Did the Method Really Resolve the Gibbs Phenomenon?

The Approximation Mechanism 逼近機制

The Gibbs phenomenon arises from the approximation mechanism. An integral criterion such as an L² projection, a least-squares principle, or another variational minimization does not impose pointwise range constraints when approximating a discontinuous or nearly discontinuous function.

Gibbs 現象來自逼近機制。對間斷或近乎間斷的函數，L² 投影、最小二乘或其他積分型變分極小化只控制整體誤差，並不約束逐點值域。

For a continuous approximation space, the approximant must connect values across a jump that it cannot represent. An integral projection controls an average error; it does not impose a pointwise range constraint. Thus an approximation may converge in L² while retaining a non-trivial pointwise overshoot.

One Phenomenon, Many Names 同一現象，不同名稱

A paper may avoid the name “Gibbs phenomenon” while discussing the same pointwise failure under another label.

一篇論文可能避開「Gibbs 現象」這個名稱，卻用另一個標籤討論同一類點值失控。

These terms are not identical in every mathematical context. Here they warn of the same issue: integral or variational control does not automatically keep the approximation inside the physically admissible pointwise range.

Why a Numerical Plot Is Not Enough 為什麼一張數值圖不夠

The decisive questions concern the approximation space and projection: Is the approximant continuous or discontinuous? What is its polynomial degree? Is the discontinuity aligned with the mesh? Does the defining norm control point values? Does the overshoot persist when the mesh is refined or the discontinuity is moved?

A boundary layer is a special aligned case because it is attached to the domain boundary. This helps explain why many DG-like methods appear free of Gibbs oscillations in boundary-layer tests. It does not establish pointwise control for general non-aligned interior discontinuities.

邊界層是一個特殊的對齊情形，因為它天然依附於計算域邊界。這有助於解釋許多 DG-like 方法為何在邊界層測試中看似沒有 Gibbs 振盪；但這不能證明它們對一般非對齊的內部間斷具有點值控制。

Why L¹ Minimization Is Not a General Cure 為什麼 L¹ 極小化不是一般性的解法

L¹-minimization is sometimes claimed to eliminate or strongly suppress the Gibbs phenomenon. Such a statement has no general validity. The absence of overshoot occurs only under conditions far more restrictive than mesh alignment alone. It depends on the approximation space, polynomial degree, local configuration, and, because an L¹-best approximation may be nonunique, even on which minimizer is selected.

L¹ 極小化有時被宣稱可以消除或顯著抑制 Gibbs 現象。這樣的說法並不具有一般有效性。無過衝只在遠比「網格與間斷對齊」更嚴格的條件下才可能出現；它還取決於逼近空間、多項式次數、局部配置，以及在 L¹ 最佳逼近不唯一時究竟選取哪一個極小化解。

In slightly more general settings, non-trivial overshoots reappear. Thus neither an L¹ objective nor mesh alignment supplies a general pointwise bound or a discrete maximum principle. The exceptional cases without overshoot cannot be promoted to a method-level claim.

在稍微一般一些的情形下，不可忽略的過衝便會重新出現。因此，使用 L¹ 目標函數或採用對齊網格，都不能提供一般性的逐點界或離散最大值原理。少數無過衝的特殊情形，不能被提升為對整個方法的一般性結論。

Moreover, L¹-minimization generally requires nonsmooth optimization rather than a single linear projection solve; it also supplies no general guarantee of pointwise overshoot control. Its computational cost and possible benefits should be assessed for the particular space and problem.

此外，L¹ 極小化的計算成本遠高於 L² 投影：它需要求解非光滑迭代優化問題，而不是一個線性系統，卻並未提供一般性的過衝控制改進。

Main Findings 主要結論

• For a bounded step function, the elementwise L² projection onto discontinuous piecewise constants consists of cell averages and has no overshoot, even on a non-matched mesh.
• Higher-order discontinuous approximations generally overshoot when a discontinuity cuts through an element.
• Continuous piecewise linear approximations exhibit non-trivial overshoots on regular meshes.
• Adaptive refinement may reduce integration-based error norms without eliminating pointwise overshoots.
• Neither L¹ minimization nor discontinuity alignment provides a general no-overshoot guarantee.

Core Paper 核心論文