Generated methods for polynomial evaluation

2019-04-08 12:59发布

站内文章 / C#
34 0 0

问题:

I'm trying to come up with an elegant way to handle some generated polynomials. Here's the situation we'll focus on (exclusively) for this question:

  1. order is a parameter in generating an nth order polynomial, where n:=order + 1.
  2. i is an integer parameter in the range 0..n
  3. The polynomial has zeros at x_j, where j = 1..n and j ≠ i (it should be clear at this point that StackOverflow needs a new feature or it's present and I don't know it)
  4. The polynomial evaluates to 1 at x_i.

Since this particular code example generates x_1 .. x_n, I'll explain how they're found in the code. The points are evenly spaced x_j = j * elementSize / order apart, where n = order + 1.

I generate a Func<double, double> to evaluate this polynomial¹.

private static Func<double, double> GeneratePsi(double elementSize, int order, int i)
{
    if (order < 1)
        throw new ArgumentOutOfRangeException("order", "order must be greater than 0.");

    if (i < 0)
        throw new ArgumentOutOfRangeException("i", "i cannot be less than zero.");
    if (i > order)
        throw new ArgumentException("i", "i cannot be greater than order");

    ParameterExpression xp = Expression.Parameter(typeof(double), "x");

    // generate the terms of the factored polynomial in form (x_j - x)
    List<Expression> factors = new List<Expression>();
    for (int j = 0; j <= order; j++)
    {
        if (j == i)
            continue;

        double p = j * elementSize / order;
        factors.Add(Expression.Subtract(Expression.Constant(p), xp));
    }

    // evaluate the result at the point x_i to get scaleInv=1.0/scale.
    double xi = i * elementSize / order;
    double scaleInv = Enumerable.Range(0, order + 1).Aggregate(0.0, (product, j) => product * (j == i ? 1.0 : (j * elementSize / order - xi)));

    /* generate an expression to evaluate
     *   (x_0 - x) * (x_1 - x) .. (x_n - x) / (x_i - x)
     * obviously the term (x_i - x) is cancelled in this result, but included here to make the result clear
     */
    Expression expr = factors.Skip(1).Aggregate(factors[0], Expression.Multiply);
    // multiplying by scale forces the condition f(x_i)=1
    expr = Expression.Multiply(Expression.Constant(1.0 / scaleInv), expr);

    Expression<Func<double, double>> lambdaMethod = Expression.Lambda<Func<double, double>>(expr, xp);
    return lambdaMethod.Compile();
}

The problem: I also need to evaluate ψ′=dψ/dx. To do this, I can rewrite ψ=scale×(x_0 - x)(x_1 - x)×..×(x_n - x)/(x_i - x) in the form ψ=α_n×x^n + α_n×x^(n-1) + .. + α_1×x + α_0. This gives ψ′=n×α_n×x^(n-1) + (n-1)×α_n×x^(n-2) + .. + 1×α_1.

For computational reasons, we can rewrite the final answer without calls to Math.Pow by writing ψ′=x×(x×(x×(..) - β_2) - β_1) - β_0.

To do all this "trickery" (all very basic algebra), I need a clean way to:

  1. Expand a factored Expression containing ConstantExpression and ParameterExpression leaves and basic mathematical operations (end up either BinaryExpression with the NodeType set to the operation) - the result here can include InvocationExpression elements to the MethodInfo for Math.Pow which we'll handle in a special manner throughout.
  2. Then I take the derivative with respect to some specified ParameterExpression. Terms in the result where the right hand side parameter to an invocation of Math.Pow was the constant 2 are replaced by the ConstantExpression(2) multiplied by what was the left hand side (the invocation of Math.Pow(x,1) is removed). Terms in the result that become zero because they were constant with respect to x are removed.
  3. Then factor out the instances of some specific ParameterExpression where they occur as the left hand side parameter to an invocation of Math.Pow. When the right hand side of the invocation becomes a ConstantExpression with the value 1, we replace the invocation with just the ParameterExpression itself.

¹ In the future, I'd like the method to take a ParameterExpression and return an Expression that evaluates based on that parameter. That way I can aggregate generated functions. I'm not there yet. ² In the future, I hope to release a general library for working with LINQ Expressions as symbolic math.

回答1:

I wrote the basics of several symbolic math features using the ExpressionVisitor type in .NET 4. It's not perfect, but it looks like the foundation of a viable solution.

  • Symbolic is a public static class exposing methods like Expand, Simplify, and PartialDerivative
  • ExpandVisitor is an internal helper type that expands expressions
  • SimplifyVisitor is an internal helper type that simplifies expressions
  • DerivativeVisitor is an internal helper type that takes the derivative of an expression
  • ListPrintVisitor is an internal helper type that converts an Expression to a prefix notation with a Lisp syntax

Symbolic

public static class Symbolic
{
    public static Expression Expand(Expression expression)
    {
        return new ExpandVisitor().Visit(expression);
    }

    public static Expression Simplify(Expression expression)
    {
        return new SimplifyVisitor().Visit(expression);
    }

    public static Expression PartialDerivative(Expression expression, ParameterExpression parameter)
    {
        bool totalDerivative = false;
        return new DerivativeVisitor(parameter, totalDerivative).Visit(expression);
    }

    public static string ToString(Expression expression)
    {
        ConstantExpression result = (ConstantExpression)new ListPrintVisitor().Visit(expression);
        return result.Value.ToString();
    }
}

Expanding expressions with ExpandVisitor

internal class ExpandVisitor : ExpressionVisitor
{
    protected override Expression VisitBinary(BinaryExpression node)
    {
        var left = Visit(node.Left);
        var right = Visit(node.Right);

        if (node.NodeType == ExpressionType.Multiply)
        {
            Expression[] leftNodes = GetAddedNodes(left).ToArray();
            Expression[] rightNodes = GetAddedNodes(right).ToArray();
            var result =
                leftNodes
                .SelectMany(x => rightNodes.Select(y => Expression.Multiply(x, y)))
                .Aggregate((sum, term) => Expression.Add(sum, term));

            return result;
        }

        if (node.Left == left && node.Right == right)
            return node;

        return Expression.MakeBinary(node.NodeType, left, right, node.IsLiftedToNull, node.Method, node.Conversion);
    }

    /// <summary>
    /// Treats the <paramref name="node"/> as the sum (or difference) of one or more child nodes and returns the
    /// the individual addends in the sum.
    /// </summary>
    private static IEnumerable<Expression> GetAddedNodes(Expression node)
    {
        BinaryExpression binary = node as BinaryExpression;
        if (binary != null)
        {
            switch (binary.NodeType)
            {
            case ExpressionType.Add:
                foreach (var n in GetAddedNodes(binary.Left))
                    yield return n;

                foreach (var n in GetAddedNodes(binary.Right))
                    yield return n;

                yield break;

            case ExpressionType.Subtract:
                foreach (var n in GetAddedNodes(binary.Left))
                    yield return n;

                foreach (var n in GetAddedNodes(binary.Right))
                    yield return Expression.Negate(n);

                yield break;

            default:
                break;
            }
        }

        yield return node;
    }
}

Taking a derivative with DerivativeVisitor

internal class DerivativeVisitor : ExpressionVisitor
{
    private ParameterExpression _parameter;
    private bool _totalDerivative;

    public DerivativeVisitor(ParameterExpression parameter, bool totalDerivative)
    {
        if (_totalDerivative)
            throw new NotImplementedException();

        _parameter = parameter;
        _totalDerivative = totalDerivative;
    }

    protected override Expression VisitBinary(BinaryExpression node)
    {
        switch (node.NodeType)
        {
        case ExpressionType.Add:
        case ExpressionType.Subtract:
            return Expression.MakeBinary(node.NodeType, Visit(node.Left), Visit(node.Right));

        case ExpressionType.Multiply:
            return Expression.Add(Expression.Multiply(node.Left, Visit(node.Right)), Expression.Multiply(Visit(node.Left), node.Right));

        case ExpressionType.Divide:
            return Expression.Divide(Expression.Subtract(Expression.Multiply(Visit(node.Left), node.Right), Expression.Multiply(node.Left, Visit(node.Right))), Expression.Power(node.Right, Expression.Constant(2)));

        case ExpressionType.Power:
            if (node.Right is ConstantExpression)
            {
                return Expression.Multiply(node.Right, Expression.Multiply(Visit(node.Left), Expression.Subtract(node.Right, Expression.Constant(1))));
            }
            else if (node.Left is ConstantExpression)
            {
                return Expression.Multiply(node, MathExpressions.Log(node.Left));
            }
            else
            {
                return Expression.Multiply(node, Expression.Add(
                    Expression.Multiply(Visit(node.Left), Expression.Divide(node.Right, node.Left)),
                    Expression.Multiply(Visit(node.Right), MathExpressions.Log(node.Left))
                    ));
            }

        default:
            throw new NotImplementedException();
        }
    }

    protected override Expression VisitConstant(ConstantExpression node)
    {
        return MathExpressions.Zero;
    }

    protected override Expression VisitInvocation(InvocationExpression node)
    {
        MemberExpression memberExpression = node.Expression as MemberExpression;
        if (memberExpression != null)
        {
            var member = memberExpression.Member;
            if (member.DeclaringType != typeof(Math))
                throw new NotImplementedException();

            switch (member.Name)
            {
            case "Log":
                return Expression.Divide(Visit(node.Expression), node.Expression);

            case "Log10":
                return Expression.Divide(Visit(node.Expression), Expression.Multiply(Expression.Constant(Math.Log(10)), node.Expression));

            case "Exp":
            case "Sin":
            case "Cos":
            default:
                throw new NotImplementedException();
            }
        }

        throw new NotImplementedException();
    }

    protected override Expression VisitParameter(ParameterExpression node)
    {
        if (node == _parameter)
            return MathExpressions.One;

        return MathExpressions.Zero;
    }
}

Simplifying expressions with SimplifyVisitor

internal class SimplifyVisitor : ExpressionVisitor
{
    protected override Expression VisitBinary(BinaryExpression node)
    {
        var left = Visit(node.Left);
        var right = Visit(node.Right);

        ConstantExpression leftConstant = left as ConstantExpression;
        ConstantExpression rightConstant = right as ConstantExpression;
        if (leftConstant != null && rightConstant != null
            && (leftConstant.Value is double) && (rightConstant.Value is double))
        {
            double leftValue = (double)leftConstant.Value;
            double rightValue = (double)rightConstant.Value;

            switch (node.NodeType)
            {
            case ExpressionType.Add:
                return Expression.Constant(leftValue + rightValue);
            case ExpressionType.Subtract:
                return Expression.Constant(leftValue - rightValue);
            case ExpressionType.Multiply:
                return Expression.Constant(leftValue * rightValue);
            case ExpressionType.Divide:
                return Expression.Constant(leftValue / rightValue);
            default:
                throw new NotImplementedException();
            }
        }

        switch (node.NodeType)
        {
        case ExpressionType.Add:
            if (IsZero(left))
                return right;
            if (IsZero(right))
                return left;
            break;

        case ExpressionType.Subtract:
            if (IsZero(left))
                return Expression.Negate(right);
            if (IsZero(right))
                return left;
            break;

        case ExpressionType.Multiply:
            if (IsZero(left) || IsZero(right))
                return MathExpressions.Zero;
            if (IsOne(left))
                return right;
            if (IsOne(right))
                return left;
            break;

        case ExpressionType.Divide:
            if (IsZero(right))
                throw new DivideByZeroException();
            if (IsZero(left))
                return MathExpressions.Zero;
            if (IsOne(right))
                return left;
            break;

        default:
            throw new NotImplementedException();
        }

        return Expression.MakeBinary(node.NodeType, left, right);
    }

    protected override Expression VisitUnary(UnaryExpression node)
    {
        var operand = Visit(node.Operand);

        ConstantExpression operandConstant = operand as ConstantExpression;
        if (operandConstant != null && (operandConstant.Value is double))
        {
            double operandValue = (double)operandConstant.Value;

            switch (node.NodeType)
            {
            case ExpressionType.Negate:
                if (operandValue == 0.0)
                    return MathExpressions.Zero;

                return Expression.Constant(-operandValue);

            default:
                throw new NotImplementedException();
            }
        }

        switch (node.NodeType)
        {
        case ExpressionType.Negate:
            if (operand.NodeType == ExpressionType.Negate)
            {
                return ((UnaryExpression)operand).Operand;
            }

            break;

        default:
            throw new NotImplementedException();
        }

        return Expression.MakeUnary(node.NodeType, operand, node.Type);
    }

    private static bool IsZero(Expression expression)
    {
        ConstantExpression constant = expression as ConstantExpression;
        if (constant != null)
        {
            if (constant.Value.Equals(0.0))
                return true;
        }

        return false;
    }

    private static bool IsOne(Expression expression)
    {
        ConstantExpression constant = expression as ConstantExpression;
        if (constant != null)
        {
            if (constant.Value.Equals(1.0))
                return true;
        }

        return false;
    }
}

Formatting expressions for display with ListPrintVisitor

internal class ListPrintVisitor : ExpressionVisitor
{
    protected override Expression VisitBinary(BinaryExpression node)
    {
        string op = null;

        switch (node.NodeType)
        {
        case ExpressionType.Add:
            op = "+";
            break;
        case ExpressionType.Subtract:
            op = "-";
            break;
        case ExpressionType.Multiply:
            op = "*";
            break;
        case ExpressionType.Divide:
            op = "/";
            break;
        default:
            throw new NotImplementedException();
        }

        var left = Visit(node.Left);
        var right = Visit(node.Right);
        string result = string.Format("({0} {1} {2})", op, ((ConstantExpression)left).Value, ((ConstantExpression)right).Value);
        return Expression.Constant(result);
    }

    protected override Expression VisitConstant(ConstantExpression node)
    {
        if (node.Value is string)
            return node;

        return Expression.Constant(node.Value.ToString());
    }

    protected override Expression VisitParameter(ParameterExpression node)
    {
        return Expression.Constant(node.Name);
    }
}

Testing the results

[TestMethod]
public void BasicSymbolicTest()
{
    ParameterExpression x = Expression.Parameter(typeof(double), "x");
    Expression linear = Expression.Add(Expression.Constant(3.0), x);
    Assert.AreEqual("(+ 3 x)", Symbolic.ToString(linear));

    Expression quadratic = Expression.Multiply(linear, Expression.Add(Expression.Constant(2.0), x));
    Assert.AreEqual("(* (+ 3 x) (+ 2 x))", Symbolic.ToString(quadratic));

    Expression expanded = Symbolic.Expand(quadratic);
    Assert.AreEqual("(+ (+ (+ (* 3 2) (* 3 x)) (* x 2)) (* x x))", Symbolic.ToString(expanded));
    Assert.AreEqual("(+ (+ (+ 6 (* 3 x)) (* x 2)) (* x x))", Symbolic.ToString(Symbolic.Simplify(expanded)));

    Expression derivative = Symbolic.PartialDerivative(expanded, x);
    Assert.AreEqual("(+ (+ (+ (+ (* 3 0) (* 0 2)) (+ (* 3 1) (* 0 x))) (+ (* x 0) (* 1 2))) (+ (* x 1) (* 1 x)))", Symbolic.ToString(derivative));

    Expression simplified = Symbolic.Simplify(derivative);
    Assert.AreEqual("(+ 5 (+ x x))", Symbolic.ToString(simplified));
}


标签： c# linq math expression-trees algebra
举报

收藏的人(0)

Ta的文章 更多文章
登录 后发表评论
0条评论
还没有人评论过~