Table 1 lists all monomials up to total degree three, indicating their membership in Q 4 and Q 8 , and identifies the degree of completeness of each space.

Table 1. Monomial content of Q4 and Q8 approximation spaces.

The degree of completeness is the largest integer p such that every monomial of total degree ≤ p belongs to the space; Q4 achieves p = 1 (missing ξ 2 and η 2 ) while Q8 achieves p = 2. This distinction governs approximation order, as made precise in Section 2.6 [1][6][7]. Figure 1 illustrates the node configurations for both elements on the master element.

Figure 1. Q4 and Q8 Node Layout. Node configurations for Q4 and Q8 interpolation elements on the normalized master element [−1, 1]². Q4 uses four corner nodes; Q8 augments these with four edge-midpoint nodes.

Figure 1 establishes the geometric foundation for the entire comparison. The left panel shows the Q4 element, which samples the surface only at its four corners: (−1, −1), (1, −1), (1, 1), and (−1, 1). Because each edge of the Q4 element is controlled by exactly two endpoint values, the interpolated profile along any edge is constrained to be a straight line. This means that any curvature present in the true surface between two corners is invisible to Q4—it has no nodal information from which to detect or reconstruct bending.

The right panel shows the Q8 element, which retains all four corner nodes and adds one midpoint node on each of the four edges: (0, −1), (1, 0), (0, 1), and (−1, 0). This seemingly modest addition has a profound effect. Each edge is now controlled by three values, i.e., two endpoints and one midpoint, which is the minimum information needed to fit a parabola. As a result, Q8 can sense and exactly reproduce quadratic edge curvature, and can approximate more general smooth edge curvature more accurately than Q4. The figure makes it visually clear that the difference between Q4 and Q8 is not a global modeling choice but a local sampling one: Q8 simply captures more geometric information per element by reading the surface at its edge midpoints.

The four Q4 nodes are placed at the element corners:

Given nodal values z k = z ( ξ k , η k ) for k = 1 , 2 , 3 , 4 , the Q4 interpolant is

where the bilinear shape functions are

These functions satisfy two fundamental properties, the Kronecker delta and the partition of unity. The Kronecker-delta property states

ensuring that the interpolant passes through the nodal values exactly. The partition of unity property states

which guarantees that constant fields are reproduced exactly and is a necessary condition for consistency [1][2].

The Q8 element uses the four corner nodes of Q4 plus four edge-midpoint nodes:

Given nodal values z k = z ( ξ k , η k ) for k = 1 , ⋯ , 8 , the Q8 interpolant is

The serendipity shape functions for corner nodes are

and for edge-midpoint nodes are

The Q8 shape functions satisfy the same Kronecker-delta and partition of unity properties [1][2]. Their polynomial span is Q 8 , which is complete to degree two [1][2][8].

It is mathematically expected that Q8 can outperform Q4 on curved surfaces because Q8 has more nodal information and a richer approximation space [9]. Therefore, the purpose of this paper is to quantify the geometric consequence of the added Q8 degrees of freedom. The comparison asks whether the extra nodal information preserves quantities that matter in surface reconstruction: height, curvature, and surface orientation. This distinction is important for computational decision-making. In many applications, Q4 may be sufficient. If the surface is nearly planar over the element, or if only corner samples are available, Q4 is efficient and appropriate. If the surface is locally smooth but curved, and if midpoint samples are available or can be estimated reliably, Q8 may be justified because it preserves curvature and normals more accurately. Modifications to the standard Q8 element have also been studied [10].

The closed-form shape functions used here may be viewed as the inverse of a fixed nodal collocation system on the reference element. Since the nodes are fixed and the nodal functionals are unisolvent on the stated polynomial spaces, the reference interpolation operators are bounded. Under affine scaling to shape-regular physical elements, these constants remain independent of the element size h .

The fundamental exactness property of both interpolants is stated as follows.

Theorem1(PolynomialreproductionofQ4andQ8). Let Ω = [ − 1 , 1 ] 2 be the reference square and let Q 4 and Q 8 denote the polynomial spaces defined in Equation (1) and Equation (2), respectively.

(i) The Q4 interpolation operator ℐ Q 4 reproduces every polynomial in Q 4 .

(ii) The Q8 interpolation operator ℐ Q 8 reproduces every polynomial in Q 8 .

(iii) There exist polynomials f ∈ Q 8 \ Q 4 that are not reproduced by ℐ Q 4 .

Proof. Consider first the Q4 element. The reference nodes { ( ξ k , η k ) } k = 1 4 are the four corners of Ω, and the Q4 shape functions { N k } k = 1 4 defined in Equation (5a-5d) are Lagrange shape functions satisfying the Kronecker delta property

and the partition of unity ∑ k = 1 4 N k ≡ 1 . The corresponding nodal evaluation functionals ℓ k : C 0 ( Ω ) → ℝ , ℓ k ( f ) = f ( ξ k , η k ) , are linearly independent on the four–dimensional space Q 4 ; hence they form a unisolvent set of degrees of freedom for Q 4 . Equivalently, the only polynomial p ∈ Q 4 satisfying ℓ k ( p ) = 0 for all k is the zero polynomial.

By definition, the Q4 interpolation operator ℐ Q 4 : C 0 ( Ω ) → Q 4 is given by

so that ℓ k ( ℐ Q 4 f ) = f ( ξ k , η k ) for all k . Now let f ∈ Q 4 . Then both f and ℐ Q 4 f belong to Q 4 and satisfy

By unisolvency, f = ℐ Q 4 f on Ω, which proves (i).

The proof of (ii) is analogous for Q8. The eight serendipity shape functions { ϕ k } k = 1 8 defined in Equation (10) and Equation (11) satisfy

at the eight Q8 nodes, and ∑ k = 1 8 ϕ k ≡ 1 . The nodal evaluation functionals ℓ k ( f ) = f ( ξ k , η k ) are unisolvent for the eight–dimensional polynomial space Q 8 . The Q8 interpolation operator

therefore satisfies ℐ Q 8 f = f for every f ∈ Q 8 , establishing (ii).

For (iii), consider the dome

which lies in Q 8 but not in Q 4 because it contains the pure quadratic monomials ξ 2 and η 2 absent from Q 4 . At each Q4 corner node ± 1 , ± 1 we have

so ℐ Q 4 f is the unique polynomial in

that vanishes at all four corners. Any bilinear polynomial with four zero corner values is identically zero on Ω, hence ℐ Q 4 f ≡ 0 . Since f ( 0 , 0 ) = 1 , it follows that ℐ Q 4 f ≠ f , and in particular ‖ f − ℐ Q 4 f ‖ L ∞ ( Ω ) ≥ 1 . This shows that Q4 does not reproduce all functions in

, completing the proof. □

Remark1. Since

, Q4 cannot reproduce any function

exactly over the element. However, for some functions outside the Q4 span, the resulting error may be small or acceptable for a given application.

Consistency. Both Q4 and Q8 satisfy the consistency condition for interpolation on the reference element: they exactly reproduce all constant and linear fields. This is guaranteed by the partition of unity and Kronecker-delta properties of the shape functions. In the interpolation setting considered here, the relevant analog of the finite element patch test is the exact reproduction of constant and linear fields, which both elements satisfy [1][6].

Stability. The closed-form shape functions for Q4 and Q8 may be viewed as the inverses of fixed nodal collocation systems on the reference element. For Q4, the four shape functions are the unique bilinear basis dual to the four corner nodal functionals. For Q8, the eight shape functions are the unique serendipity basis dual to the eight nodal functionals (four corners and four edge midpoints). In both cases, the nodal functionals are unisolvent on their respective polynomial spaces, meaning the only function in the space that vanishes at all nodes is the zero function [7][8]. This unisolvence guarantees that the interpolation operators ℐ Q 4 and ℐ Q 8 are well-defined and bounded on the reference element. Under affine scaling to shape-regular physical elements, the operator norms remain bounded independently of the element size h [7][11].

Remark2. No Vandermonde system is solved during interpolation at runtime. The shape functions are precomputed in closed form, and interpolation at any point ( ξ , η ) requires only the evaluation of eight scalar functions and a weighted sum. The stability statement above refers to the conditioning of the underlying nodal collocation map, not to an iterative or linear-system computation.

Theorem2(ElementwiseinterpolationerrorforQ4andQ8). Let { K h } be a family of shape-regular quadrilateral elements obtained by affine images of the reference square Ω = [ − 1 , 1 ] 2 , with element diameter h = diam ( K h ) . Let ℐ h denote either the Q4 or Q8 nodal interpolation operator, defined by Equation (4) or Equation (9) on Ω and extended to K h by composition with the reference mapping.

Assume f ∈ H p + 1 ( K h ) , where p is the polynomial reproduction degree of the interpolant ( p = 1 for Q4, p = 2 for Q8; see Table 1).

Then there exists a constant C > 0 , independent of h and f , such that for j = 0 , 1 , 2 the following bound holds:

Proof. On the reference element Ω, the Q4 and Q8 interpolation operators ℐ Q 4 and ℐ Q 8 are linear projectors onto

and

, respectively. By Theorem 1 and the degree-of-completeness summary in Table 1, ℐ Q 4 reproduces all polynomials of degree at most p = 1 , and ℐ Q 8 reproduces all polynomials of degree at most p = 2 .

Each ℐ h is defined elementwise by pullback from Ω: if F h : Ω → K h is the affine map, then ℐ h f = ( ℐ r e f ( f ∘ F h ) ) ∘ F h − 1 , where ℐ r e f is ℐ Q 4 or ℐ Q 8 on Ω. The shape-regularity of { K h } implies that the Jacobians of F h and F h − 1 are uniformly bounded in h , and the norms ‖   ⋅   ‖ H m ( K h ) and ‖   ⋅   ‖ H m ( Ω ) are equivalent up to constants depending only on the shape-regularity and on m .

On Ω, the operators ℐ Q 4 and ℐ Q 8 are bounded linear maps on H p + 1 ( Ω ) , with operator norms independent of h . They reproduce all polynomials of degree at most p . Therefore, the hypotheses of the Bramble–Hilbert lemma are satisfied on each element K h : there exists a polynomial q ∈ ℙ p ( K h ) such that

with C 1 independent of h and f .

Since ℐ h reproduces all polynomials of degree at most p , we have ℐ h q = q . Thus,

Taking the H j ( K h ) norm and using the linearity and boundedness of ℐ h on H p + 1 ( K h ) , there exists C 2 > 0 , independent of h and f , such that

Combining this with the Bramble–Hilbert estimate above yields

where C = ( 1 + C 2 ) C 1 depends only on the element shape-regularity and the interpolation operator, but not on h or f . This proves Equation (19). □

Corollary1(ErrorordersforQ4andQ8). For Q4, p = 1 , so Equation (19) gives

whereas for Q8, p = 2 , one obtains

These convergence orders are summarized in Table 2 and, in particular, show that Q8 achieves one higher order of convergence than Q4 in each Sobolev norm level j = 0 , 1 , 2 .

Table 2. Theoretical convergence orders for Q4 and Q8 in Sobolev norms.

^†The O ( h 0 ) bound means the error estimate does not decrease with h; it does not imply the error grows or diverges.

Remark3(Scopeofthenumericalexperiments). The single-element experiments in Section 6 confirm polynomial reproduction and single-element approximation behavior on the fixed master element with h = 2 . They do not directly demonstrate asymptotic convergence orders, which are statements about a family of elements with h → 0 and would require a mesh-refinement study. The error ratios observed in Table 3 are consistent with the qualitative advantage predicted by the bounds in Theorem 2 and Table 2, but they should not be interpreted as empirical measurements of convergence rates. A mesh-refinement study is therefore identified as a direction for future work in Section 9.

A height field z = f ( ξ , η ) over the master element defines a surface patch through the Monge (graph) parameterization [12]:

The two tangent vectors are

The upward-pointing surface normal is obtained as their cross product:

This derivation shows that the formula n = [ − f ξ , − f η , 1 ] ⊤ is not an arbitrary choice but the upward-pointing normal direction to the Monge patch, obtained by convention. The unit normal is

The first fundamental form of the Monge patch has coefficients

and the second fundamental form has coefficients

The Hessian of the height field z = f ( ξ , η ) is the symmetric matrix

The Hessian-based curvature proxy is the Frobenius norm of H f :

In the small-slope regime where f ξ 2 + f η 2 ≪ 1 , the shape operator simplifies to S ≈ H f and the following relations hold:

These relations hold in the small-slope approximation. For surfaces with larger gradients, κ H remains a well-defined and interpolation-sensitive second-derivative measure, but it should be interpreted as a height-field curvature proxy rather than the exact principal curvature magnitude of the Monge surface.

The proxy κ H is preferred over the Laplacian Δ f = f ξ ξ + f η η for the following reason. Consider the saddle z = ξ 2 / 2 − η 2 / 2 , which has f ξ ξ = 1 and f η η = − 1 . Then Δ f = 0 identically, incorrectly suggesting zero curvature, while

correctly indicating strong bending. The curvature proxy error metric is therefore

The principal curvatures κ 1 and κ 2 are the eigenvalues of the shape operator S = I − 1 I I , where I and I I are the matrices of the first and second fundamental forms. For the dome surface z = 1 − ( ξ 2 + η 2 ) / 2 , the Hessian eigenvalues of the height field are − 1 , − 1 , giving a constant Hessian-based curvature proxy κ H = 2 . In the small-slope regime, the Hessian eigenvalues approximate the principal curvatures, but the exact principal curvatures of the Monge surface vary with ( ξ , η ) because the first fundamental form depends on the gradient.

The pointwise height error is

The height RMSE over the discrete grid Ω h of 201 × 201 points is

providing a discrete approximation to the L 2 ( Ω ) norm of the interpolation error and directly connecting the numerical measure to the theoretical L 2 bound of Theorem 2.

Using the Monge patch formulation of Section 3.1, the unit normal to the reconstructed surface is

and the normal-angle error at each grid point is

reported in degrees. Since n depends on the first derivatives of z , the normal-angle error is governed by the H 1 interpolation error. By Theorem 2, Q8 provides O ( h 2 ) convergence in H 1 versus O ( h ) for Q4, predicting systematically smaller normal-angle errors for Q8, as confirmed in Table 3.

This surface belongs to

but not to

. By Theorem 1 (ii), Q8 reproduces it exactly; by part (iii), Q4 does not. The Hessian is

giving Hessian eigenvalues −1 and −1, a constant Hessian-based curvature proxy κ H = 2 , and a small-slope approximation to mean curvature of −1. The exact principal curvatures of the Monge surface vary with (ξ, η) because the first fundamental form depends on the gradient; the Hessian eigenvalues equal the principal curvatures only in the small-slope regime.

This surface belongs to

. The term ξ η is in

; the terms ξ 2 / 3 and − η 2 / 3 are not. The Hessian is

The small-slope approximation to Gaussian curvature (the Hessian determinant) is

confirming the saddle character. (The exact Gaussian curvature of the Monge surface includes a denominator of ( 1 + f ξ 2 + f η 2 ) 2 and is not identically equal to det ( H f ) .) The Hessian curvature magnitude is

This surface is outside both Q4 and Q8 polynomial spaces. Its Taylor expansion near ( 0 , 0 ) begins

which contains cubic and higher terms not present in either

or

. The ripple is included to test approximation behavior rather than exact polynomial reproduction. This is important because real surfaces in depth imaging, scanning, and biomedical reconstruction are rarely exact low-order polynomials.

The Hessian is

which varies strongly over Ω, making curvature preservation a non-trivial challenge for a single-element interpolant.

Figure 2 shows the dome surface reconstruction comparison panels. The dome surface z = 1 − ( ξ 2 + η 2 ) / 2 is a smooth, symmetric, concave-down paraboloid. Its defining feature is pure quadratic curvature in both coordinate directions, with a peak at the origin and equal values of zero at all four corners. Because all four corners share the same height (z = 0), Q4 receives identical corner values and can extract no curvature information from them whatsoever. The Q4 bilinear polynomial, which spans only {1, ξ, η, ξη}, is therefore forced to reconstruct a flat surface—the bilinear interpolant through four equal corner values is identically zero everywhere, giving a height RMSE of 0.6967 and a maximum height error of 1.0000 at the dome’s peak. The Q4 reconstruction panel in the center confirms this: the dome's central bulge is entirely absent, replaced by a flat or nearly flat surface.

Figure 2. Dome surface comparison. True dome surface and Q4 and Q8 reconstructions. Q4 collapses the quadratic curvature to a near-flat bilinear surface; Q8 reproduces the dome to numerical precision.

The Q8 reconstruction panel on the right tells the opposite story. The midpoint nodes sample the dome at its edges, where z takes non-zero values that directly encode the parabolic profile. Q8’s polynomial span includes ξ 2 and η 2 , so it possesses exactly the terms needed to fit the dome. The result is exact reproduction to machine precision: a height RMSE of 6.21 × 10⁻¹⁷ and a normal-angle error of 2.49 × 10⁻⁷ degrees, both numerically zero. This figure is the clearest illustration in the paper of the categorical difference between the two interpolants: Q4 fails completely on a surface it cannot represent, while Q8 succeeds exactly.

Figure 3 shows the saddle surface reconstruction comparison panels. The saddle surface z = ξ η + ( ξ 2 − η 2 ) / 3 is a more demanding and nuanced test than the dome, as it contains two structurally distinct components.

Figure 3. Saddle surface comparison. True saddle surface and Q4 and Q8 reconstructions. Q4 recovers the bilinear twisting component but fails to reproduce the quadratic curvature terms; Q8 reconstructs the full saddle to numerical precision.

The first is the bilinear term ξ η , which belongs to the Q4 polynomial span. The second is the quadratic terms ξ 2 and η 2 , which do not. This mixture means Q4 is not completely blind to the saddle; it can detect and partially reproduce the twisting behavior, but it systematically misses the additional bending captured by the quadratic terms.

The Q4 reconstruction panel shows a surface with the correct general orientation and twisting character of a saddle, yet it visibly deviates from the true shape, particularly away from the element center. The resulting height RMSE is 0.1419 and the normal-angle error is 20.97 degrees. These errors are smaller than the dome case, consistent with the fact that Q4 at least captures the ξ η component, but they remain substantial and would meaningfully distort any downstream geometric measurement.

The Q8 reconstruction again achieves machine-precision accuracy, with height RMSE of 6.78 × 10⁻¹⁷ and normal-angle error of 2.19 × 10⁻⁷ degrees. This confirms that Q8’s polynomial span encompasses both the bilinear and the quadratic components of the saddle simultaneously. The saddle experiment is important because it demonstrates that partial polynomial overlap, Q4 capturing ξ η but not ξ 2 or η 2 , does not translate into acceptable geometric fidelity. An interpolant must contain the full polynomial content of the surface to reconstruct it accurately. The saddle belongs to

. The Q4 height RMSE of 0.1419 is smaller than the dome case because Q4 captures the ξ η bilinear component. Nevertheless, the normal-angle error of 20.97˚ indicates substantial geometric distortion from the missing quadratic terms. Q8 achieves machine-precision accuracy across all metrics, confirming Theorem 1 (ii).

Figure 4 shows the transcendental ripple reconstruction comparison panels. The ripple surface z = 1 / 2 sin ( 2 ξ ) cos ( 2 η ) is the most practically important of the three test cases because it lies outside the polynomial span of both Q4 and Q8. No finite serendipity polynomial can represent a transcendental function exactly, so this experiment measures approximation quality rather than exact reproduction. This situation is representative of real-world surfaces encountered in depth imaging, structured-light scanning, and biomedical reconstruction, which are almost never exact low-order polynomials.

Figure 4. Transcendental ripple comparison. True transcendental ripple surface and Q4 and Q8 reconstructions. Neither method is exact, but Q8 substantially reduces height and shape error relative to Q4.

The Q4 reconstruction panel shows a severely smoothed and distorted version of the ripple. Q4’s bilinear form cannot capture the oscillatory curvature of the ripple, producing a height RMSE of 0.3310 and a normal-angle error of 37.47 degrees—the largest normal error among all test cases. The surface orientation is almost entirely wrong over much of the element.

The Q8 reconstruction is not exact, as expected, but it is substantially better. The reconstructed surface follows the ripple’s shape far more closely, reducing height RMSE to 0.0843 (a 74.5% reduction or a factor of approximately 3.93) and normal-angle error to 15.02 degrees. The maximum height error falls from 0.6576 to 0.1955. Visually, the Q8 panel captures the curvature undulation of the ripple in a way that Q4 does not. This result supports a balanced conclusion: Q8 does not solve the approximation problem for arbitrary smooth surfaces, but it provides a meaningfully better local geometric model. For applications where approximate but curvature-aware interpolation is sufficient, such as filling small missing-depth regions on smooth surfaces, Q8 offers a practical advantage over Q4 even when the surface is non-polynomial.

Since

, both interpolants are approximate. Q8 reduces the height RMSE from 0.3310 to 0.0843, a factor of approximately 3.93. This ratio is consistent with the expected qualitative advantage of Q8 over Q4 but should be interpreted as a single-element approximation rather than a direct measure of the asymptotic convergence order. A separate mesh-refinement study with h→0 would be required to estimate convergence rates numerically. The normal-angle error decreases from 37.47˚ to 15.02˚, and the curvature MAE decreases from 1.8504 to 0.9856, corresponding to a 46.7% reduction.

Figure 5 shows the Hessian curvature magnitude error maps for all three surfaces. The Hessian curvature magnitude κ H = ( z ξ ξ ) 2 + 2 ( z ξ η ) 2 + ( z η η ) 2 measures the local bending strength of a surface at every point. Unlike the Laplacian, which can mask saddle-like curvature through cancellation of positive and negative components, the Hessian magnitude provides a reliable scalar indicator of how strongly the surface bends regardless of the sign of curvature.

The left column of each row in Figure 5 shows the true κ H field. For the dome, κ H is nearly uniform across the element, reflecting its constant quadratic bending. For the saddle, κ H is spatially varying, with stronger bending near the edges due to the combined quadratic and bilinear structure. For the ripple, κ H is highly spatially varying, reflecting the oscillatory nature of the true surface.

Figure 5. Hessian curvature magnitude error maps. Color maps of true Hessian curvature magnitude κ H and absolute curvature error for Q4 and Q8 across all three test surfaces. Q4 errors are large and spatially structured; Q8 errors are near zero for the dome and saddle, and substantially reduced for the ripple.

The middle column shows Q4 curvature error. For the dome, Q4’s error is large and spatially uniform—since Q4 produces a flat reconstruction, it contributes zero curvature everywhere and the error equals the true κ H at every point, giving a curvature MAE of 1.4142. For the saddle, Q4 errors are concentrated in regions where the quadratic terms dominate over the bilinear term. For the ripple, Q4 curvature error is large and spatially structured, with a MAE of 1.8504, larger even than on the dome or saddle because the ripple’s oscillatory curvature is entirely outside Q4’s reach.

The right column shows Q8 curvature error. For the dome and saddle, Q8 error maps are visually indistinguishable from zero, the color scales register only machine-precision residuals (MAE on the order of 10⁻¹³). For the ripple, Q8 error is visible but meaningfully smaller than Q4, with MAE falling from 1.8504 to 0.9856. The shared color scale used for the Q4 and Q8 error columns within each row ensures that the visual comparison is fair. This figure reinforces the key finding: Q8 preserves not only surface height but also the second-derivative structure that defines local shape.

Figure 6 shows the centerline profiles. It shifts attention from surface height to second-derivative geometry. Figure 6 provides a one-dimensional cross-sectional view of the reconstruction quality, slicing each surface along the centerline η = 0. This representation complements the 3D surface plots and 2D error maps by showing the exact numerical profile of height and curvature along a single line, making it easy to see where and by how much each method deviates from the truth.

Dome (top row):The left panel shows the true dome profile along η = 0 , which is the parabola z = 1 − ξ 2 / 2 , with a peak of 1.0 at the origin and values of 0.5 at ξ = ± 1 . The Q4 profile is the horizontal line z = 0 , because all four corner nodes of the full element have a value of zero. Along the centerline, this misses both the peak at ξ = 0 and the true centerline endpoint values z = 0.5 at ξ = ± 1 . The Q8 profile overlays the true curve exactly. The curvature panel shows the same pattern: Q8 matches the constant Hessian curvature proxy, while Q4 contributes zero curvature.

Saddle (middle row):Along η = 0 , the true saddle reduces to z = ξ 2 / 3 , so Q4 misses the parabolic centerline curvature entirely, while Q8 reproduces it exactly. The Q4 interpolant captures the bilinear component ξ η , which vanishes on this centerline, so the Q4 profile is the horizontal line z = 0 , missing the parabolic profile. Q8 follows the true curve exactly, and the curvature panel confirms that Q8 preserves the nonzero ξ-direction centerline curvature ( d 2 z / d ξ 2 = 2 / 3 along η = 0) while Q4 does not ( d 2 z Q 4 / d ξ 2 = 0 on this centerline, since z_Q4 = ξη vanishes identically when η = 0). It is important to note, however, that the full 2D Hessian proxy for Q4 is nonzero because Q4 retains the mixed-derivative contribution from the ξη term. The curvature MAE of 0.285 in Table 3 reflects the difference between Q4 and the true Hessian-based curvature magnitude.

Ripple (bottom row): The ripple centerline z = 1/2sin(2ξ) is an oscillating curve that neither method reproduces exactly. However, the Q8 profile follows the oscillation much more closely than Q4, tracking the peaks and troughs with substantially smaller deviation. The curvature profile shows the same: Q8 captures the sign and approximate magnitude of the curvature variation, whereas Q4 produces a much flatter, less informative curvature estimate. The remaining discrepancy between Q8 and the true ripple is an honest representation of the approximation error inherent to any finite polynomial element on a transcendental surface.

Figure 6. Centerline profiles. Surface height and Hessian curvature magnitude profiles along η = 0 for all three test surfaces, comparing the true surface against Q4 and Q8 reconstructions.

The ripple surface is the most application-relevant test because it lies outside both interpolation spaces. Neither Q4 nor Q8 can reconstruct it exactly. Nevertheless, Q8 substantially improves the approximation. The height RMSE decreases from 0.3310 for Q4 to 0.0843 for Q8, a reduction of approximately 74.5%. The maximum height error decreases from 0.6576 to 0.1955. The normal-angle error decreases from 37.47 degrees to 15.02 degrees. The curvature proxy error also decreases, although it remains nonzero. This is expected because the ripple contains transcendental variation that cannot be represented exactly by Q8. The observed error ratio of approximately 3.93 is consistent with the expected qualitative advantage of Q8 over Q4, but it should be interpreted as a single-element approximation result rather than a direct measurement of asymptotic convergence order. A separate mesh-refinement study with h → 0 would be required to estimate convergence rates numerically.

Q4 evaluates 4 shape functions; Q8 evaluates 8. Both remain O ( 1 ) per interpolation point and require no iteration, global system assembly, or training. The more consequential cost is the requirement of reliable edge-midpoint nodal values for Q8. If these are unavailable or noisier than corner samples, the Q8 approximation advantage may be compromised.

The criterion below is valid under three conditions. First, the surface must be sufficiently smooth: f ∈ H 2 ( Ω ) for the Q4 bounds and f ∈ H 3 ( Ω ) for the Q8 bounds, so that the Sobolev seminorms in Theorem 2 are finite. Second, the element mapping must be affine (or at least regularly bounded), so that the Bramble–Hilbert constants scale independently of h. Third, the Q8 midpoint nodal values must be reliable; as noted in Table 4, high noise at midpoint nodes undermines the approximation advantage that motivates the Q8 choice. When any of these conditions are violated, the table below should be applied with caution.

The theoretical comparison in Table 2 shows that Q8 achieves one higher order of convergence than Q4 at every Sobolev norm level: L 2 (heights), H 1 (slopes and normals), and H 2 (curvature). Together with the Mongepatch curvature analysis in Section 3, this yields a clear set of selection criteria for sparse surface reconstruction:

Height-onlyaccuracy( L 2 ). Q4 achieves O ( h 2 ) convergence in L 2 , whereas Q8 achieves O ( h 3 ) . When the application uses only height information on locally planar or bilinear patches, the extra order of Q8 may not justify its additional nodal and computational cost, and Q4 is typically sufficient.Normalandslopeaccuracy( H 1 ). Surface normals depend on first derivatives of the height field, so normal errors are governed by the H 1 interpolation error. Q4 yields O ( h ) convergence in H 1 , whereas Q8 yields O ( h 2 ) . For tasks that use normals downstream—such as shading, registration, normal-based matching, or biomechanical analysis—Q8 is theoretically preferred, because it provides faster decay of orientation error under mesh refinement.Curvatureandshapeaccuracy( H 2 ). The Hessian-based curvature proxy κ H is equivalent to the H 2 seminorm of the Monge-patch height field, so curvature errors are governed by the H 2 interpolation error. In this norm, Q4 has only an O ( h 0 ) bound, meaning the theoretical curvature-level error does not decrease with refinement, while Q8 attains O ( h ) . For curvature-sensitive reconstruction, or whenever bending and local shape carry physical meaning, Q8 is the better-suited interpolant within the Q4 - Q8 family [7][13][14].

In practical terms, Q4 is appropriate for height-only interpolation on locally planar or bilinear elements, especially when only corner samples are available or computational cost is dominant. Q8 is justified when reliable edge-midpoint samples are available and higher-order control of surface normals or curvature is required, as in the dome, saddle, and ripple examples in Section 6.

Table 4 aggregates the theoretical and numerical findings into practical guidance for choosing between Q4 and Q8 in sparse surface reconstruction. Rather than presenting Q8 as universally superior, the table highlights the specific conditions under which each interpolant is mathematically and computationally appropriate. Rows associated with height accuracy reflect the L 2 orders from Theorem 2; rows involving normals and edge geometry reflect the H 1 behavior; and rows involving curvature or the Hessian-based proxy κ H reflect the H 2 behavior. This makes the engineering recommendations traceable directly back to the Sobolev-norm comparison in Table 2 and the curvature analysis in Section 3.

Table 4. Practical selection criteria for Q4 and Q8 interpolation.

The first two rows address locally planar and bilinear surfaces. In these cases, Q4 is sufficient, and Q8 is unnecessary. Q4’s polynomial span exactly contains flat and bilinear fields, so adding midpoint nodes provides no additional accuracy while increasing sampling cost. In nearly planar regions, common in man-made environments, flat tissue surfaces, or background depth maps, Q4 is the pragmatic choice.

Rows three and four address curvature and edge shape. When pure bending or parabolic edge profiles are present, Q8 is clearly preferred. These rows capture the core finding of the paper: Q8’s inclusion of ξ 2 and η 2 in its polynomial span is the critical structural advantage over Q4 for curved-surface reconstruction.

Rows five through seven address downstream usage and data availability. If surface normals are used downstream—for rendering, registration, shape matching, or biomechanical analysis—Q8 is preferred because it produces more accurate normals. If only corner samples are available, Q8 cannot be used without midpoint estimation; Q4 is then the appropriate choice. If midpoint samples are available, Q8 should be considered.

Row eight addresses computational simplicity. In real-time systems, embedded implementations, or very large-scale reconstructions where per-element cost matters, Q4’s lower arithmetic count may be decisive. Q8’s extra shape functions are modest in absolute terms, but in computationally constrained settings, Q4 may still be preferred.

Rows nine through eleven handle special cases. Smooth anatomical or depth surfaces—the primary application targets—favor Q8 when curvature matters. Sharp discontinuities, such as object boundaries, anatomical folds, or depth edges, should not be handled by either Q4 or Q8 alone; edge-aware segmentation or boundary-detection pre-processing is needed before any element-based interpolation. High nodal noise warrants caution: while Q8 provides better curvature fidelity with clean data, using more samples can amplify noise if the midpoint measurements are unreliable, and regularized or filtered variants may be needed.