One photon in a billion

An incident photon at frequency ν₀ can leave a molecule in a higher vibrational state, emerging red-shifted by exactly the vibrational frequency of the bond that absorbed the difference. The shift, not the absolute wavelength, is the observable — which is why a Raman spectrum is plotted against wavenumber in cm⁻¹ and looks the same whichever laser you excite it with.

The problem is the cross-section. Spontaneous Raman scattering is about twelve orders of magnitude weaker than fluorescence, so a dilute biological sample produces a signal that sits underneath its own background. Two things rescue it. First, a plasmonic surface: conduction electrons in a gold or silver nanostructure oscillate collectively with the incident field and concentrate it into gaps a few nanometres across. Because the enhancement scales roughly as the fourth power of the local field, a molecule sitting in one of those hot spots can scatter many orders of magnitude more strongly than the same molecule in free solution. Second, analysis: with the right preprocessing and classification, information spread across a whole spectrum can separate classes that no single band resolves.

Both rescues have a cost. Enhancement that varies wildly across a substrate turns quantitative measurement into guesswork, and a classifier trained on inconsistently preprocessed spectra will happily learn the preprocessing. Much of the work is in controlling those two failure modes.