关于提升Pari/GP中lindep函数积分近似适用性及解决异常输出问题的技术问询

Let's break down why you're getting that unexpected complex vector, how to fix it, and how to make lindep work better for integral approximation scenarios.

Why Isn't lindep Returning the Expected [12,12,-3]? #

The core issue here is that lindep relies on floating-point approximations to detect linear dependencies, and two key factors are throwing it off:

1. Insufficient precision mismatch: You used a 20-digit approximation of your integral, but Pari/GP's default real precision (usually around 28 digits) might not be enough to distinguish the tiny error in your approximation from a "false" linear dependency with large integers.
2. Default threshold behavior: By default, lindep looks for any linear combination that's numerically close to zero—not necessarily the clean, small-integer combination you're expecting. When your integral approximation has even a tiny error, the function can latch onto a large-integer vector that fits the noisy data better than your theoretical solution.

How to Get the Expected Vector #

Try these targeted fixes to guide lindep toward your desired solution:

1. Crank up the precision first #

Before calling lindep, set a higher real precision to minimize floating-point error:

default(realprecision, 50);  // Boost to 50 digits of precision
lindep([zeta(2), zeta(3), 11.3879638800312828875]);

This gives the function more data to work with, making it easier to spot the true small-integer dependency.

2. Tighten the error threshold #

lindep accepts a second parameter: a threshold for how close to zero the linear combination needs to be. Force it to only accept very tight fits:

lindep([zeta(2), zeta(3), 11.3879638800312828875], 1e-15);

This tells the function to ignore large-integer combinations that only approximate zero, and prioritize solutions that match your theoretical dependency closely.

3. Use exact theoretical values (if possible) #

If you already know the relationship from Wolfram Alpha, skip the approximation entirely by calculating the exact target value from the zeta functions:

exact_integral = (12*zeta(2) + 12*zeta(3))/3;
lindep([zeta(2), zeta(3), exact_integral]);

This eliminates approximation error entirely, so lindep will immediately return the clean [12,12,-3] vector.

Improving lindep for Integral Approximations #

For future integral-based linear dependency checks, use these best practices:

• Maximize integral precision: Calculate your integral to 30+ digits instead of 20. The more precise your approximation, the less likely lindep is to get distracted by noise.
• Pre-filter for small integers: If you suspect the solution has small integer coefficients, manually test candidate combinations first (e.g., check if 12*zeta(2) +12*zeta(3) -3*your_integral is near zero). Then use lindep with a tight threshold to confirm.
• Use lattice reduction directly: lindep uses the LLL algorithm under the hood. For more control, construct a lattice matrix manually with your scaled approximation values, then run qflll to prioritize small-integer solutions:

  // Scale values to convert to integer lattice (1e15 ensures precision)
  M = matrix(3,3, i,j, if(i==3, 1e15*vec([zeta(2), zeta(3), 11.3879638800312828875])[j], if(i==j,1,0)));
  qflll(M);  // Returns reduced lattice basis, including your desired small-integer vector
  

内容的提问来源于stack exchange，提问作者Max Lonysa Muller