Blindness for w

From: Alexander Conley (AJConley@lbl.gov)
Date: Mon Oct 04 2004 - 13:10:43 PDT

  • Next message: Alexander Conley: "Removing the blindness"

    After some playing around it looks like my blindness scheme
    doesn't work nearly as well in w space as it did in Om/Ol space.
    The culprit seems to be the shape of the error contours, which are
    almost pathological in the w/om plane -- at least for the SN data alone.

    I do have a version that kind of works, which I will use, but
    improvements will clearly have to be made if something like
    SNAP tries to use this technique.

    The problem comes down to how you specify the original,
    non-blind values of Om/Ol (or w/Om). In the first case, after
    some experimentation I discovered that the most reliable
    method seemed to be to estimate Om/Ol from the 1D marginalized
    probability distributions. Originally I simply used the best fitting
    4 dimensional (om,ol,script_m,alpha) point, but this proved to
    be a bit sensitive. The best fit could lie close to a grid point
    boundary,
    so sometimes a very small shift in om/ol could lead to a larger than
    anticipated change. The slightly random nature of the fitting process
    seemed to also contribute to some variablility. The marginalized
    distributions were more stable, so I went with them.

    Sadly, the marginalized distributions in w/om space (at least
    for SN data only) are pretty useless. The curvature means that the
    marginalized point and the best fitting 4D point are frequently rather
    far away from each other (which is not true in the Om/Ol case).
    The marginalized point seems to frequently end up in the tails of the
    2D error contour, which isn't really surprising when one looks at it.
    Another problem is that the error contour seems to run badly off to
    negative om values, which I can't include in my parameter space. This
    is also an issue in Om/Ol, but much less of one.

    Test show that using the marginalized w/Om simply doesn't work.
    Applying a shift to w/om of +0.2, -0.1 shifted the best fit values by
    +0.12, -0.02, which is not sufficiently close to the desired effect by
    a long shot.

    Using the best fit 4D point works much better, but the previous worries
    about the sensitivity to the gridding in 4D apply.

    Note that I do not expect the marginalized distributions to be a problem
    once the CMB, etc. priors are added. This should have the effect of
    making the error contour much more normal looking, in which case
    the marginalizations should be useful again.

    In theory it would be possible to apply the blindness after including
    these
    additional priors, but I think this is a bad idea. I suspect that it
    would not
    be too difficult to back out some information about the secret offset if
    the priors were included at this stage, particularly because the 2dFGRS
    contours are so vertical in this plane. That is, I think somebody
    could violate
    the blindness scheme if I tried this. In order to really make this
    work you
    have to have end-to-end control of the results, which I don't if the CMB
    and galaxy surveys are included.

    In a conversation last week you suggested trying to apply some sort of
    polynomial warping to 'straighten-out' the error contour. This is an
    interesting idea, but since when you get right down to it I don't really
    care about w for this paper, I am not going to explore it. Instead I
    will
    use the 4D point with the provisio that it simply isn't as stable, and
    leave
    the 'unbending' to somebody who really needs this information.

    Alex



    This archive was generated by hypermail 2.1.4 : Mon Oct 04 2004 - 13:10:42 PDT