate diagnostic into a quantity extracted from the reconstruction measurement itself.

Technical Approach

The method uses a two-dimensional vectorial phase-retrieval algorithm (2D VPR). It works in time and spatial-frequency coordinates, where compact supports constrain the two unknown pulses. The algorithm masks nonzero regions using rectangular supports, evaluates a leakage-score error over possible support dimensions, and selects the global minimum. For the reported reconstruction, the support search finds [8.3 fs, 1.0 mm⁻¹] for HHG and [22.2 fs, 1.0 mm⁻¹] for FEL; one grid step is 2.78 fs in time and 0.52 mm⁻¹ in spatial frequency.

This support selection is not merely an implementation detail. A compact support must contain the full pulse, yet an overly large or misplaced one permits ambiguous solutions. The supplementary analysis reports a local minimum at [16.7 fs, 3.1 mm⁻¹] that yields an incorrect reconstruction. The search is computationally substantial: with $N_t=N_k=9$, 65,561 VPR iterations are evaluated. The paper reports that 32 parallel nodes, one task per core, complete that search in about 5 minutes, compared with several hours for sequential execution.

Figure 3 illustrates why the two-dimensional data matter. The leakage maps expose a global optimum for the supports of both objects, while the spatial-frequency axis adds information unavailable to a purely spectral measurement.

Results and Analysis

For delay tagging, Gaussian fits to the cross-correlation side lobes yield a normalized accuracy distribution centered at 0.37 fs. The extracted single-shot delay correlates with beamline BAM and LAM monitor measurements, with 6.6 fs RMSE after excluding shots classified as FEL-jitter outliers and 8.6 fs over the full dataset. The comparison supports the claim that DBH resolves relative timing far more finely than the monitor agreement scale, but it should not be read as a direct absolute-time calibration: the monitors themselves set the benchmark used in this plot.

Figure 2 shows both aspects. The narrow retrieved-delay distribution substantiates attosecond-scale fitting precision, whereas the scatter against the diagnostics exposes the practical timing noise of the beamline and the effect of outliers.

The supplementary material also documents a zero-delay case, where temporal fringes vanish and the 2D route cannot be applied because cross-correlation signals overlap. A one-dimensional VPR variant then uses separately collected HHG, FEL, and interference spectra. It reconstructs an FEL pulse with 10.4 fs FWHM against a 7.0 fs transform limit and a group-delay dispersion of about 15 fs². That fallback is informative but weaker: it uses fewer data points, lacks the redundancy of the 2D map, is more noise-sensitive, and requires intensity rescaling outside the interference region.

The evidence therefore supports a useful experimental demonstration, especially for facilities that need shot-resolved timing metadata alongside pulse characterization. Its strongest result is integration rather than a demonstrated universal accuracy across FEL modes, photon energies, and pulse shapes. The paper shows a working reconstruction and timing workflow for the reported sub-10 fs regime; broader deployment will depend on how reliably the support optimization and fringe contrast transfer to other operating conditions.