A Gaussian below the bar is read as its negated exponent above it

[Rafael-SOWNet]

e^(-x^2)/x^2 was integrated by the Gaussian's moments, and 1/(e^(x^2) x^2), the same function, was not: the rule for the moments took the exponential from above the bar only. By parts writes the factor beside erf(b x) in erf(b x)/(e^(b^2 x^2) x^2) the second way, and that integral was declined for want of it. A Gaussian below the bar is read as the exponential of its negated exponent now.

Measured

Input	Was (2.5.0)	Now
"1/(e^(x^2)*x^2)".Integrate("x")	integral(1 / (e ^ x ^ 2 * x ^ 2), x)	-1 / x * e ^ (-x ^ 2) + (-2) * pi ^ (1/2) / 2 * erf(x) + C
"erf(b*x)/(e^(b^2*x^2)*x^2)".Integrate("x")	UnrecognizedFunctionParseException: there is no function erf	an antiderivative in erf and Ei(-2 b^2 x^2)

• Rubi, master at 2db42780 and this change on it, run side by side, 0 wrong everywhere:
  • Family 8: 367 → 371 of 420. The four are erf and erfc of b x over e^(b^2 x^2) x^2 and over e^(b^2 x^2) x^4.
  • Family 2: 544 of 650 on both.
  • The independent suites: 1756 of 1814 on both.
  • Families 1 and 3–7 at five a file: 799 of 897 on both.
• Run alone, the four are 0 of 4 on master and 4 of 4 here. Each is answered in 0.2 to 0.9 s, where master declined it in 2.4 to 10.5 s.
• The unit tests pass: 14,249, none failed (net10.0). On master with An exponential or a hyperbolic function over several linears is split into partial fractions #1640 the classes this touches pass (125), and the native AOT publish of the C++ wrapper builds with no trim or AOT warnings.
• The performance gate passes on c6024927, which is this change on 2db42780. Allocation matches the baseline on all 19 gated benchmarks.

GaussianIntegralTest has two rows with the Gaussian below the bar, each differentiated back.

Part of #1501.

🤖 Generated with Claude Code