From: Greg Aldering (aldering@panisse.lbl.gov)
Date: Fri May 23 2003 - 09:33:17 PDT
On Fri, 23 May 2003, Robert A. Knop Jr. wrote:
> On Thu, May 22, 2003 at 06:20:20PM -0700, Greg Aldering wrote:
> > I also added a too-faint cut, requiring that SN not be fainter than N
> > times its magnitude uncertainty (including an intrinsic error of
> > 0.17). In Rob's last fits, all the residuals were within 3-sigma for
> > the low-extinction subset, even though no cut on deviation was
> > applied.
>
> In fact, there was one that was too faint -- 9733, which got thrown out
> from the main set. It's a 6-sigma outlier to the faint side from the
> primary (non-extinction-corrected) fit.
>
> > Thus, my cut considers the likelihood that if there were a
> > significantly dim outlier we might have chosed to reject it even
> > without a well-measured color to prove that it is reddened.
>
> ...so, in this case, we may well have, although we don't have the data
> to show that 9733 is faint just because it's reddened.
Yes, 9733 is exactly the case I had in mind. Also, Saul and I noted that
in P99 there was a fit done in which two outliers just slightly worse
than 2-sigma - one too bright and the other too faint - were excluded.
That is what motivated me to run a simulation with a 2-sigma cut on
the (corrected) brightness.
> > I find that our method does have a bias, but that the bias is small
> > compared to our other systematics. However, it might not be small
> > relative to the Riess prior (that has to be checked).
>
> The bias with the Riess prior is complicated, of course, since colors
> which are divergent simply because of measurement errors get modified.
> This bias tends to operate in the *other* direction from the lowe bias,
> assuming that the high-redshift supernoave have worse E(B-V) error bars
> than do the low-redshift supernoave (i.e. high-redshift supernovae get
> made brighter on the average, since all negative E(B-V) values get
> suppressed to zero but positive E(B-V) values are still allowed to be
> corrected for, albeit at a reduced level).
>
> > Here is what I get:
> >
> > N-sigma <A_B> <A_B> Bias
> > dim cut low-z high-z low-high
> > -------------------------------------
> > 2.0 0.068 0.081 0.013
> > 2.5 0.075 0.094 0.019
> > 3.0 0.080 0.104 0.024
>
> Was this using the distribution E(B-V) error bars from the full
> high-redshift set?
Yes, I used everything in your "Gerson table" of May 18th (the last
one you circulated). Note that I only use the error bars on the color
and the magnitude as being representative for purposes of the Monte
Carlo simulation. Do you see a problem with that approach?
- Greg
This archive was generated by hypermail 2.1.4 : Fri May 23 2003 - 09:33:39 PDT