凸四边形区域的数值精度

时间:2014-02-14 16:37:10

标签: c# .net .net-4.5 numerical double-precision

我已经实现了一种计算R3中凸四边形区域的方法。该方法工作正常,但我在小数点后8位有数值精度问题。看一下方法:

internal static double GetTriangleArea(double ax, double ay, double az,
            double bx, double by, double bz,
            double cx, double cy, double cz
            )
        {
            /**
             * AB = B-A = (ux, uy, uz)
             * AC = C-A = (wx, wy, wz)
             * 
             * S = 0.5*sqrt{(uy*wz - uz*wy)² + (uz*wx - ux*wz)² + (ux*wy - uy*wx)²}
             * */

            var ux = bx - ax;
            var uy = by - ay;
            var uz = bz - az;

            var wx = cx - ax;
            var wy = cy - ay;
            var wz = cz - az;

            var t1 = uy*wz - uz*wy;
            var t2 = uz*wx - ux*wz;
            var t3 = ux*wy - uy*wx;
            var s = 0.5*Math.Sqrt(t1*t1 + t2*t2 + t3*t3);

            return s;
        }

        internal static double GetConvexQuadrilateralArea(double ax, double ay, double az,
            double bx, double by, double bz,
            double cx, double cy, double cz,
            double dx, double dy, double dz)
        {
            var triangle1 = GetTriangleArea(ax, ay, az, bx, by, bz, cx, cy, cz);
            var triangle2 = GetTriangleArea(ax, ay, az, cx, cy, cz, dx,dy,dz);

            return triangle1 + triangle2;
        }

这是测试:

[TestMethod]
        public void ParallelogramOfBaseBAndHeightHMustHaveAreaEqualToBTimesH()
        {
            var random = new Random(1);
            const double scale = 10000;
            for (var counter = 0; counter < 1000; counter++)
            {
                double baseLength = random.NextDouble() * scale;
                double height = random.NextDouble() * scale;

                double dx = random.NextDouble()*scale;

                var a = new[] { 0, 0, 0 };
                var b = new[] { baseLength, 0, 0 };
                var c = new[] { baseLength+dx, height, 0 };
                var d = new[] { 0F+dx, height, 0 };

                double expected = baseLength * height;

                var result = MathUtils.GetConvexQuadrilateralArea(a[0], a[1], a[2], b[0], b[1], b[2], c[0], c[1], c[2],
                    d[0], d[1], d[2]);

                Assert.AreEqual(expected, result, Epsilon*scale,
                    string.Format("sideA: {0}, height: {1}, dx: {2}", baseLength, height, dx));
            }
        }

此测试失败,并显示以下消息:预期差异不大于&lt; 1E-09&gt;预期值74813926.2967871与实际值74813926.2967871之间。 sideA:8552.44307245707,身高:8747.66726454146,dx:4721.64729829954。

我的问题是:有没有办法在仍然使用双精度数字时提高实现的数值精度?

1 个答案:

答案 0 :(得分:2)

双打有52位分数。在十进制中,大约是15.7位数。您的数字在点的左侧有八位数字,因此您只能期望在点的右侧有七位正确 - 这只是由于表示,甚至没有考虑计算产生的任何累积误差

所以,答案是:不,没有办法。您必须使用更精确的格式。