This gap may diminish if you're looking at a pool of people who passed some previous threshold, eg they have a degree in math. But even then, some of the gap remains, due to the shape of normal distributions. Think of the heights of men and women, and how even if you looked only at people six feet and taller, the proportion of women in that pool who just barely meet the threshold of 6' would be much higher than the similar group of men; despite the cutoff, the average height of the women would still be less than the average height of the men, within that group. Draw overlapping bell curves to see this visually, and analyze the areas under the graph of each distribution, to the right of a given cutoff point.
We do have a law against invidious racial discrimination. The current implementation of the law states that in practice, if some selection mechanism has a passing rate for one minority that is less than 80% of that whites, that is evidence of discrimination in the absence of validation of that mechanism. This is known as the four fifths rule. A test in which 10% of whites and 7% of blacks passed would fall afoul of this law. Say you set a 115 IQ threshold. That is +1 S.D. from the white mean roughly, and +2 S.D. from the black mean, meaning 16% of whites and 2.3% of blacks would pass, if drawing from the general population. That wouldn't even pass a 'One fifth rule,' let alone a four fifths one.
And validation is no sure thing either - police and fire departments have had these tests validated and still have the results thrown out regularly, due to disparate racial impact, i.e. relatively few blacks and Hispanics (average IQ: ~90) passing the test.
On a related note, explicit racial and gender quotas are more efficient than the race- sex- and IQ-blind, approach in use today. A quota would mandate in effect different thresholds for different groups, but it would allow one at least to use this selection mechanism. Within given groups, one would obtain more efficient outcomes. The problem with quotas is that they make racial preferences explicit, while abandoning such selection methods entirely is only an implicit, de facto racial preference, a rather subtle and hidden one at that.