The SPSS documentation doesn't mention any limits on the domain of the input.  http://publib.boulder.ibm.com/infocenter/spssstat/v20r0m0/index.jsp?topic=%2Fcom.ibm.spss.statistics.help%2Fidh_mean.htm
(it says the weights must be positive, but does not mention the values).

In the special case, where a value is zero. It can be shown algebraically that the harmonic mean is also zero:

h = a.b / (a+b)
  = a.0  / a

So I retract my previous statement.  I think now that if the sum-of-reciprocals method is used, there should be a special case when one or more values are zero.  But I still don't see why negative values should be treated specially.

I cannot find in the academic literature any indication that harmonic means are defined only for positive values.
This would also fit with the parallel resistor analogy (yes negative resistors do exist - kindof - see Horowitz and Hill).

+      /* Sum of reciprocals only needed for variables with only positive values */
+      m->sumreci = (value <= 0. || m->sumreci == SYSMIS) ? SYSMIS : m->sumreci + 1 / value * weight;

I don't understand why there is this exception for values less than or equal to zero.

The harmonic mean is well defined for negative values, so we don't need to make a special case there.  Where a value is exactly equal to zero, the reciprocal of course is infinite.  However since we are dealing with floating point numbers, this will be recognised by the floating point library and the end result will be (correctly) NaN.  So as far as I can see, we don't need to special case zero either.

Harmonic means for descriptive statistics (descriptives and frequencies). Calculated in moments.c. I separated counting of harmonic means in a new function void moments_calc_additional.

Example:

GET FILE="Speed_and_distance.sav".

WEIGHT BY Distance_km.

DESCRIPTIVES
    /VARIABLES= Speed_kmph
    /STATISTICS=HARMONIC.


Files modified:
descriptives.c
frequencies.q
moments.c
moments.h
psppire-dialog-action-descriptives.c
psppire-dialog-action-frequencies.c