This paper proposes a novel imputation-free deep incomplete multi-view clustering (DVIMC) method based on variational autoencoders (VAEs). The method addresses the challenges of incomplete multi-view clustering by leveraging the Product-of-Experts (PoE) approach and a coherence objective to effectively aggregate information from available views and mitigate the impact of information imbalance. DVIMC learns a common representation that captures shared clustering structures without requiring imputation of missing views. The method is designed to handle multi-view datasets with arbitrary missing views and achieves competitive clustering performance compared to state-of-the-art methods. The key contributions of this work include: (1) proposing a flexible DVIMC method based on VAEs, which is the first imputation-free work based on VAEs in the field of incomplete multi-view clustering; (2) employing the PoE approach combined with a coherence constraint to address incomplete learning and mitigate clustering bias caused by information imbalance; and (3) incorporating shared clustering assignment learning with coherence loss to simultaneously obtain common representations with clustering-friendly structures and optimal clustering results for incomplete multi-view data. The proposed method is evaluated on four real-world datasets: Caltech7-5V, Scene-15, Multi-Fashion, and NoisyMNIST. Experimental results show that DVIMC achieves competitive clustering performance compared to existing methods, particularly outperforming other VAE-based methods in all metrics. The method is also effective in learning shared clustering structures from incomplete multi-view data, as demonstrated by T-SNE visualization of the common representation. The method is trained using a variational lower bound (ELBO) objective, which includes a coherence objective loss to enforce consistency of information encoded by the aggregated representation. The overall objective loss combines the ELBO with a regularization parameter to balance the two objectives. The method is implemented using a deep multi-view Gaussian mixture model, where each view is processed by a view-specific encoder and decoder. The PoE approach is used to aggregate information from each view and derive the common representation. The coherence objective loss is introduced to mitigate the influence of information imbalance and ensure the consistency of the aggregated representation. The experiments demonstrate that DVIMC significantly outperforms other methods in clustering performance, especially when the number of missing views is large. The method is also effective in distinguishing samples from different clusters, as shown by the T-SNE visualization of the common representation. The parameter analysis shows that the regularization parameter α should be set to a moderate value to achieve optimal clustering performance. The method is applicable to a wide range of incomplete multi-view clustering tasks and is designed to be flexible and efficient in handling multi-view data with arbitrary missing views.This paper proposes a novel imputation-free deep incomplete multi-view clustering (DVIMC) method based on variational autoencoders (VAEs). The method addresses the challenges of incomplete multi-view clustering by leveraging the Product-of-Experts (PoE) approach and a coherence objective to effectively aggregate information from available views and mitigate the impact of information imbalance. DVIMC learns a common representation that captures shared clustering structures without requiring imputation of missing views. The method is designed to handle multi-view datasets with arbitrary missing views and achieves competitive clustering performance compared to state-of-the-art methods. The key contributions of this work include: (1) proposing a flexible DVIMC method based on VAEs, which is the first imputation-free work based on VAEs in the field of incomplete multi-view clustering; (2) employing the PoE approach combined with a coherence constraint to address incomplete learning and mitigate clustering bias caused by information imbalance; and (3) incorporating shared clustering assignment learning with coherence loss to simultaneously obtain common representations with clustering-friendly structures and optimal clustering results for incomplete multi-view data. The proposed method is evaluated on four real-world datasets: Caltech7-5V, Scene-15, Multi-Fashion, and NoisyMNIST. Experimental results show that DVIMC achieves competitive clustering performance compared to existing methods, particularly outperforming other VAE-based methods in all metrics. The method is also effective in learning shared clustering structures from incomplete multi-view data, as demonstrated by T-SNE visualization of the common representation. The method is trained using a variational lower bound (ELBO) objective, which includes a coherence objective loss to enforce consistency of information encoded by the aggregated representation. The overall objective loss combines the ELBO with a regularization parameter to balance the two objectives. The method is implemented using a deep multi-view Gaussian mixture model, where each view is processed by a view-specific encoder and decoder. The PoE approach is used to aggregate information from each view and derive the common representation. The coherence objective loss is introduced to mitigate the influence of information imbalance and ensure the consistency of the aggregated representation. The experiments demonstrate that DVIMC significantly outperforms other methods in clustering performance, especially when the number of missing views is large. The method is also effective in distinguishing samples from different clusters, as shown by the T-SNE visualization of the common representation. The parameter analysis shows that the regularization parameter α should be set to a moderate value to achieve optimal clustering performance. The method is applicable to a wide range of incomplete multi-view clustering tasks and is designed to be flexible and efficient in handling multi-view data with arbitrary missing views.