Fourier series and transforms
An MRI scanner never takes a picture of you. It records how strongly your body’s hydrogen nuclei respond to magnetic patterns of different fineness, a list of numbers physicists call k-space. The picture you see is computed afterwards, by a single mathematical operation: the Fourier transform.
The idea behind it is that any signal, a sound, an image or a wavefunction, can be built by adding up pure waves of different frequencies. Knowing how much of each frequency is present is as good as knowing the signal itself. Quantum mechanics leans on this so heavily that position and momentum turn out to be a Fourier pair.
Building a square wave from circles
A square wave has corners, and sine waves have none, yet the square wave can be written as a sum of sines:
Each sine is a circle turning at its own speed, and the plate stacks them tip to tail. Use 1 terms and watch the pen on the last circle trace something closer and closer to a square.
How are the coefficients found? Different sines are orthogonal: the average of over a full period is zero unless . So to find how much of is in a signal, multiply by and average. Everything else drops out.
The overshoot that never goes away
Look closely at the corners. However many terms you add, the sum overshoots the jump by about 9% before settling down. More terms squeeze the overshoot into a narrower spike, but its height does not shrink. This is the Gibbs phenomenon.
The limit is exactly of the jump. You can see it ringing around sharp edges in compressed images and MRI scans. Established
Narrow in time, wide in frequency
For a signal that is not periodic, the sum becomes an integral, the Fourier transform:
Take a pulse shaped like a bell curve and set its width to 1.00. The top of the plate shows the pulse and the bottom shows its spectrum. Squeeze the pulse and the spectrum spreads; widen it and the spectrum narrows.
Measured as standard deviations, the product of the two widths can never be smaller than one half, and the bell curve is the shape that reaches the limit exactly:
A short drum hit contains many frequencies; a pure note has to last. In Part One, with , this same line of mathematics becomes Heisenberg’s uncertainty principle.
Two useful theorems
The convolution theorem says that blurring a signal, which means averaging each point with its neighbours, becomes simple multiplication of the transforms. Camera blur, the smearing of a detector and the spreading of a wave packet are all convolutions, and the transform turns each into multiplication.
Parseval’s theorem says the total energy is the same whether you add it up in time or in frequency: nothing is lost by changing view.
With the convention above, . Proof: write as a double integral over and , do the integral first, and use .
Back to the scanner
The plate shows a simple test image, the scanner’s k-space data for it, and the picture rebuilt from only part of that data. Keep the central 6 columns and rows of k-space and throw away the rest.
The middle of k-space holds the coarse shapes; the outer parts hold fine detail and sharp edges. Fast MRI scans sample the middle densely and the edges sparsely, then let the reconstruction fill in. On a computer, the fast Fourier transform does this in steps instead of , which is what made digital imaging practical.
What you now have
Signals as sums of waves, coefficients from orthogonality, a 9% overshoot you cannot remove, and a trade-off between width in time and width in frequency that will become the uncertainty principle.
· · ·
Three questions to carry with you.
And one to work out. What is the coefficient of , the first term, in the square wave above?
Sources: E. M. Stein and R. Shakarchi, Fourier Analysis. R. N. Bracewell, The Fourier Transform and Its Applications. D. W. McRobbie et al., MRI from Picture to Proton. Checked by code: Gibbs overshoot 8.949% (limit), Gaussian ΔtΔω = 0.5, b₁ = 4/π.