Reference: The Ultimate Resolved Source Cheatsheet

OK, this isn’t actually the ultimate resolved source cheatsheet … yet. I’m just trying to centralize my notes on dealing with extended sources. I’ve been doing this lately, and I had to rederive a bunch of things that I know I’ve figured out before. Hopefully, next time I’ll remember that I wrote all this stuff down here.

Units🔗

Radio astronomical images typically come in two unit systems: Jy/pixel and Jy/beam. For various questions one usually wants to convert to Jy/arcsec² or Jy/sr. _Jy/pixel is not trivially equivalent to Jy/arcsec², because the pixel size can change as a function of position in many map projections.

There must be something profound about why maps come in Jy/bm and models come in Jy/pixel, but I don’t see it right now.

Source fluxes are best measured as total fluxes which come in units of plain Jy. This should be obvious! Flux is conserved! In the general case peak fluxes are always dependent on the image properties.

The important exception, however, is unresolved sources which by their nature have a peak flux equal to their total flux no matter what unit system you’re using. Total flux of 1 Jy, peak flux of 1 Jy/px spread over 1 pixel, or peak flux of 1 Jy/bm spread out over 1 beam.

Resolved sources will have a peak flux that depends on the particular image units and projection, because the total flux will be spread out over a number of pixels that depends on the image particulars. In a Jy/px image, you need to figure out how many pixels the source subtends, and mutatis mutandis for Jy/bm. Generically, pkflux = totflux * arcsec2perpix / arcsec2insource. If you have a 2D Gaussian source where the axes are FWHMs in arcsec, arcsec2insource = 2 pi major minor / (8 ln 2).

I think the best way to work with Jy/bm is to go through Jy/px: the beam volume in pixels can be derived from the beam parameters and the pixel volume: pixperbm = arcsec2perbm / arcsec2perpix. For various operation I feel like this quantity will cancel out, though. Once again, due to the position-dependent pixel volume in most maps, this parameter is not constant across the map.

2D Gaussian Parametrizations🔗

There are a few ways to parametrize a 2D Gaussian. I’ll omit centering and overall scaling in the following equations for simplicity.

Tersest. The simplest expression in terms of symbols is:

z = \exp\left(Ax^2 + Bxy + Cy^2\right)

This parametrization is computationally efficient but not intuitive. One particular danger to note is that the parameters behave a bit oddly, and there are domain limitations that make it awkward for numerical optimizers (which are likely to tweak the parameters outside of their domains): A and C must be negative, and, if I’m doing my algebra right, you must have B² < 4AC.

Mathematical Major/Minor/PA. We often want to express the shape of the 2D Gaussian as a major axis σ1, minor axis σ2, and position angle θ. Here σ1 ≥ σ2 > 0 and θ is the rotation between the major axis and +x (with +θ being towards +y). Because of the symmetry of the shape, there’s a 180° degeneracy in θ. We find

A = -\frac{1}{2}\left[\frac{\cos^2 \theta}{\sigma_1^2} + \frac{\sin^2 \theta}{\sigma_2^2}\right]

,

B = \sin \theta \cos\theta \left(\sigma_2^{-2}- \sigma_1^{-2}\right)

, and

C = -\frac{1}{2}\left[\frac{\sin^2\theta}{\sigma_1^2} + \frac{\cos^2 \theta}{\sigma_2^2}\right]

. Note that if θ = 0, then

z = \exp\left(-\frac{1}{2}\left[\frac{x^2}{\sigma_1^2} + \frac{y^2}{\sigma_2^2}\right]\right)

. Inverting, we find

\theta = \frac{1}{2}\arctan \frac{B}{A-C}

,

\sigma_1^{-2} = -\sqrt{(A-C)^2 + B^2} - A - C

, and

\sigma_2^{-2} = +\sqrt{(A-C)^2 + B^2} - A - C

. During a roundtrip, θ may flip by 180° depending on the numerics, but as mentioned above this doesn’t change anything.

Astronomical Major/Minor/PA. This is the same as the above, except that by the definition of PA, the x axis is declination and the y axis is right ascension. Be careful to write

y = \exp \left (A\delta^2 + B\alpha \delta + C\alpha^2\right)

.

Numerically-friendly. If you’re fitting for Gaussian parameters, it’s very helpful to numerical stability to have nice continuous behavior over all of your parameters. The B parameters above seem to be problematic, so here’s a parametrization that is pretty cheap to compute and behaves nicely:

A=-\frac{1}{2}\left(p\cos^2\theta + q\sin^2\theta\right)

,

B=\cos\theta \sin\theta(q-p)

, and

C=-\frac{1}{2}\left(p\sin^2\theta + q\cos^2\theta\right)

. It shouldn’t be hard to see that

\sigma_1 = p^{-1/2}

and

\sigma_2 = q^{-1/2}

. Here, the optimizer may end up with either of the σ values being the larger. If σ2 > σ1, you can exchange them and add (or subtract) π/2 to θ.

Probabalistic. Change gears for a second. What if we interpret our 2D Gaussian as expressing the joint behavior of variables x and y, with standard deviations σx and σy and covariance Cxy? In this case, we still ignore mean values, but there is one correct normalization:

p(x,y|\sigma_x,\sigma_y,C_{xy}) = \frac{1}{2\pi\sqrt{\sigma_x^2\sigma_y^2-C_{xy}^2}} \exp(Ax^2 + Bxy + Cy^2),

where

A=-\frac{\sigma_y^2}{2(\sigma_x^2\sigma_y^2-C_{xy}^2)},

B = \frac{C_{xy}}{\sigma_x^2\sigma_y^2-C_{xy}^2},

C=-\frac{\sigma_x^2}{2(\sigma_x^2\sigma_y^2-C_{xy}^2)}.

These unpleasant expressions are difficult to invert. If you’re trying to solve for a covariance matrix from some data, use numpy.cov or read Estimation of covariance matrices. It is surprising to me how difficult it is to relate σx, σy, and Cxy to other quantities — everything seems to be done best by going through σ1, σ2, and θ (see ellpar below).

Alternate numerically-friendly probabalistic. The numerically-friendly version above still requires p, q > 0 and is periodic with θ. If the above probabalistic approach is workable, use

\sigma_x = e^u

\sigma_y = e^v

C_{xy} = e^{u+v}\tanh w

. For initialization,

w = \mathop{\mathrm{arctanh}} \frac{C_{xy}}{\sigma_x \sigma_y}.

Wikipedia. The Wikipedia page on Gaussian Functions uses:

z = \exp\left(-\left[ax^2 + 2bxy + cy^2\right]\right)

Here the condition on the parameters is that the matrix [a b \ b c] be positive definite, or that

az_1^2 + bz_1z_2 + cz_2^2 > 0, \forall z_1, z_2

(or, equivalently I think, that a and c be positive and _b_² < ac).

Wikipedia gives a conversion formula to this system from major/minor/PA, but there are two caveats: the Wikipedia system names σ{1,2} as σ{x,y} somewhat misleadingly, and inverts θ to be a clockwise rotation from +x to -y. Avoid those equations since the ones I give above are clearer.

Gaussian Convolutions🔗

Resolved sources are often modeled as 2D Gaussians. These are convolved with the 2D Gaussian of your synthesized beam, so one often wants to compute this convolution or deconvolution analytically. The file src/subs/gaupar.for in MIRIAD has routines for going both directions, and they are not trivial. The deconvolution case needs special handling for sources that are close to (or smaller than!) point sources. The existing heuristics seem only so-so. See pwpy/scilib/astutil.py for a Python deconvolution implementation.

To manage flux through such a deconvolution, work in total flux units. Nothing changes! The peak flux is scaled in two ways: once, to convert from Jy/bm to Jy/px (assuming this is what you’re doing), which involves the beam volume and the pixel volume; and once to convert the source area, which involves the original volume and the deconvolved volume.

Gaussian Coordinate Transformations🔗

When you have a Gaussian model for a source, this usally needs to be rendered into actual pixels or something.

Doing this purely geometrically is difficult. For instance, in a given map projection, the RA/Dec axes may not be precisely aligned with the X/Y pixel axes. If you want to draw a Gaussian source as an ellipse, you need to rotate it a little bit. If the source is very large, these effects change over the source extent and things get scary. It’s better to calculate the RA/Dec of each pixel and brute-force your way through many of these issues.

To do this brute forcing, you need to find the  value of the parameter that goes in the exponential. You can think of this as computing a distance in from the origin in the transformed space where the Gaussian is a circle. You un-translate, then un-rotate, then un-scale:

def gaussdist(x0, y0, maj, min, pa, x, y):
    # x0, y0: Gaussian center
    # maj, min: major/minor axes
    # pa: position angle, from +x toward +y

    dx, dy = x - x0, y - y0
    c, s = np.cos(pa), np.sin(pa)
    a = c * dx + s * dy
    b = -s * dx + c * dy
    d2 = (a / maj)**2 + (b / min)**2

    # d2: squared "distance" from Gaussian such
    # that z = exp(-0.5 * d2)
    return d2

The astronomical “East from North” PA convention means that you can replace X with declination and Y with right ascension. This formulation can be turned into the one in the Wikipedia article on Gaussian Function but is a lot simpler to think about.

A: Converting Jy/pixel to Jy/arcsec²🔗

To do this you need to calculate the size of a pixel at the position of the source you’re considering. Code to do this is in [pwpy/intfbin/msimgen](https://github.com/pkgw/pwpy/blob/master/i ntfbin/msimgen). It looks something like this:

def pixelvolume(pixel2world, pixelcoords):
    delta = 1e-5
    w1 = pixel2world(pixelcoords)
    pixelcoords[X] += delta
    pixelcoords[Y] += delta
    w2 = pixel2world(pixelcoords)
    dra = w2[RA] - w1[RA]
    ddec = w2[DEC] - w2[DEC]
    return (dra**2 + ddec**2) / (2 * delta**2)

B: Error Ellipses From Covariant Variables🔗

This is almost entirely unrelated to everything else here, except that it involves 2D Gaussians. Given two correlated variables, what are the parameters of the ellipse that describes a confidence region?

def ellpar(sx, sy, cxy):
    # sx: std dev (not variance) of x var
    # sy: std dev (not variance) of y var
    # cxy: covariance (not corr. coeff.) of x and y

    from numpy import arctan2, sqrt

    pa = 0.5 * arctan2(2 * cxy, sx**2 - sy**2)
    h = sqrt((sx**2 - sy**2)**2 + 4*cxy**2)  
    maj = sqrt(2 * (sx**2 * sy**2 - cxy**2) / (sx**2 + sy**2 - h))
    min = sqrt(2 * (sx**2 * sy**2 - cxy**2) / (sx**2 + sy**2 + h))

    # maj/min: major/minor axes of ellipse
    # note: sigmas, not FWHMs
    # pa: position angle of ellipse, rotating from +x to +y
    return maj, min, pa

Say we want to draw an actual ellipse representing this bivariate distribution, so that the area within the ellipse represents a certain confidence interval on predicted values. Parametrize this with a value f(CL), where we draw an ellipse that has axes of sizes fmaj and fmin. By transforming X and Y to have a mean of 0 and stddev of 1, then considering the probability distribution of r = x²+y², we can find the relationship between the axis lengths and the confidence interval they subsume. P(r) = r exp (-r²/2) so

f(\textrm{CL}) = \sqrt{-2 \ln (1-\textrm{CL})}.

We see that if we just draw the ellipse according to values associated with the underlying distribution, we get the 39% confidence interval. If we want one in terms of the usual σ limits, we must scale by

f(\sigma) = \sqrt{-2 \ln \mathop{\mathrm{erfc}} (\f rac{\sigma}{\sqrt{2}})}.

For 1 sigma this about 1.5 and for 2 sigma it’s about 2.5. Note: Earlier I had this analysis all wrong!

C: Tracing Ellipses🔗

To trace an ellipse, all we need to do is express it in parametric form. Let th be an offset angle from the major axis of the ellipse. In code form:

def ellpoint(th, x0, y0, maj, min, pa):
    from numpy import cos, sin

    x = x0 + maj * cos(th) * cos(pa) - min * sin(th) * sin(pa)
    y = y0 + maj * cos(th) * sin(pa) + min * sin(th) * cos(pa)
    return x, y

That’s all there is to it.

Questions or comments? For better or worse this website isn’t interactive, so send me an email or, uh, Toot me.

To get notified of new posts, try subscribing to my lightweight newsletter or my RSS/Atom feed. No thirsty influencering — you get alerts about what I’m writing; I get warm fuzzies from knowing that someone’s reading!

Later: Reference: Deleting wireless network connections in Fedora 16

Earlier: Reference: SMA MIRIAD merge

See a list of all posts.

View the revision history of this page.