stanfordnlp/CoreNLP · error · java.lang.IllegalArgumentException

Vector of incorrect size passed to applyInitialHessian in…

Error message

Vector of incorrect size passed to applyInitialHessian in QNInfo class

What it means

QNMinimizer.QNInfo.applyInitialHessian scales the input vector element-wise by the stored diagonal d. If the passed vector's length differs from d.length, the element-wise division is impossible, so it throws this IllegalArgumentException. If d is null, the vector is returned unscaled (no error).

Solutions

Pass an x vector whose length matches the QNInfo's stored d (same model dimension)
Create a fresh QNInfo/QNMinimizer for the new problem dimension instead of reusing one
Clear/reset the stored history (set d null or construct a new QNInfo) when the dimension changes

Example fix

// before
qnInfo.applyInitialHessian(new double[100]); // qnInfo built for dim 50
// after
double[] x = new double[50];
qnInfo.applyInitialHessian(x); // match stored dimension

Defensive patterns

Strategy: validation

Validate before calling

if (x.length != expectedDim) throw new IllegalArgumentException("x.length=" + x.length + " but QNInfo dimension=" + expectedDim);

Try / catch

try { qnInfo.applyInitialHessian(x); } catch (IllegalArgumentException e) { if (e.getMessage().contains("incorrect size passed to applyInitialHessian")) { qnInfo = new QNMinimizer.QNInfo(x.length); qnInfo.applyInitialHessian(x); } else throw e; }

Prevention

Never reuse QNInfo/QNMinimizer instances across problems of different dimension
Assert vector lengths equal the model dimension before optimizer calls
Reset optimizer state whenever the model size changes

When it happens

Trigger: Calling applyInitialHessian (directly or through the QNMinimizer pipeline) with a state vector x whose dimension differs from the dimension of the previously stored Hessian diagonal/surrogate history d.

Common situations: Reusing a QNInfo object across problems of different dimensionality; changing model size between optimization runs while reusing the same minimizer/QNInfo; passing a gradient subset instead of the full vector.

Understand the failure class

Background: Tensor shape mismatch errors ("must have shape", "expected shape ... got ..."): when tensor dimensions disagree with what an op or layer was told to expect — this error's family across 6 libraries.

Related errors

AI-assisted analysis of stanfordnlp/CoreNLP@1b7edd19c4 (2026-09-10). Data as JSON: /api/errors/8004855438d826e3. Report an issue: GitHub.

Appendix: source

Thrown at src/edu/stanford/nlp/optimization/QNMinimizer.java:724

      return used;
    }
  } // end class ScalarQNInfo

  class DiagonalQNInfo extends QNInfo {
    DiagonalQNInfo(int size) {
      super(size);
    }

    DiagonalQNInfo(List<double[]> sList, List<double[]> yList) {
      super(sList, yList);
    }

    double[] applyInitialHessian(double[] x, StringBuilder sb) {
      sb.append('D');
      if(d != null) {
        // Check sizes
        if(x.length != d.length)
          throw new IllegalArgumentException("Vector of incorrect size passed to applyInitialHessian in QNInfo class");
        // Scale element-wise
        for(int i = 0; i < x.length; i++)
          x[i] /= d[i];
      }
      return x;
    }

    int update(double[] newS, double[] newY, double yy, double sy, double sg, double step) {
      if(sy < 0) {
        // NOTE: if applying QNMinimizer to a non convex problem, we would still
        // like to update the matrix
        // or we could get stuck in a series of skipped updates.
        if(!quiet)
          log.info(" Negative curvature detected, update skipped ");
        return used;
      }
      if(yy == 0.0) {
        if(!quiet)

View on GitHub (pinned to 1b7edd19c4)