After a long battle with the SVG documentation here is a triangle colored by linearly interpolated colors specified at each vertex.
<svg xmlns="http://www.w3.org/2000/svg" version="1.200000" width="100%" height="100%" viewBox="0 0 100.000000 86.600000" xmlns:xlink="http://www.w3.org/1999/xlink">
<g transform="matrix(1 0 0 -1 0 86.600000)">
<defs>
<linearGradient id="fadeA-1" gradientUnits="userSpaceOnUse" x1="50.000000" y1="0.000000" x2="50.000000" y2="86.600000">
<stop offset="0%" stop-color="#FF0000"/>
<stop offset="100%" stop-color="#000000" />
</linearGradient>
<linearGradient id="fadeB-1" gradientUnits="userSpaceOnUse" x1="0.000000" y1="86.60000" x2="75.000000" y2="43.300000">
<stop offset="0%" stop-color="#00FF00"/>
<stop offset="100%" stop-color="#000000" />
</linearGradient>
<linearGradient id="fadeC-1" gradientUnits="userSpaceOnUse" x1="100.000000" y1="86.60000" x2="25.000000" y2="43.300000">
<stop offset="0%" stop-color="#0000FF"/>
<stop offset="100%" stop-color="#000000" />
</linearGradient>
<path id="pathA-1" d="M 50.000000,0.000000 L 0.000000,86.600000 100.000000,86.600000 Z" fill="url(#fadeA-1)"/>
<path id="pathB-1" d="M 50.000000,0.000000 L 0.000000,86.600000 100.000000,86.600000 Z" fill="url(#fadeB-1)"/>
<filter id="Default">
<feImage xlink:href="#pathA-1" result="layerA" x="0" y="0" />
<feImage xlink:href="#pathB-1" result="layerB" x="0" y="0" />
<feComposite in="layerA" in2="layerB" operator="arithmetic" k1="0" k2="1.0" k3="1.0" k4="0" result="temp"/>
<feComposite in="temp" in2="SourceGraphic" operator="arithmetic" k1="0" k2="1.0" k3="1.0" k4="0"/>
</filter>
</defs>
<g stroke="none" stroke-width="0" shape-rendering="crispEdges" >
<path d="M 50.000000,0.000000 L 0.000000,86.600000 100.000000,86.600000 Z" fill="url(#fadeC-1)" filter="url(#Default)" />
</g>
</g>
</svg>
What you should see above if your browser supports SVG
Update: This still isn't quite linear interpolation. For some reason I get a tendency toward black in the middle. I tried using fancier arithmetic (multiplication) with masks but this made things worse (quantization artifacts). There is also a nasty relative coordinates issue with the filters I used above.
Update: For reference, here's what I'd like to see:
