Handle NaNs in y in same manner as 'plot' command

[alexperrone]

There has been several discussions of handling NaNs on the Matlab toolbox page. While there are several options for handling NaNs in all cases, e.g. in x, or y, or in the patch, etc. The simplest and most typical case is to handle NaNs in y in the same manner as 'plot' command so that non-contiguous lines are supported.

Consider this plot:

x = 1:10;
y = x + rand(1, 10);
boundary = 2 * ones(1, 10);
boundedline(x, y, boundary);  % requires boundedline function to be on path

Now when we add a NaN in y, this is not supported:

y(5) = NaN;
boundedline(x, y, boundary);  % error

This is the desired behavior:

This is akin to the way the 'plot' command would handle this case:

plot(x, y);  % with y(5) = NaN

[kakearney]

okay, I've received enough feedback on this topic that I'll try to implement the behavior above. I think I'll allow for a couple different behaviors: 1) remove NaNs/Infs
2) Keep them in the line, resulting in gap, but inpaint past them in the patches 3) Create a gap in both the line and the patch

If anyone has additional behaviors they'd prefer, please post here.

[okomarov]

I reckon 3 should be the default.

[kakearney]

In the end, #1 ended up being more complicated than anticipated to add to the code without major changes, so I've implemented just 2 and 3 in the most recent update. See the new 'nan' input parameter to specify this. I've kept fill-and-smooth as the default behavior for back-compatibility purposes (since that was the behavior in more recent versions).

So far testing has gone smoothly, and I'm in the process of writing up a more in-depth discussion of this function and its new options for my blog; will post here when that's done.

[alexperrone]

The option 1) remove NaNs/Infs was posted on the toolbox page by Eelke Spaak, this didn't work?

To fix this, I added the following around line 313, right before the actual calls to line() and patch(): % remove nans for k = 1:numel(xp) naninds = isnan(xp{k}) | isnan(yp{k}); xp{k} = xp{k}(~naninds); yp{k} = yp{k}(~naninds); end

Since it connects the points which were NaN, it basically linearly interpolates.

[kakearney]

Removing from just the patches would be pretty straightforward, but somewhat redundant because the results are almost identical to the 'fill' option, which simply uses a slight smoother interpolant to fill the point rather than the linear one the graphics fill in. I was worried that users would interpret "remove" to also apply to the line, and that would make coding a bit more complicated. But I've gone ahead and added the remove-from-patches-only option (to be uploaded shortly).

Known limitations from my test cases so far:

• gap: Doesn't handle one-sided NaNs. If there's a NaN in either side of the bounds, it's a gap right now.
• fill: Doesn't handle one-sided NaNs at the beginning or end of the line well. Creates a sloped edge in the patch.
• remove: Same one-sided NaNs on end issue as 'fill'. Also, two-sided NaNs on the ends results in a gap, which may or may not be the user's desired outcome.

On Wed, Feb 17, 2016 at 9:51 AM, alexperrone notifications@github.com wrote:

The option 1) remove NaNs/Infs was posted on the toolbox page by Eelke Spaak, this didn't work?

To fix this, I added the following around line 313, right before the actual calls to line() and patch(): % remove nans for k = 1:numel(xp) naninds = isnan(xp{k}) | isnan(yp{k}); xp{k} = xp{k}(~naninds); yp{k} = yp{k}(~naninds); end

Since it connects the points which were NaN, it basically linearly interpolates.

— Reply to this email directly or view it on GitHub https://github.com/kakearney/boundedline-pkg/issues/1#issuecomment-185323675 .